1
2 /*---------------------------------------------------------------*/
3 /*--- begin host_ppc_isel.c ---*/
4 /*---------------------------------------------------------------*/
5
6 /*
7 This file is part of Valgrind, a dynamic binary instrumentation
8 framework.
9
10 Copyright (C) 2004-2017 OpenWorks LLP
11 info@open-works.net
12
13 This program is free software; you can redistribute it and/or
14 modify it under the terms of the GNU General Public License as
15 published by the Free Software Foundation; either version 2 of the
16 License, or (at your option) any later version.
17
18 This program is distributed in the hope that it will be useful, but
19 WITHOUT ANY WARRANTY; without even the implied warranty of
20 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
21 General Public License for more details.
22
23 You should have received a copy of the GNU General Public License
24 along with this program; if not, write to the Free Software
25 Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA
26 02110-1301, USA.
27
28 The GNU General Public License is contained in the file COPYING.
29
30 Neither the names of the U.S. Department of Energy nor the
31 University of California nor the names of its contributors may be
32 used to endorse or promote products derived from this software
33 without prior written permission.
34 */
35
36 #include "libvex_basictypes.h"
37 #include "libvex_ir.h"
38 #include "libvex.h"
39
40 #include "ir_match.h"
41 #include "main_util.h"
42 #include "main_globals.h"
43 #include "host_generic_regs.h"
44 #include "host_generic_simd64.h"
45 #include "host_ppc_defs.h"
46
47 /* GPR register class for ppc32/64 */
48 #define HRcGPR(_mode64) ((_mode64) ? HRcInt64 : HRcInt32)
49
50
51 /*---------------------------------------------------------*/
52 /*--- Register Usage Conventions ---*/
53 /*---------------------------------------------------------*/
54 /*
55 Integer Regs
56 ------------
57 GPR0 Reserved
58 GPR1 Stack Pointer
59 GPR2 not used - TOC pointer
60 GPR3:10 Allocateable
61 GPR11 if mode64: not used - calls by ptr / env ptr for some langs
62 GPR12 if mode64: not used - exceptions / global linkage code
63 GPR13 not used - Thread-specific pointer
64 GPR14:28 Allocateable
65 GPR29 Unused by us (reserved for the dispatcher)
66 GPR30 AltiVec temp spill register
67 GPR31 GuestStatePointer
68
69 Of Allocateable regs:
70 if (mode64)
71 GPR3:10 Caller-saved regs
72 else
73 GPR3:12 Caller-saved regs
74 GPR14:29 Callee-saved regs
75
76 GPR3 [Return | Parameter] - carrying reg
77 GPR4:10 Parameter-carrying regs
78
79
80 Floating Point Regs
81 -------------------
82 FPR0:31 Allocateable
83
84 FPR0 Caller-saved - scratch reg
85 if (mode64)
86 FPR1:13 Caller-saved - param & return regs
87 else
88 FPR1:8 Caller-saved - param & return regs
89 FPR9:13 Caller-saved regs
90 FPR14:31 Callee-saved regs
91
92
93 Vector Regs (on processors with the VMX feature)
94 -----------
95 VR0-VR1 Volatile scratch registers
96 VR2-VR13 Volatile vector parameters registers
97 VR14-VR19 Volatile scratch registers
98 VR20-VR31 Non-volatile registers
99 VRSAVE Non-volatile 32-bit register
100 */
101
102
103 /*---------------------------------------------------------*/
104 /*--- PPC FP Status & Control Register Conventions ---*/
105 /*---------------------------------------------------------*/
106 /*
107 Vex-generated code expects to run with the FPU set as follows: all
108 exceptions masked. The rounding mode is set appropriately before
109 each floating point insn emitted (or left unchanged if known to be
110 correct already). There are a few fp insns (fmr,fneg,fabs,fnabs),
111 which are unaffected by the rm and so the rounding mode is not set
112 prior to them.
113
114 At least on MPC7447A (Mac Mini), frsqrte is also not affected by
115 rounding mode. At some point the ppc docs get sufficiently vague
116 that the only way to find out is to write test programs.
117 */
118 /* Notes on the FP instruction set, 6 Feb 06.
119
120 What exns -> CR1 ? Sets FPRF ? Observes RM ?
121 -------------------------------------------------------------
122
123 fmr[.] if . n n
124 fneg[.] if . n n
125 fabs[.] if . n n
126 fnabs[.] if . n n
127
128 fadd[.] if . y y
129 fadds[.] if . y y
130 fcfid[.] (Si64->dbl) if . y y
131 fcfidU[.] (Ui64->dbl) if . y y
132 fcfids[.] (Si64->sngl) if . Y Y
133 fcfidus[.] (Ui64->sngl) if . Y Y
134 fcmpo (cmp, result n n n
135 fcmpu to crfD) n n n
136 fctid[.] (dbl->i64) if . ->undef y
137 fctidz[.] (dbl->i64) if . ->undef rounds-to-zero
138 fctiw[.] (dbl->i32) if . ->undef y
139 fctiwz[.] (dbl->i32) if . ->undef rounds-to-zero
140 fdiv[.] if . y y
141 fdivs[.] if . y y
142 fmadd[.] if . y y
143 fmadds[.] if . y y
144 fmsub[.] if . y y
145 fmsubs[.] if . y y
146 fmul[.] if . y y
147 fmuls[.] if . y y
148
149 (note: for fnm*, rounding happens before final negation)
150 fnmadd[.] if . y y
151 fnmadds[.] if . y y
152 fnmsub[.] if . y y
153 fnmsubs[.] if . y y
154
155 fre[.] if . y y
156 fres[.] if . y y
157
158 frsqrte[.] if . y apparently not
159
160 fsqrt[.] if . y y
161 fsqrts[.] if . y y
162 fsub[.] if . y y
163 fsubs[.] if . y y
164
165
166 fpscr: bits 30-31 (ibm) is RM
167 24-29 (ibm) are exnmasks/non-IEEE bit, all zero
168 15-19 (ibm) is FPRF: class, <, =, >, UNord
169
170 ppc fe(guest) makes fpscr read as all zeros except RM (and maybe FPRF
171 in future)
172
173 mcrfs - move fpscr field to CR field
174 mtfsfi[.] - 4 bit imm moved to fpscr field
175 mtfsf[.] - move frS[low 1/2] to fpscr but using 8-bit field mask
176 mtfsb1[.] - set given fpscr bit
177 mtfsb0[.] - clear given fpscr bit
178 mffs[.] - move all fpscr to frD[low 1/2]
179
180 For [.] presumably cr1 is set with exn summary bits, as per
181 main FP insns
182
183 A single precision store truncates/denormalises the in-register value,
184 but does not round it. This is so that flds followed by fsts is
185 always the identity.
186 */
187
188
189 /*---------------------------------------------------------*/
190 /*--- misc helpers ---*/
191 /*---------------------------------------------------------*/
192
193 /* These are duplicated in guest-ppc/toIR.c */
unop(IROp op,IRExpr * a)194 static IRExpr* unop ( IROp op, IRExpr* a )
195 {
196 return IRExpr_Unop(op, a);
197 }
198
mkU32(UInt i)199 static IRExpr* mkU32 ( UInt i )
200 {
201 return IRExpr_Const(IRConst_U32(i));
202 }
203
bind(Int binder)204 static IRExpr* bind ( Int binder )
205 {
206 return IRExpr_Binder(binder);
207 }
208
isZeroU8(IRExpr * e)209 static Bool isZeroU8 ( IRExpr* e )
210 {
211 return e->tag == Iex_Const
212 && e->Iex.Const.con->tag == Ico_U8
213 && e->Iex.Const.con->Ico.U8 == 0;
214 }
215
216
217 /*---------------------------------------------------------*/
218 /*--- ISelEnv ---*/
219 /*---------------------------------------------------------*/
220
221 /* This carries around:
222
223 - A mapping from IRTemp to IRType, giving the type of any IRTemp we
224 might encounter. This is computed before insn selection starts,
225 and does not change.
226
227 - A mapping from IRTemp to HReg. This tells the insn selector
228 which virtual register(s) are associated with each IRTemp
229 temporary. This is computed before insn selection starts, and
230 does not change. We expect this mapping to map precisely the
231 same set of IRTemps as the type mapping does.
232
233 - vregmapLo holds the primary register for the IRTemp.
234 - vregmapMedLo holds the secondary register for the IRTemp,
235 if any is needed. That's only for Ity_I64 temps
236 in 32 bit mode or Ity_I128 temps in 64-bit mode.
237 - vregmapMedHi is only for dealing with Ity_I128 temps in
238 32 bit mode. It holds bits 95:64 (Intel numbering)
239 of the IRTemp.
240 - vregmapHi is also only for dealing with Ity_I128 temps
241 in 32 bit mode. It holds the most significant bits
242 (127:96 in Intel numbering) of the IRTemp.
243
244 - The code array, that is, the insns selected so far.
245
246 - A counter, for generating new virtual registers.
247
248 - The host subarchitecture we are selecting insns for.
249 This is set at the start and does not change.
250
251 - A Bool to tell us if the host is 32 or 64bit.
252 This is set at the start and does not change.
253
254 - An IRExpr*, which may be NULL, holding the IR expression (an
255 IRRoundingMode-encoded value) to which the FPU's rounding mode
256 was most recently set. Setting to NULL is always safe. Used to
257 avoid redundant settings of the FPU's rounding mode, as
258 described in set_FPU_rounding_mode below.
259
260 - A VexMiscInfo*, needed for knowing how to generate
261 function calls for this target.
262
263 - The maximum guest address of any guest insn in this block.
264 Actually, the address of the highest-addressed byte from any
265 insn in this block. Is set at the start and does not change.
266 This is used for detecting jumps which are definitely
267 forward-edges from this block, and therefore can be made
268 (chained) to the fast entry point of the destination, thereby
269 avoiding the destination's event check.
270 */
271
272 typedef
273 struct {
274 /* Constant -- are set at the start and do not change. */
275 IRTypeEnv* type_env;
276 // 64-bit mode 32-bit mode
277 HReg* vregmapLo; // Low 64-bits [63:0] Low 32-bits [31:0]
278 HReg* vregmapMedLo; // high 64-bits[127:64] Next 32-bits [63:32]
279 HReg* vregmapMedHi; // unused Next 32-bits [95:64]
280 HReg* vregmapHi; // unused highest 32-bits [127:96]
281 Int n_vregmap;
282
283 /* 27 Jan 06: Not currently used, but should be */
284 UInt hwcaps;
285
286 Bool mode64;
287
288 const VexAbiInfo* vbi; // unused
289
290 Bool chainingAllowed;
291 Addr64 max_ga;
292
293 /* These are modified as we go along. */
294 HInstrArray* code;
295 Int vreg_ctr;
296
297 IRExpr* previous_rm;
298 }
299 ISelEnv;
300
301
lookupIRTemp(ISelEnv * env,IRTemp tmp)302 static HReg lookupIRTemp ( ISelEnv* env, IRTemp tmp )
303 {
304 vassert(tmp >= 0);
305 vassert(tmp < env->n_vregmap);
306 return env->vregmapLo[tmp];
307 }
308
lookupIRTempPair(HReg * vrHI,HReg * vrLO,ISelEnv * env,IRTemp tmp)309 static void lookupIRTempPair ( HReg* vrHI, HReg* vrLO,
310 ISelEnv* env, IRTemp tmp )
311 {
312 vassert(tmp >= 0);
313 vassert(tmp < env->n_vregmap);
314 vassert(! hregIsInvalid(env->vregmapMedLo[tmp]));
315 *vrLO = env->vregmapLo[tmp];
316 *vrHI = env->vregmapMedLo[tmp];
317 }
318
319 /* Only for used in 32-bit mode */
lookupIRTempQuad(HReg * vrHi,HReg * vrMedHi,HReg * vrMedLo,HReg * vrLo,ISelEnv * env,IRTemp tmp)320 static void lookupIRTempQuad ( HReg* vrHi, HReg* vrMedHi, HReg* vrMedLo,
321 HReg* vrLo, ISelEnv* env, IRTemp tmp )
322 {
323 vassert(!env->mode64);
324 vassert(tmp >= 0);
325 vassert(tmp < env->n_vregmap);
326 vassert(! hregIsInvalid(env->vregmapMedLo[tmp]));
327 *vrHi = env->vregmapHi[tmp];
328 *vrMedHi = env->vregmapMedHi[tmp];
329 *vrMedLo = env->vregmapMedLo[tmp];
330 *vrLo = env->vregmapLo[tmp];
331 }
332
addInstr(ISelEnv * env,PPCInstr * instr)333 static void addInstr ( ISelEnv* env, PPCInstr* instr )
334 {
335 addHInstr(env->code, instr);
336 if (vex_traceflags & VEX_TRACE_VCODE) {
337 ppPPCInstr(instr, env->mode64);
338 vex_printf("\n");
339 }
340 }
341
newVRegI(ISelEnv * env)342 static HReg newVRegI ( ISelEnv* env )
343 {
344 HReg reg
345 = mkHReg(True/*vreg*/, HRcGPR(env->mode64), 0/*enc*/, env->vreg_ctr);
346 env->vreg_ctr++;
347 return reg;
348 }
349
newVRegF(ISelEnv * env)350 static HReg newVRegF ( ISelEnv* env )
351 {
352 HReg reg = mkHReg(True/*vreg*/, HRcFlt64, 0/*enc*/, env->vreg_ctr);
353 env->vreg_ctr++;
354 return reg;
355 }
356
newVRegV(ISelEnv * env)357 static HReg newVRegV ( ISelEnv* env )
358 {
359 HReg reg = mkHReg(True/*vreg*/, HRcVec128, 0/*enc*/, env->vreg_ctr);
360 env->vreg_ctr++;
361 return reg;
362 }
363
364
365 /*---------------------------------------------------------*/
366 /*--- ISEL: Forward declarations ---*/
367 /*---------------------------------------------------------*/
368
369 /* These are organised as iselXXX and iselXXX_wrk pairs. The
370 iselXXX_wrk do the real work, but are not to be called directly.
371 For each XXX, iselXXX calls its iselXXX_wrk counterpart, then
372 checks that all returned registers are virtual. You should not
373 call the _wrk version directly.
374
375 'Word' refers to the size of the native machine word, that is,
376 32-bit int in 32-bit mode and 64-bit int in 64-bit mode. '2Word'
377 therefore refers to a double-width (64/128-bit) quantity in two
378 integer registers.
379 */
380 /* 32-bit mode: compute an I8/I16/I32 into a GPR.
381 64-bit mode: compute an I8/I16/I32/I64 into a GPR. */
382 static HReg iselWordExpr_R_wrk ( ISelEnv* env, const IRExpr* e,
383 IREndness IEndianess );
384 static HReg iselWordExpr_R ( ISelEnv* env, const IRExpr* e,
385 IREndness IEndianess );
386
387 /* 32-bit mode: Compute an I8/I16/I32 into a RH
388 (reg-or-halfword-immediate).
389 64-bit mode: Compute an I8/I16/I32/I64 into a RH
390 (reg-or-halfword-immediate).
391 It's important to specify whether the immediate is to be regarded
392 as signed or not. If yes, this will never return -32768 as an
393 immediate; this guaranteed that all signed immediates that are
394 return can have their sign inverted if need be.
395 */
396 static PPCRH* iselWordExpr_RH_wrk ( ISelEnv* env,
397 Bool syned, const IRExpr* e,
398 IREndness IEndianess );
399 static PPCRH* iselWordExpr_RH ( ISelEnv* env,
400 Bool syned, const IRExpr* e,
401 IREndness IEndianess );
402
403 /* 32-bit mode: compute an I32 into a RI (reg or 32-bit immediate).
404 64-bit mode: compute an I64 into a RI (reg or 64-bit immediate). */
405 static PPCRI* iselWordExpr_RI_wrk ( ISelEnv* env, const IRExpr* e,
406 IREndness IEndianess );
407 static PPCRI* iselWordExpr_RI ( ISelEnv* env, const IRExpr* e,
408 IREndness IEndianess );
409
410 /* In 32 bit mode ONLY, compute an I8 into a
411 reg-or-5-bit-unsigned-immediate, the latter being an immediate in
412 the range 1 .. 31 inclusive. Used for doing shift amounts. */
413 static PPCRH* iselWordExpr_RH5u_wrk ( ISelEnv* env, const IRExpr* e,
414 IREndness IEndianess );
415 static PPCRH* iselWordExpr_RH5u ( ISelEnv* env, const IRExpr* e,
416 IREndness IEndianess );
417
418 /* In 64-bit mode ONLY, compute an I8 into a
419 reg-or-6-bit-unsigned-immediate, the latter being an immediate in
420 the range 1 .. 63 inclusive. Used for doing shift amounts. */
421 static PPCRH* iselWordExpr_RH6u_wrk ( ISelEnv* env, const IRExpr* e,
422 IREndness IEndianess );
423 static PPCRH* iselWordExpr_RH6u ( ISelEnv* env, const IRExpr* e,
424 IREndness IEndianess );
425
426 /* 32-bit mode: compute an I32 into an AMode.
427 64-bit mode: compute an I64 into an AMode.
428
429 Requires to know (xferTy) the type of data to be loaded/stored
430 using this amode. That is so that, for 64-bit code generation, any
431 PPCAMode_IR returned will have an index (immediate offset) field
432 that is guaranteed to be 4-aligned, if there is any chance that the
433 amode is to be used in ld/ldu/lda/std/stdu.
434
435 Since there are no such restrictions on 32-bit insns, xferTy is
436 ignored for 32-bit code generation. */
437 static PPCAMode* iselWordExpr_AMode_wrk ( ISelEnv* env, const IRExpr* e,
438 IRType xferTy,
439 IREndness IEndianess );
440 static PPCAMode* iselWordExpr_AMode ( ISelEnv* env, const IRExpr* e,
441 IRType xferTy,
442 IREndness IEndianess );
443
444 static void iselInt128Expr_to_32x4_wrk ( HReg* rHi, HReg* rMedHi,
445 HReg* rMedLo, HReg* rLo,
446 ISelEnv* env, const IRExpr* e,
447 IREndness IEndianess );
448 static void iselInt128Expr_to_32x4 ( HReg* rHi, HReg* rMedHi,
449 HReg* rMedLo, HReg* rLo,
450 ISelEnv* env, const IRExpr* e,
451 IREndness IEndianess );
452
453
454 /* 32-bit mode ONLY: compute an I64 into a GPR pair. */
455 static void iselInt64Expr_wrk ( HReg* rHi, HReg* rLo,
456 ISelEnv* env, const IRExpr* e,
457 IREndness IEndianess );
458 static void iselInt64Expr ( HReg* rHi, HReg* rLo,
459 ISelEnv* env, const IRExpr* e,
460 IREndness IEndianess );
461
462 /* 64-bit mode ONLY: compute an I128 into a GPR64 pair. */
463 static void iselInt128Expr_wrk ( HReg* rHi, HReg* rLo,
464 ISelEnv* env, const IRExpr* e,
465 IREndness IEndianess );
466
467 static void iselInt128Expr ( HReg* rHi, HReg* rLo,
468 ISelEnv* env, const IRExpr* e,
469 IREndness IEndianess );
470
471 static PPCCondCode iselCondCode_wrk ( ISelEnv* env, const IRExpr* e,
472 IREndness IEndianess );
473 static PPCCondCode iselCondCode ( ISelEnv* env, const IRExpr* e,
474 IREndness IEndianess );
475
476 static HReg iselDblExpr_wrk ( ISelEnv* env, const IRExpr* e,
477 IREndness IEndianess );
478 static HReg iselDblExpr ( ISelEnv* env, const IRExpr* e,
479 IREndness IEndianess );
480
481 static HReg iselFltExpr_wrk ( ISelEnv* env, const IRExpr* e,
482 IREndness IEndianess );
483 static HReg iselFltExpr ( ISelEnv* env, const IRExpr* e,
484 IREndness IEndianess );
485
486 static HReg iselVecExpr_wrk ( ISelEnv* env, const IRExpr* e,
487 IREndness IEndianess );
488 static HReg iselVecExpr ( ISelEnv* env, const IRExpr* e,
489 IREndness IEndianess );
490
491 /* 64-bit mode ONLY. */
492 static HReg iselDfp32Expr_wrk ( ISelEnv* env, const IRExpr* e,
493 IREndness IEndianess );
494 static HReg iselDfp32Expr ( ISelEnv* env, const IRExpr* e,
495 IREndness IEndianess );
496 static HReg iselDfp64Expr_wrk ( ISelEnv* env, const IRExpr* e,
497 IREndness IEndianess );
498 static HReg iselDfp64Expr ( ISelEnv* env, const IRExpr* e,
499 IREndness IEndianess );
500 static HReg iselFp128Expr_wrk ( ISelEnv* env, const IRExpr* e,
501 IREndness IEndianess);
502 static HReg iselFp128Expr ( ISelEnv* env, const IRExpr* e,
503 IREndness IEndianess);
504
505 /* 64-bit mode ONLY: compute an D128 into a GPR64 pair. */
506 static void iselDfp128Expr_wrk ( HReg* rHi, HReg* rLo, ISelEnv* env,
507 const IRExpr* e, IREndness IEndianess );
508 static void iselDfp128Expr ( HReg* rHi, HReg* rLo, ISelEnv* env,
509 const IRExpr* e, IREndness IEndianess );
510
511 /*---------------------------------------------------------*/
512 /*--- ISEL: Misc helpers ---*/
513 /*---------------------------------------------------------*/
514
515 /* Make an int reg-reg move. */
516
mk_iMOVds_RR(HReg r_dst,HReg r_src)517 static PPCInstr* mk_iMOVds_RR ( HReg r_dst, HReg r_src )
518 {
519 vassert(hregClass(r_dst) == hregClass(r_src));
520 vassert(hregClass(r_src) == HRcInt32 ||
521 hregClass(r_src) == HRcInt64);
522 return PPCInstr_Alu(Palu_OR, r_dst, r_src, PPCRH_Reg(r_src));
523 }
524
525 /* Advance/retreat %r1 by n. */
526
add_to_sp(ISelEnv * env,UInt n)527 static void add_to_sp ( ISelEnv* env, UInt n )
528 {
529 HReg sp = StackFramePtr(env->mode64);
530 vassert(n <= 1024 && (n%16) == 0);
531 addInstr(env, PPCInstr_Alu( Palu_ADD, sp, sp,
532 PPCRH_Imm(True,toUShort(n)) ));
533 }
534
sub_from_sp(ISelEnv * env,UInt n)535 static void sub_from_sp ( ISelEnv* env, UInt n )
536 {
537 HReg sp = StackFramePtr(env->mode64);
538 vassert(n <= 1024 && (n%16) == 0);
539 addInstr(env, PPCInstr_Alu( Palu_SUB, sp, sp,
540 PPCRH_Imm(True,toUShort(n)) ));
541 }
542
543 /*
544 returns a quadword aligned address on the stack
545 - copies SP, adds 16bytes, aligns to quadword.
546 use sub_from_sp(32) before calling this,
547 as expects to have 32 bytes to play with.
548 */
get_sp_aligned16(ISelEnv * env)549 static HReg get_sp_aligned16 ( ISelEnv* env )
550 {
551 HReg r = newVRegI(env);
552 HReg align16 = newVRegI(env);
553 addInstr(env, mk_iMOVds_RR(r, StackFramePtr(env->mode64)));
554 // add 16
555 addInstr(env, PPCInstr_Alu( Palu_ADD, r, r,
556 PPCRH_Imm(True,toUShort(16)) ));
557 // mask to quadword
558 addInstr(env,
559 PPCInstr_LI(align16, 0xFFFFFFFFFFFFFFF0ULL, env->mode64));
560 addInstr(env, PPCInstr_Alu(Palu_AND, r,r, PPCRH_Reg(align16)));
561 return r;
562 }
563
564
565
566 /* Load 2*I32 regs to fp reg */
mk_LoadRR32toFPR(ISelEnv * env,HReg r_srcHi,HReg r_srcLo)567 static HReg mk_LoadRR32toFPR ( ISelEnv* env,
568 HReg r_srcHi, HReg r_srcLo )
569 {
570 HReg fr_dst = newVRegF(env);
571 PPCAMode *am_addr0, *am_addr1;
572
573 vassert(!env->mode64);
574 vassert(hregClass(r_srcHi) == HRcInt32);
575 vassert(hregClass(r_srcLo) == HRcInt32);
576
577 sub_from_sp( env, 16 ); // Move SP down 16 bytes
578 am_addr0 = PPCAMode_IR( 0, StackFramePtr(env->mode64) );
579 am_addr1 = PPCAMode_IR( 4, StackFramePtr(env->mode64) );
580
581 // store hi,lo as Ity_I32's
582 addInstr(env, PPCInstr_Store( 4, am_addr0, r_srcHi, env->mode64 ));
583 addInstr(env, PPCInstr_Store( 4, am_addr1, r_srcLo, env->mode64 ));
584
585 // load as float
586 addInstr(env, PPCInstr_FpLdSt(True/*load*/, 8, fr_dst, am_addr0));
587
588 add_to_sp( env, 16 ); // Reset SP
589 return fr_dst;
590 }
591
592 /* Load I64 reg to fp reg */
mk_LoadR64toFPR(ISelEnv * env,HReg r_src)593 static HReg mk_LoadR64toFPR ( ISelEnv* env, HReg r_src )
594 {
595 HReg fr_dst = newVRegF(env);
596 PPCAMode *am_addr0;
597
598 vassert(env->mode64);
599 vassert(hregClass(r_src) == HRcInt64);
600
601 sub_from_sp( env, 16 ); // Move SP down 16 bytes
602 am_addr0 = PPCAMode_IR( 0, StackFramePtr(env->mode64) );
603
604 // store as Ity_I64
605 addInstr(env, PPCInstr_Store( 8, am_addr0, r_src, env->mode64 ));
606
607 // load as float
608 addInstr(env, PPCInstr_FpLdSt(True/*load*/, 8, fr_dst, am_addr0));
609
610 add_to_sp( env, 16 ); // Reset SP
611 return fr_dst;
612 }
613
614
615 /* Given an amode, return one which references 4 bytes further
616 along. */
617
advance4(ISelEnv * env,PPCAMode * am)618 static PPCAMode* advance4 ( ISelEnv* env, PPCAMode* am )
619 {
620 PPCAMode* am4 = dopyPPCAMode( am );
621 if (am4->tag == Pam_IR
622 && am4->Pam.IR.index + 4 <= 32767) {
623 am4->Pam.IR.index += 4;
624 } else {
625 vpanic("advance4(ppc,host)");
626 }
627 return am4;
628 }
629
630
631 /* Given a guest-state array descriptor, an index expression and a
632 bias, generate a PPCAMode pointing at the relevant piece of
633 guest state. */
634 static
genGuestArrayOffset(ISelEnv * env,IRRegArray * descr,IRExpr * off,Int bias,IREndness IEndianess)635 PPCAMode* genGuestArrayOffset ( ISelEnv* env, IRRegArray* descr,
636 IRExpr* off, Int bias, IREndness IEndianess )
637 {
638 HReg rtmp, roff;
639 Int elemSz = sizeofIRType(descr->elemTy);
640 Int nElems = descr->nElems;
641 Int shift = 0;
642
643 /* Throw out any cases we don't need. In theory there might be a
644 day where we need to handle others, but not today. */
645
646 if (nElems != 16 && nElems != 32)
647 vpanic("genGuestArrayOffset(ppc host)(1)");
648
649 switch (elemSz) {
650 case 4: shift = 2; break;
651 case 8: shift = 3; break;
652 default: vpanic("genGuestArrayOffset(ppc host)(2)");
653 }
654
655 if (bias < -100 || bias > 100) /* somewhat arbitrarily */
656 vpanic("genGuestArrayOffset(ppc host)(3)");
657 if (descr->base < 0 || descr->base > 5000) /* somewhat arbitrarily */
658 vpanic("genGuestArrayOffset(ppc host)(4)");
659
660 /* Compute off into a reg, %off. Then return:
661
662 addi %tmp, %off, bias (if bias != 0)
663 andi %tmp, nElems-1
664 sldi %tmp, shift
665 addi %tmp, %tmp, base
666 ... Baseblockptr + %tmp ...
667 */
668 roff = iselWordExpr_R(env, off, IEndianess);
669 rtmp = newVRegI(env);
670 addInstr(env, PPCInstr_Alu(
671 Palu_ADD,
672 rtmp, roff,
673 PPCRH_Imm(True/*signed*/, toUShort(bias))));
674 addInstr(env, PPCInstr_Alu(
675 Palu_AND,
676 rtmp, rtmp,
677 PPCRH_Imm(False/*unsigned*/, toUShort(nElems-1))));
678 addInstr(env, PPCInstr_Shft(
679 Pshft_SHL,
680 env->mode64 ? False : True/*F:64-bit, T:32-bit shift*/,
681 rtmp, rtmp,
682 PPCRH_Imm(False/*unsigned*/, toUShort(shift))));
683 addInstr(env, PPCInstr_Alu(
684 Palu_ADD,
685 rtmp, rtmp,
686 PPCRH_Imm(True/*signed*/, toUShort(descr->base))));
687 return
688 PPCAMode_RR( GuestStatePtr(env->mode64), rtmp );
689 }
690
691
692 /*---------------------------------------------------------*/
693 /*--- ISEL: Function call helpers ---*/
694 /*---------------------------------------------------------*/
695
696 /* Used only in doHelperCall. See big comment in doHelperCall re
697 handling of register-parameter args. This function figures out
698 whether evaluation of an expression might require use of a fixed
699 register. If in doubt return True (safe but suboptimal).
700 */
701 static
mightRequireFixedRegs(IRExpr * e)702 Bool mightRequireFixedRegs ( IRExpr* e )
703 {
704 switch (e->tag) {
705 case Iex_RdTmp: case Iex_Const: case Iex_Get:
706 return False;
707 default:
708 return True;
709 }
710 }
711
712
713 /* Do a complete function call. |guard| is a Ity_Bit expression
714 indicating whether or not the call happens. If guard==NULL, the
715 call is unconditional. |retloc| is set to indicate where the
716 return value is after the call. The caller (of this fn) must
717 generate code to add |stackAdjustAfterCall| to the stack pointer
718 after the call is done. */
719
720 static
doHelperCall(UInt * stackAdjustAfterCall,RetLoc * retloc,ISelEnv * env,IRExpr * guard,IRCallee * cee,IRType retTy,IRExpr ** args,IREndness IEndianess)721 void doHelperCall ( /*OUT*/UInt* stackAdjustAfterCall,
722 /*OUT*/RetLoc* retloc,
723 ISelEnv* env,
724 IRExpr* guard,
725 IRCallee* cee, IRType retTy, IRExpr** args,
726 IREndness IEndianess)
727 {
728 PPCCondCode cc;
729 HReg argregs[PPC_N_REGPARMS];
730 HReg tmpregs[PPC_N_REGPARMS];
731 Bool go_fast;
732 Int n_args, i, argreg;
733 UInt argiregs;
734 Bool mode64 = env->mode64;
735
736 /* Set default returns. We'll update them later if needed. */
737 *stackAdjustAfterCall = 0;
738 *retloc = mk_RetLoc_INVALID();
739
740 /* These are used for cross-checking that IR-level constraints on
741 the use of IRExpr_VECRET() and IRExpr_GSPTR() are observed. */
742 UInt nVECRETs = 0;
743 UInt nGSPTRs = 0;
744
745 /* Marshal args for a call and do the call.
746
747 This function only deals with a tiny set of possibilities, which
748 cover all helpers in practice. The restrictions are that only
749 arguments in registers are supported, hence only PPC_N_REGPARMS x
750 (mode32:32 | mode64:64) integer bits in total can be passed.
751 In fact the only supported arg type is (mode32:I32 | mode64:I64).
752
753 The return type can be I{64,32,16,8} or V{128,256}. In the
754 latter two cases, it is expected that |args| will contain the
755 special node IRExpr_VECRET(), in which case this routine
756 generates code to allocate space on the stack for the vector
757 return value. Since we are not passing any scalars on the
758 stack, it is enough to preallocate the return space before
759 marshalling any arguments, in this case.
760
761 |args| may also contain IRExpr_GSPTR(), in which case the value
762 in the guest state pointer register is passed as the
763 corresponding argument.
764
765 Generating code which is both efficient and correct when
766 parameters are to be passed in registers is difficult, for the
767 reasons elaborated in detail in comments attached to
768 doHelperCall() in priv/host-x86/isel.c. Here, we use a variant
769 of the method described in those comments.
770
771 The problem is split into two cases: the fast scheme and the
772 slow scheme. In the fast scheme, arguments are computed
773 directly into the target (real) registers. This is only safe
774 when we can be sure that computation of each argument will not
775 trash any real registers set by computation of any other
776 argument.
777
778 In the slow scheme, all args are first computed into vregs, and
779 once they are all done, they are moved to the relevant real
780 regs. This always gives correct code, but it also gives a bunch
781 of vreg-to-rreg moves which are usually redundant but are hard
782 for the register allocator to get rid of.
783
784 To decide which scheme to use, all argument expressions are
785 first examined. If they are all so simple that it is clear they
786 will be evaluated without use of any fixed registers, use the
787 fast scheme, else use the slow scheme. Note also that only
788 unconditional calls may use the fast scheme, since having to
789 compute a condition expression could itself trash real
790 registers.
791
792 Note this requires being able to examine an expression and
793 determine whether or not evaluation of it might use a fixed
794 register. That requires knowledge of how the rest of this insn
795 selector works. Currently just the following 3 are regarded as
796 safe -- hopefully they cover the majority of arguments in
797 practice: IRExpr_Tmp IRExpr_Const IRExpr_Get.
798 */
799
800 /* Note that the cee->regparms field is meaningless on PPC32/64 host
801 (since there is only one calling convention) and so we always
802 ignore it. */
803
804 n_args = 0;
805 for (i = 0; args[i]; i++)
806 n_args++;
807
808 if (n_args > PPC_N_REGPARMS) {
809 vpanic("doHelperCall(PPC): cannot currently handle > 8 args");
810 // PPC_N_REGPARMS
811 }
812
813 /* This is kind of stupid .. the arrays are sized as PPC_N_REGPARMS
814 but we then assume that that value is 8. */
815 vassert(PPC_N_REGPARMS == 8);
816
817 argregs[0] = hregPPC_GPR3(mode64);
818 argregs[1] = hregPPC_GPR4(mode64);
819 argregs[2] = hregPPC_GPR5(mode64);
820 argregs[3] = hregPPC_GPR6(mode64);
821 argregs[4] = hregPPC_GPR7(mode64);
822 argregs[5] = hregPPC_GPR8(mode64);
823 argregs[6] = hregPPC_GPR9(mode64);
824 argregs[7] = hregPPC_GPR10(mode64);
825 argiregs = 0;
826
827 tmpregs[0] = tmpregs[1] = tmpregs[2] =
828 tmpregs[3] = tmpregs[4] = tmpregs[5] =
829 tmpregs[6] = tmpregs[7] = INVALID_HREG;
830
831 /* First decide which scheme (slow or fast) is to be used. First
832 assume the fast scheme, and select slow if any contraindications
833 (wow) appear. */
834
835 go_fast = True;
836
837 /* We'll need space on the stack for the return value. Avoid
838 possible complications with nested calls by using the slow
839 scheme. */
840 if (retTy == Ity_V128 || retTy == Ity_V256)
841 go_fast = False;
842
843 if (go_fast && guard) {
844 if (guard->tag == Iex_Const
845 && guard->Iex.Const.con->tag == Ico_U1
846 && guard->Iex.Const.con->Ico.U1 == True) {
847 /* unconditional */
848 } else {
849 /* Not manifestly unconditional -- be conservative. */
850 go_fast = False;
851 }
852 }
853
854 if (go_fast) {
855 for (i = 0; i < n_args; i++) {
856 IRExpr* arg = args[i];
857 if (UNLIKELY(arg->tag == Iex_GSPTR)) {
858 /* that's OK */
859 }
860 else if (UNLIKELY(arg->tag == Iex_VECRET)) {
861 /* This implies ill-formed IR, since if the IR was
862 well-formed, the return-type test above would have
863 filtered it out. */
864 vpanic("doHelperCall(PPC): invalid IR");
865 }
866 else if (mightRequireFixedRegs(arg)) {
867 go_fast = False;
868 break;
869 }
870 }
871 }
872
873 /* At this point the scheme to use has been established. Generate
874 code to get the arg values into the argument rregs. */
875
876 if (go_fast) {
877
878 /* FAST SCHEME */
879 argreg = 0;
880
881 for (i = 0; i < n_args; i++) {
882 IRExpr* arg = args[i];
883 vassert(argreg < PPC_N_REGPARMS);
884
885 if (arg->tag == Iex_GSPTR) {
886 argiregs |= (1 << (argreg+3));
887 addInstr(env, mk_iMOVds_RR( argregs[argreg],
888 GuestStatePtr(mode64) ));
889 argreg++;
890 } else {
891 vassert(arg->tag != Iex_VECRET);
892 IRType ty = typeOfIRExpr(env->type_env, arg);
893 vassert(ty == Ity_I32 || ty == Ity_I64);
894 if (!mode64) {
895 if (ty == Ity_I32) {
896 argiregs |= (1 << (argreg+3));
897 addInstr(env,
898 mk_iMOVds_RR( argregs[argreg],
899 iselWordExpr_R(env, arg,
900 IEndianess) ));
901 } else { // Ity_I64 in 32-bit mode
902 HReg rHi, rLo;
903 if ((argreg%2) == 1)
904 // ppc32 ELF abi spec for passing LONG_LONG
905 argreg++; // XXX: odd argreg => even rN
906 vassert(argreg < PPC_N_REGPARMS-1);
907 iselInt64Expr(&rHi,&rLo, env, arg, IEndianess);
908 argiregs |= (1 << (argreg+3));
909 addInstr(env, mk_iMOVds_RR( argregs[argreg++], rHi ));
910 argiregs |= (1 << (argreg+3));
911 addInstr(env, mk_iMOVds_RR( argregs[argreg], rLo));
912 }
913 } else { // mode64
914 argiregs |= (1 << (argreg+3));
915 addInstr(env, mk_iMOVds_RR( argregs[argreg],
916 iselWordExpr_R(env, arg,
917 IEndianess) ));
918 }
919 argreg++;
920 } /* if (arg == IRExprP__BBPR) */
921 }
922
923 /* Fast scheme only applies for unconditional calls. Hence: */
924 cc = mk_PPCCondCode( Pct_ALWAYS, Pcf_NONE );
925
926 } else {
927
928 /* SLOW SCHEME; move via temporaries */
929 argreg = 0;
930
931 /* If we have a vector return type, allocate a place for it on
932 the stack and record its address. Rather than figure out the
933 complexities of PPC{32,64} ELF ABI stack frame layout, simply
934 drop the SP by 1024 and allocate the return point in the
935 middle. I think this should comfortably clear any ABI
936 mandated register save areas. Note that it doesn't maintain
937 the backchain as it should, since we're not doing st{d,w}u to
938 adjust the SP, but .. that doesn't seem to be a big deal.
939 Since we're not expecting to have to unwind out of here. */
940 HReg r_vecRetAddr = INVALID_HREG;
941 if (retTy == Ity_V128) {
942 r_vecRetAddr = newVRegI(env);
943 sub_from_sp(env, 512);
944 addInstr(env, mk_iMOVds_RR( r_vecRetAddr, StackFramePtr(mode64) ));
945 sub_from_sp(env, 512);
946 }
947 else if (retTy == Ity_V256) {
948 vassert(0); //ATC
949 r_vecRetAddr = newVRegI(env);
950 sub_from_sp(env, 512);
951 addInstr(env, mk_iMOVds_RR( r_vecRetAddr, StackFramePtr(mode64) ));
952 sub_from_sp(env, 512);
953 }
954
955 vassert(n_args >= 0 && n_args <= 8);
956 for (i = 0; i < n_args; i++) {
957 IRExpr* arg = args[i];
958 vassert(argreg < PPC_N_REGPARMS);
959 if (UNLIKELY(arg->tag == Iex_GSPTR)) {
960 tmpregs[argreg] = newVRegI(env);
961 addInstr(env, mk_iMOVds_RR( tmpregs[argreg],
962 GuestStatePtr(mode64) ));
963 nGSPTRs++;
964 }
965 else if (UNLIKELY(arg->tag == Iex_VECRET)) {
966 /* We stashed the address of the return slot earlier, so just
967 retrieve it now. */
968 vassert(!hregIsInvalid(r_vecRetAddr));
969 tmpregs[i] = r_vecRetAddr;
970 nVECRETs++;
971 }
972 else {
973 IRType ty = typeOfIRExpr(env->type_env, arg);
974 vassert(ty == Ity_I32 || ty == Ity_I64);
975 if (!mode64) {
976 if (ty == Ity_I32) {
977 tmpregs[argreg] = iselWordExpr_R(env, arg, IEndianess);
978 } else { // Ity_I64 in 32-bit mode
979 HReg rHi, rLo;
980 if ((argreg%2) == 1)
981 // ppc32 ELF abi spec for passing LONG_LONG
982 argreg++; // XXX: odd argreg => even rN
983 vassert(argreg < PPC_N_REGPARMS-1);
984 iselInt64Expr(&rHi,&rLo, env, arg, IEndianess);
985 tmpregs[argreg++] = rHi;
986 tmpregs[argreg] = rLo;
987 }
988 } else { // mode64
989 tmpregs[argreg] = iselWordExpr_R(env, arg, IEndianess);
990 }
991 }
992 argreg++;
993 }
994
995 /* Now we can compute the condition. We can't do it earlier
996 because the argument computations could trash the condition
997 codes. Be a bit clever to handle the common case where the
998 guard is 1:Bit. */
999 cc = mk_PPCCondCode( Pct_ALWAYS, Pcf_NONE );
1000 if (guard) {
1001 if (guard->tag == Iex_Const
1002 && guard->Iex.Const.con->tag == Ico_U1
1003 && guard->Iex.Const.con->Ico.U1 == True) {
1004 /* unconditional -- do nothing */
1005 } else {
1006 cc = iselCondCode( env, guard, IEndianess );
1007 }
1008 }
1009
1010 /* Move the args to their final destinations. */
1011 for (i = 0; i < argreg; i++) {
1012 if (hregIsInvalid(tmpregs[i])) // Skip invalid regs
1013 continue;
1014 /* None of these insns, including any spill code that might
1015 be generated, may alter the condition codes. */
1016 argiregs |= (1 << (i+3));
1017 addInstr( env, mk_iMOVds_RR( argregs[i], tmpregs[i] ) );
1018 }
1019
1020 }
1021
1022 /* Do final checks, set the return values, and generate the call
1023 instruction proper. */
1024 if (retTy == Ity_V128 || retTy == Ity_V256) {
1025 vassert(nVECRETs == 1);
1026 } else {
1027 vassert(nVECRETs == 0);
1028 }
1029
1030 vassert(nGSPTRs == 0 || nGSPTRs == 1);
1031
1032 vassert(*stackAdjustAfterCall == 0);
1033 vassert(is_RetLoc_INVALID(*retloc));
1034 switch (retTy) {
1035 case Ity_INVALID:
1036 /* Function doesn't return a value. */
1037 *retloc = mk_RetLoc_simple(RLPri_None);
1038 break;
1039 case Ity_I64:
1040 *retloc = mk_RetLoc_simple(mode64 ? RLPri_Int : RLPri_2Int);
1041 break;
1042 case Ity_I32: case Ity_I16: case Ity_I8:
1043 *retloc = mk_RetLoc_simple(RLPri_Int);
1044 break;
1045 case Ity_V128:
1046 /* Result is 512 bytes up the stack, and after it has been
1047 retrieved, adjust SP upwards by 1024. */
1048 *retloc = mk_RetLoc_spRel(RLPri_V128SpRel, 512);
1049 *stackAdjustAfterCall = 1024;
1050 break;
1051 case Ity_V256:
1052 vassert(0); // ATC
1053 /* Ditto */
1054 *retloc = mk_RetLoc_spRel(RLPri_V256SpRel, 512);
1055 *stackAdjustAfterCall = 1024;
1056 break;
1057 default:
1058 /* IR can denote other possible return types, but we don't
1059 handle those here. */
1060 vassert(0);
1061 }
1062
1063 /* Finally, generate the call itself. This needs the *retloc value
1064 set in the switch above, which is why it's at the end. */
1065
1066 Addr64 target = mode64 ? (Addr)cee->addr
1067 : toUInt((Addr)(cee->addr));
1068 addInstr(env, PPCInstr_Call( cc, target, argiregs, *retloc ));
1069 }
1070
1071
1072 /*---------------------------------------------------------*/
1073 /*--- ISEL: FP rounding mode helpers ---*/
1074 /*---------------------------------------------------------*/
1075
1076 ///* Set FPU's rounding mode to the default */
1077 //static
1078 //void set_FPU_rounding_default ( ISelEnv* env )
1079 //{
1080 // HReg fr_src = newVRegF(env);
1081 // HReg r_src = newVRegI(env);
1082 //
1083 // /* Default rounding mode = 0x0
1084 // Only supporting the rounding-mode bits - the rest of FPSCR is 0x0
1085 // - so we can set the whole register at once (faster)
1086 // note: upper 32 bits ignored by FpLdFPSCR
1087 // */
1088 // addInstr(env, PPCInstr_LI(r_src, 0x0, env->mode64));
1089 // if (env->mode64) {
1090 // fr_src = mk_LoadR64toFPR( env, r_src ); // 1*I64 -> F64
1091 // } else {
1092 // fr_src = mk_LoadRR32toFPR( env, r_src, r_src ); // 2*I32 -> F64
1093 // }
1094 // addInstr(env, PPCInstr_FpLdFPSCR( fr_src ));
1095 //}
1096
1097 /* Convert IR rounding mode to PPC encoding */
roundModeIRtoPPC(ISelEnv * env,HReg r_rmIR)1098 static HReg roundModeIRtoPPC ( ISelEnv* env, HReg r_rmIR )
1099 {
1100 /*
1101 rounding mode | PPC | IR
1102 -----------------------------------------------
1103 to nearest, ties to even | 000 | 000
1104 to zero | 001 | 011
1105 to +infinity | 010 | 010
1106 to -infinity | 011 | 001
1107 +++++ Below are the extended rounding modes for decimal floating point +++++
1108 to nearest, ties away from 0 | 100 | 100
1109 to nearest, ties toward 0 | 101 | 111
1110 to away from 0 | 110 | 110
1111 to prepare for shorter precision | 111 | 101
1112 */
1113 HReg r_rmPPC = newVRegI(env);
1114 HReg r_tmp1 = newVRegI(env);
1115 HReg r_tmp2 = newVRegI(env);
1116
1117 vassert(hregClass(r_rmIR) == HRcGPR(env->mode64));
1118
1119 // r_rmPPC = XOR(r_rmIR, r_rmIR << 1) & 3
1120 //
1121 // slwi tmp1, r_rmIR, 1
1122 // xor tmp1, r_rmIR, tmp1
1123 // andi r_rmPPC, tmp1, 3
1124
1125 addInstr(env, PPCInstr_Shft(Pshft_SHL, True/*32bit shift*/,
1126 r_tmp1, r_rmIR, PPCRH_Imm(False,1)));
1127
1128 addInstr( env, PPCInstr_Alu( Palu_AND,
1129 r_tmp2, r_tmp1, PPCRH_Imm( False, 3 ) ) );
1130
1131 addInstr( env, PPCInstr_Alu( Palu_XOR,
1132 r_rmPPC, r_rmIR, PPCRH_Reg( r_tmp2 ) ) );
1133
1134 return r_rmPPC;
1135 }
1136
1137
1138 /* Set the FPU's rounding mode: 'mode' is an I32-typed expression
1139 denoting a value in the range 0 .. 7, indicating a round mode
1140 encoded as per type IRRoundingMode. Set the PPC FPSCR to have the
1141 same rounding. When the dfp_rm arg is True, set the decimal
1142 floating point rounding mode bits (29:31); otherwise, set the
1143 binary floating point rounding mode bits (62:63).
1144
1145 For speed & simplicity, we're setting the *entire* FPSCR here.
1146
1147 Setting the rounding mode is expensive. So this function tries to
1148 avoid repeatedly setting the rounding mode to the same thing by
1149 first comparing 'mode' to the 'mode' tree supplied in the previous
1150 call to this function, if any. (The previous value is stored in
1151 env->previous_rm.) If 'mode' is a single IR temporary 't' and
1152 env->previous_rm is also just 't', then the setting is skipped.
1153
1154 This is safe because of the SSA property of IR: an IR temporary can
1155 only be defined once and so will have the same value regardless of
1156 where it appears in the block. Cool stuff, SSA.
1157
1158 A safety condition: all attempts to set the RM must be aware of
1159 this mechanism - by being routed through the functions here.
1160
1161 Of course this only helps if blocks where the RM is set more than
1162 once and it is set to the same value each time, *and* that value is
1163 held in the same IR temporary each time. In order to assure the
1164 latter as much as possible, the IR optimiser takes care to do CSE
1165 on any block with any sign of floating point activity.
1166 */
1167 static
_set_FPU_rounding_mode(ISelEnv * env,IRExpr * mode,Bool dfp_rm,IREndness IEndianess)1168 void _set_FPU_rounding_mode ( ISelEnv* env, IRExpr* mode, Bool dfp_rm,
1169 IREndness IEndianess )
1170 {
1171 HReg fr_src = newVRegF(env);
1172 HReg r_src;
1173
1174 vassert(typeOfIRExpr(env->type_env,mode) == Ity_I32);
1175
1176 /* Do we need to do anything? */
1177 if (env->previous_rm
1178 && env->previous_rm->tag == Iex_RdTmp
1179 && mode->tag == Iex_RdTmp
1180 && env->previous_rm->Iex.RdTmp.tmp == mode->Iex.RdTmp.tmp) {
1181 /* no - setting it to what it was before. */
1182 vassert(typeOfIRExpr(env->type_env, env->previous_rm) == Ity_I32);
1183 return;
1184 }
1185
1186 /* No luck - we better set it, and remember what we set it to. */
1187 env->previous_rm = mode;
1188
1189 /* Only supporting the rounding-mode bits - the rest of FPSCR is
1190 0x0 - so we can set the whole register at once (faster). */
1191
1192 // Resolve rounding mode and convert to PPC representation
1193 r_src = roundModeIRtoPPC( env, iselWordExpr_R(env, mode, IEndianess) );
1194
1195 // gpr -> fpr
1196 if (env->mode64) {
1197 if (dfp_rm) {
1198 HReg r_tmp1 = newVRegI( env );
1199 addInstr( env,
1200 PPCInstr_Shft( Pshft_SHL, False/*64bit shift*/,
1201 r_tmp1, r_src, PPCRH_Imm( False, 32 ) ) );
1202 fr_src = mk_LoadR64toFPR( env, r_tmp1 );
1203 } else {
1204 fr_src = mk_LoadR64toFPR( env, r_src ); // 1*I64 -> F64
1205 }
1206 } else {
1207 if (dfp_rm) {
1208 HReg r_zero = newVRegI( env );
1209 addInstr( env, PPCInstr_LI( r_zero, 0, env->mode64 ) );
1210 fr_src = mk_LoadRR32toFPR( env, r_src, r_zero );
1211 } else {
1212 fr_src = mk_LoadRR32toFPR( env, r_src, r_src ); // 2*I32 -> F64
1213 }
1214 }
1215
1216 // Move to FPSCR
1217 addInstr(env, PPCInstr_FpLdFPSCR( fr_src, dfp_rm ));
1218 }
1219
set_FPU_rounding_mode(ISelEnv * env,IRExpr * mode,IREndness IEndianess)1220 static void set_FPU_rounding_mode ( ISelEnv* env, IRExpr* mode,
1221 IREndness IEndianess )
1222 {
1223 _set_FPU_rounding_mode(env, mode, False, IEndianess);
1224 }
1225
set_FPU_DFP_rounding_mode(ISelEnv * env,IRExpr * mode,IREndness IEndianess)1226 static void set_FPU_DFP_rounding_mode ( ISelEnv* env, IRExpr* mode,
1227 IREndness IEndianess )
1228 {
1229 _set_FPU_rounding_mode(env, mode, True, IEndianess);
1230 }
1231
1232 static
FPU_rounding_mode_isOdd(IRExpr * mode)1233 Bool FPU_rounding_mode_isOdd (IRExpr* mode) {
1234 /* If the rounding mode is set to odd, the the expr must be a constant U8
1235 * value equal to 8. Otherwise, it must be a bin op expressiong that
1236 * calculates the value.
1237 */
1238
1239 if (mode->tag != Iex_Const)
1240 return False;
1241
1242 vassert(mode->Iex.Const.con->tag == Ico_U32);
1243 vassert(mode->Iex.Const.con->Ico.U32 == 0x8);
1244 return True;
1245 }
1246
1247 /*---------------------------------------------------------*/
1248 /*--- ISEL: vector helpers ---*/
1249 /*---------------------------------------------------------*/
1250
1251 /* Generate all-zeroes into a new vector register.
1252 */
generate_zeroes_V128(ISelEnv * env)1253 static HReg generate_zeroes_V128 ( ISelEnv* env )
1254 {
1255 HReg dst = newVRegV(env);
1256 addInstr(env, PPCInstr_AvBinary(Pav_XOR, dst, dst, dst));
1257 return dst;
1258 }
1259
1260 /* Generate all-ones into a new vector register.
1261 */
generate_ones_V128(ISelEnv * env)1262 static HReg generate_ones_V128 ( ISelEnv* env )
1263 {
1264 HReg dst = newVRegV(env);
1265 PPCVI5s * src = PPCVI5s_Imm(-1);
1266 addInstr(env, PPCInstr_AvSplat(8, dst, src));
1267 return dst;
1268 }
1269
1270
1271 /*
1272 Generates code for AvSplat
1273 - takes in IRExpr* of type 8|16|32
1274 returns vector reg of duplicated lanes of input
1275 - uses AvSplat(imm) for imms up to simm6.
1276 otherwise must use store reg & load vector
1277 */
mk_AvDuplicateRI(ISelEnv * env,IRExpr * e,IREndness IEndianess)1278 static HReg mk_AvDuplicateRI( ISelEnv* env, IRExpr* e, IREndness IEndianess )
1279 {
1280 HReg r_src;
1281 HReg dst = newVRegV(env);
1282 PPCRI* ri = iselWordExpr_RI(env, e, IEndianess);
1283 IRType ty = typeOfIRExpr(env->type_env,e);
1284 UInt sz = (ty == Ity_I8) ? 8 : (ty == Ity_I16) ? 16 : 32;
1285 vassert(ty == Ity_I8 || ty == Ity_I16 || ty == Ity_I32);
1286
1287 /* special case: immediate */
1288 if (ri->tag == Pri_Imm) {
1289 Int simm32 = (Int)ri->Pri.Imm;
1290
1291 /* figure out if it's do-able with imm splats. */
1292 if (simm32 >= -32 && simm32 <= 31) {
1293 Char simm6 = (Char)simm32;
1294 if (simm6 > 15) { /* 16:31 inclusive */
1295 HReg v1 = newVRegV(env);
1296 HReg v2 = newVRegV(env);
1297 addInstr(env, PPCInstr_AvSplat(sz, v1, PPCVI5s_Imm(-16)));
1298 addInstr(env, PPCInstr_AvSplat(sz, v2, PPCVI5s_Imm(simm6-16)));
1299 addInstr(env,
1300 (sz== 8) ? PPCInstr_AvBin8x16(Pav_SUBU, dst, v2, v1) :
1301 (sz==16) ? PPCInstr_AvBin16x8(Pav_SUBU, dst, v2, v1)
1302 : PPCInstr_AvBin32x4(Pav_SUBU, dst, v2, v1) );
1303 return dst;
1304 }
1305 if (simm6 < -16) { /* -32:-17 inclusive */
1306 HReg v1 = newVRegV(env);
1307 HReg v2 = newVRegV(env);
1308 addInstr(env, PPCInstr_AvSplat(sz, v1, PPCVI5s_Imm(-16)));
1309 addInstr(env, PPCInstr_AvSplat(sz, v2, PPCVI5s_Imm(simm6+16)));
1310 addInstr(env,
1311 (sz== 8) ? PPCInstr_AvBin8x16(Pav_ADDU, dst, v2, v1) :
1312 (sz==16) ? PPCInstr_AvBin16x8(Pav_ADDU, dst, v2, v1)
1313 : PPCInstr_AvBin32x4(Pav_ADDU, dst, v2, v1) );
1314 return dst;
1315 }
1316 /* simplest form: -16:15 inclusive */
1317 addInstr(env, PPCInstr_AvSplat(sz, dst, PPCVI5s_Imm(simm6)));
1318 return dst;
1319 }
1320
1321 /* no luck; use the Slow way. */
1322 r_src = newVRegI(env);
1323 addInstr(env, PPCInstr_LI(r_src, (Long)simm32, env->mode64));
1324 }
1325 else {
1326 r_src = ri->Pri.Reg;
1327 }
1328
1329 {
1330 /* Store r_src multiple times (sz dependent); then load the dest vector. */
1331 HReg r_aligned16;
1332 PPCAMode *am_offset, *am_offset_zero;
1333
1334 sub_from_sp( env, 32 ); // Move SP down
1335 /* Get a 16-aligned address within our stack space */
1336 r_aligned16 = get_sp_aligned16( env );
1337
1338 Int i;
1339 Int stride = (sz == 8) ? 1 : (sz == 16) ? 2 : 4;
1340 UChar num_bytes_to_store = stride;
1341 am_offset_zero = PPCAMode_IR( 0, r_aligned16 );
1342 am_offset = am_offset_zero;
1343 for (i = 0; i < 16; i+=stride, am_offset = PPCAMode_IR( i, r_aligned16)) {
1344 addInstr(env, PPCInstr_Store( num_bytes_to_store, am_offset, r_src, env->mode64 ));
1345 }
1346
1347 /* Effectively splat the r_src value to dst */
1348 addInstr(env, PPCInstr_AvLdSt( True/*ld*/, 16, dst, am_offset_zero ) );
1349 add_to_sp( env, 32 ); // Reset SP
1350
1351 return dst;
1352 }
1353 }
1354
1355
1356 /* for each lane of vSrc: lane == nan ? laneX = all 1's : all 0's */
isNan(ISelEnv * env,HReg vSrc,IREndness IEndianess)1357 static HReg isNan ( ISelEnv* env, HReg vSrc, IREndness IEndianess )
1358 {
1359 HReg zeros, msk_exp, msk_mnt, expt, mnts, vIsNan;
1360
1361 vassert(hregClass(vSrc) == HRcVec128);
1362
1363 zeros = mk_AvDuplicateRI(env, mkU32(0), IEndianess);
1364 msk_exp = mk_AvDuplicateRI(env, mkU32(0x7F800000), IEndianess);
1365 msk_mnt = mk_AvDuplicateRI(env, mkU32(0x7FFFFF), IEndianess);
1366 expt = newVRegV(env);
1367 mnts = newVRegV(env);
1368 vIsNan = newVRegV(env);
1369
1370 /* 32bit float => sign(1) | exponent(8) | mantissa(23)
1371 nan => exponent all ones, mantissa > 0 */
1372
1373 addInstr(env, PPCInstr_AvBinary(Pav_AND, expt, vSrc, msk_exp));
1374 addInstr(env, PPCInstr_AvBin32x4(Pav_CMPEQU, expt, expt, msk_exp));
1375 addInstr(env, PPCInstr_AvBinary(Pav_AND, mnts, vSrc, msk_mnt));
1376 addInstr(env, PPCInstr_AvBin32x4(Pav_CMPGTU, mnts, mnts, zeros));
1377 addInstr(env, PPCInstr_AvBinary(Pav_AND, vIsNan, expt, mnts));
1378 return vIsNan;
1379 }
1380
1381
1382 /*---------------------------------------------------------*/
1383 /*--- ISEL: Integer expressions (64/32/16/8 bit) ---*/
1384 /*---------------------------------------------------------*/
1385
1386 /* Select insns for an integer-typed expression, and add them to the
1387 code list. Return a reg holding the result. This reg will be a
1388 virtual register. THE RETURNED REG MUST NOT BE MODIFIED. If you
1389 want to modify it, ask for a new vreg, copy it in there, and modify
1390 the copy. The register allocator will do its best to map both
1391 vregs to the same real register, so the copies will often disappear
1392 later in the game.
1393
1394 This should handle expressions of 64, 32, 16 and 8-bit type.
1395 All results are returned in a (mode64 ? 64bit : 32bit) register.
1396 For 16- and 8-bit expressions, the upper (32/48/56 : 16/24) bits
1397 are arbitrary, so you should mask or sign extend partial values
1398 if necessary.
1399 */
1400
iselWordExpr_R(ISelEnv * env,const IRExpr * e,IREndness IEndianess)1401 static HReg iselWordExpr_R ( ISelEnv* env, const IRExpr* e,
1402 IREndness IEndianess )
1403 {
1404 HReg r = iselWordExpr_R_wrk(env, e, IEndianess);
1405 /* sanity checks ... */
1406 # if 0
1407 vex_printf("\n"); ppIRExpr(e); vex_printf("\n");
1408 # endif
1409
1410 vassert(hregClass(r) == HRcGPR(env->mode64));
1411 vassert(hregIsVirtual(r));
1412 return r;
1413 }
1414
1415 /* DO NOT CALL THIS DIRECTLY ! */
iselWordExpr_R_wrk(ISelEnv * env,const IRExpr * e,IREndness IEndianess)1416 static HReg iselWordExpr_R_wrk ( ISelEnv* env, const IRExpr* e,
1417 IREndness IEndianess )
1418 {
1419 Bool mode64 = env->mode64;
1420 MatchInfo mi;
1421 DECLARE_PATTERN(p_32to1_then_1Uto8);
1422
1423 IRType ty = typeOfIRExpr(env->type_env,e);
1424 vassert(ty == Ity_I8 || ty == Ity_I16 ||
1425 ty == Ity_I32 || ((ty == Ity_I64) && mode64));
1426
1427 switch (e->tag) {
1428
1429 /* --------- TEMP --------- */
1430 case Iex_RdTmp:
1431 return lookupIRTemp(env, e->Iex.RdTmp.tmp);
1432
1433 /* --------- LOAD --------- */
1434 case Iex_Load: {
1435 HReg r_dst;
1436 PPCAMode* am_addr;
1437 if (e->Iex.Load.end != IEndianess)
1438 goto irreducible;
1439 r_dst = newVRegI(env);
1440 am_addr = iselWordExpr_AMode( env, e->Iex.Load.addr, ty/*of xfer*/,
1441 IEndianess );
1442 addInstr(env, PPCInstr_Load( toUChar(sizeofIRType(ty)),
1443 r_dst, am_addr, mode64 ));
1444 return r_dst;
1445 /*NOTREACHED*/
1446 }
1447
1448 /* --------- BINARY OP --------- */
1449 case Iex_Binop: {
1450 PPCAluOp aluOp;
1451 PPCShftOp shftOp;
1452
1453 /* Is it an addition or logical style op? */
1454 switch (e->Iex.Binop.op) {
1455 case Iop_Add8: case Iop_Add16: case Iop_Add32: case Iop_Add64:
1456 aluOp = Palu_ADD; break;
1457 case Iop_Sub8: case Iop_Sub16: case Iop_Sub32: case Iop_Sub64:
1458 aluOp = Palu_SUB; break;
1459 case Iop_And8: case Iop_And16: case Iop_And32: case Iop_And64:
1460 aluOp = Palu_AND; break;
1461 case Iop_Or8: case Iop_Or16: case Iop_Or32: case Iop_Or64:
1462 aluOp = Palu_OR; break;
1463 case Iop_Xor8: case Iop_Xor16: case Iop_Xor32: case Iop_Xor64:
1464 aluOp = Palu_XOR; break;
1465 default:
1466 aluOp = Palu_INVALID; break;
1467 }
1468 /* For commutative ops we assume any literal
1469 values are on the second operand. */
1470 if (aluOp != Palu_INVALID) {
1471 HReg r_dst = newVRegI(env);
1472 HReg r_srcL = iselWordExpr_R(env, e->Iex.Binop.arg1, IEndianess);
1473 PPCRH* ri_srcR = NULL;
1474 /* get right arg into an RH, in the appropriate way */
1475 switch (aluOp) {
1476 case Palu_ADD: case Palu_SUB:
1477 ri_srcR = iselWordExpr_RH(env, True/*signed*/,
1478 e->Iex.Binop.arg2, IEndianess);
1479 break;
1480 case Palu_AND: case Palu_OR: case Palu_XOR:
1481 ri_srcR = iselWordExpr_RH(env, False/*signed*/,
1482 e->Iex.Binop.arg2, IEndianess);
1483 break;
1484 default:
1485 vpanic("iselWordExpr_R_wrk-aluOp-arg2");
1486 }
1487 addInstr(env, PPCInstr_Alu(aluOp, r_dst, r_srcL, ri_srcR));
1488 return r_dst;
1489 }
1490
1491 /* a shift? */
1492 switch (e->Iex.Binop.op) {
1493 case Iop_Shl8: case Iop_Shl16: case Iop_Shl32: case Iop_Shl64:
1494 shftOp = Pshft_SHL; break;
1495 case Iop_Shr8: case Iop_Shr16: case Iop_Shr32: case Iop_Shr64:
1496 shftOp = Pshft_SHR; break;
1497 case Iop_Sar8: case Iop_Sar16: case Iop_Sar32: case Iop_Sar64:
1498 shftOp = Pshft_SAR; break;
1499 default:
1500 shftOp = Pshft_INVALID; break;
1501 }
1502 /* we assume any literal values are on the second operand. */
1503 if (shftOp != Pshft_INVALID) {
1504 HReg r_dst = newVRegI(env);
1505 HReg r_srcL = iselWordExpr_R(env, e->Iex.Binop.arg1, IEndianess);
1506 PPCRH* ri_srcR = NULL;
1507 /* get right arg into an RH, in the appropriate way */
1508 switch (shftOp) {
1509 case Pshft_SHL: case Pshft_SHR: case Pshft_SAR:
1510 if (!mode64)
1511 ri_srcR = iselWordExpr_RH5u(env, e->Iex.Binop.arg2, IEndianess);
1512 else
1513 ri_srcR = iselWordExpr_RH6u(env, e->Iex.Binop.arg2, IEndianess);
1514 break;
1515 default:
1516 vpanic("iselIntExpr_R_wrk-shftOp-arg2");
1517 }
1518 /* widen the left arg if needed */
1519 if (shftOp == Pshft_SHR || shftOp == Pshft_SAR) {
1520 if (ty == Ity_I8 || ty == Ity_I16) {
1521 PPCRH* amt = PPCRH_Imm(False,
1522 toUShort(ty == Ity_I8 ? 24 : 16));
1523 HReg tmp = newVRegI(env);
1524 addInstr(env, PPCInstr_Shft(Pshft_SHL,
1525 True/*32bit shift*/,
1526 tmp, r_srcL, amt));
1527 addInstr(env, PPCInstr_Shft(shftOp,
1528 True/*32bit shift*/,
1529 tmp, tmp, amt));
1530 r_srcL = tmp;
1531 }
1532 }
1533 /* Only 64 expressions need 64bit shifts,
1534 32bit shifts are fine for all others */
1535 if (ty == Ity_I64) {
1536 vassert(mode64);
1537 addInstr(env, PPCInstr_Shft(shftOp, False/*64bit shift*/,
1538 r_dst, r_srcL, ri_srcR));
1539 } else {
1540 addInstr(env, PPCInstr_Shft(shftOp, True/*32bit shift*/,
1541 r_dst, r_srcL, ri_srcR));
1542 }
1543 return r_dst;
1544 }
1545
1546 /* How about a div? */
1547 if (e->Iex.Binop.op == Iop_DivS32 ||
1548 e->Iex.Binop.op == Iop_DivU32 ||
1549 e->Iex.Binop.op == Iop_DivS32E ||
1550 e->Iex.Binop.op == Iop_DivU32E) {
1551 Bool syned = toBool((e->Iex.Binop.op == Iop_DivS32) || (e->Iex.Binop.op == Iop_DivS32E));
1552 HReg r_dst = newVRegI(env);
1553 HReg r_srcL = iselWordExpr_R(env, e->Iex.Binop.arg1, IEndianess);
1554 HReg r_srcR = iselWordExpr_R(env, e->Iex.Binop.arg2, IEndianess);
1555 addInstr( env,
1556 PPCInstr_Div( ( ( e->Iex.Binop.op == Iop_DivU32E )
1557 || ( e->Iex.Binop.op == Iop_DivS32E ) ) ? True
1558 : False,
1559 syned,
1560 True/*32bit div*/,
1561 r_dst,
1562 r_srcL,
1563 r_srcR ) );
1564 return r_dst;
1565 }
1566 if (e->Iex.Binop.op == Iop_DivS64 ||
1567 e->Iex.Binop.op == Iop_DivU64 || e->Iex.Binop.op == Iop_DivS64E
1568 || e->Iex.Binop.op == Iop_DivU64E ) {
1569 Bool syned = toBool((e->Iex.Binop.op == Iop_DivS64) ||(e->Iex.Binop.op == Iop_DivS64E));
1570 HReg r_dst = newVRegI(env);
1571 HReg r_srcL = iselWordExpr_R(env, e->Iex.Binop.arg1, IEndianess);
1572 HReg r_srcR = iselWordExpr_R(env, e->Iex.Binop.arg2, IEndianess);
1573 vassert(mode64);
1574 addInstr( env,
1575 PPCInstr_Div( ( ( e->Iex.Binop.op == Iop_DivS64E )
1576 || ( e->Iex.Binop.op
1577 == Iop_DivU64E ) ) ? True
1578 : False,
1579 syned,
1580 False/*64bit div*/,
1581 r_dst,
1582 r_srcL,
1583 r_srcR ) );
1584 return r_dst;
1585 }
1586
1587 /* No? Anyone for a mul? */
1588 if (e->Iex.Binop.op == Iop_Mul32
1589 || e->Iex.Binop.op == Iop_Mul64) {
1590 Bool syned = False;
1591 Bool sz32 = (e->Iex.Binop.op != Iop_Mul64);
1592 HReg r_dst = newVRegI(env);
1593 HReg r_srcL = iselWordExpr_R(env, e->Iex.Binop.arg1, IEndianess);
1594 HReg r_srcR = iselWordExpr_R(env, e->Iex.Binop.arg2, IEndianess);
1595 addInstr(env, PPCInstr_MulL(syned, False/*lo32*/, sz32,
1596 r_dst, r_srcL, r_srcR));
1597 return r_dst;
1598 }
1599
1600 /* 32 x 32 -> 64 multiply */
1601 if (mode64
1602 && (e->Iex.Binop.op == Iop_MullU32
1603 || e->Iex.Binop.op == Iop_MullS32)) {
1604 HReg tLo = newVRegI(env);
1605 HReg tHi = newVRegI(env);
1606 HReg r_dst = newVRegI(env);
1607 Bool syned = toBool(e->Iex.Binop.op == Iop_MullS32);
1608 HReg r_srcL = iselWordExpr_R(env, e->Iex.Binop.arg1, IEndianess);
1609 HReg r_srcR = iselWordExpr_R(env, e->Iex.Binop.arg2, IEndianess);
1610 addInstr(env, PPCInstr_MulL(False/*signedness irrelevant*/,
1611 False/*lo32*/, True/*32bit mul*/,
1612 tLo, r_srcL, r_srcR));
1613 addInstr(env, PPCInstr_MulL(syned,
1614 True/*hi32*/, True/*32bit mul*/,
1615 tHi, r_srcL, r_srcR));
1616 addInstr(env, PPCInstr_Shft(Pshft_SHL, False/*64bit shift*/,
1617 r_dst, tHi, PPCRH_Imm(False,32)));
1618 addInstr(env, PPCInstr_Alu(Palu_OR,
1619 r_dst, r_dst, PPCRH_Reg(tLo)));
1620 return r_dst;
1621 }
1622
1623 /* El-mutanto 3-way compare? */
1624 if (e->Iex.Binop.op == Iop_CmpORD32S
1625 || e->Iex.Binop.op == Iop_CmpORD32U) {
1626 Bool syned = toBool(e->Iex.Binop.op == Iop_CmpORD32S);
1627 HReg dst = newVRegI(env);
1628 HReg srcL = iselWordExpr_R(env, e->Iex.Binop.arg1, IEndianess);
1629 PPCRH* srcR = iselWordExpr_RH(env, syned, e->Iex.Binop.arg2,
1630 IEndianess);
1631 addInstr(env, PPCInstr_Cmp(syned, True/*32bit cmp*/,
1632 7/*cr*/, srcL, srcR));
1633 addInstr(env, PPCInstr_MfCR(dst));
1634 addInstr(env, PPCInstr_Alu(Palu_AND, dst, dst,
1635 PPCRH_Imm(False,7<<1)));
1636 return dst;
1637 }
1638
1639 if (e->Iex.Binop.op == Iop_CmpORD64S
1640 || e->Iex.Binop.op == Iop_CmpORD64U) {
1641 Bool syned = toBool(e->Iex.Binop.op == Iop_CmpORD64S);
1642 HReg dst = newVRegI(env);
1643 HReg srcL = iselWordExpr_R(env, e->Iex.Binop.arg1, IEndianess);
1644 PPCRH* srcR = iselWordExpr_RH(env, syned, e->Iex.Binop.arg2,
1645 IEndianess);
1646 vassert(mode64);
1647 addInstr(env, PPCInstr_Cmp(syned, False/*64bit cmp*/,
1648 7/*cr*/, srcL, srcR));
1649 addInstr(env, PPCInstr_MfCR(dst));
1650 addInstr(env, PPCInstr_Alu(Palu_AND, dst, dst,
1651 PPCRH_Imm(False,7<<1)));
1652 return dst;
1653 }
1654
1655 if (e->Iex.Binop.op == Iop_Max32U) {
1656 HReg r1 = iselWordExpr_R(env, e->Iex.Binop.arg1, IEndianess);
1657 HReg r2 = iselWordExpr_R(env, e->Iex.Binop.arg2, IEndianess);
1658 HReg rdst = newVRegI(env);
1659 PPCCondCode cc = mk_PPCCondCode( Pct_TRUE, Pcf_7LT );
1660 addInstr(env, mk_iMOVds_RR(rdst, r1));
1661 addInstr(env, PPCInstr_Cmp(False/*unsigned*/, True/*32bit cmp*/,
1662 7/*cr*/, rdst, PPCRH_Reg(r2)));
1663 addInstr(env, PPCInstr_CMov(cc, rdst, PPCRI_Reg(r2)));
1664 return rdst;
1665 }
1666
1667 if (e->Iex.Binop.op == Iop_32HLto64) {
1668 HReg r_Hi = iselWordExpr_R(env, e->Iex.Binop.arg1, IEndianess);
1669 HReg r_Lo = iselWordExpr_R(env, e->Iex.Binop.arg2, IEndianess);
1670 HReg r_Tmp = newVRegI(env);
1671 HReg r_dst = newVRegI(env);
1672 HReg msk = newVRegI(env);
1673 vassert(mode64);
1674 /* r_dst = OR( r_Hi<<32, r_Lo ) */
1675 addInstr(env, PPCInstr_Shft(Pshft_SHL, False/*64bit shift*/,
1676 r_dst, r_Hi, PPCRH_Imm(False,32)));
1677 addInstr(env, PPCInstr_LI(msk, 0xFFFFFFFF, mode64));
1678 addInstr(env, PPCInstr_Alu( Palu_AND, r_Tmp, r_Lo,
1679 PPCRH_Reg(msk) ));
1680 addInstr(env, PPCInstr_Alu( Palu_OR, r_dst, r_dst,
1681 PPCRH_Reg(r_Tmp) ));
1682 return r_dst;
1683 }
1684
1685 if ((e->Iex.Binop.op == Iop_CmpF64) ||
1686 (e->Iex.Binop.op == Iop_CmpD64) ||
1687 (e->Iex.Binop.op == Iop_CmpD128)) {
1688 HReg fr_srcL;
1689 HReg fr_srcL_lo;
1690 HReg fr_srcR;
1691 HReg fr_srcR_lo;
1692
1693 HReg r_ccPPC = newVRegI(env);
1694 HReg r_ccIR = newVRegI(env);
1695 HReg r_ccIR_b0 = newVRegI(env);
1696 HReg r_ccIR_b2 = newVRegI(env);
1697 HReg r_ccIR_b6 = newVRegI(env);
1698
1699 if (e->Iex.Binop.op == Iop_CmpF64) {
1700 fr_srcL = iselDblExpr(env, e->Iex.Binop.arg1, IEndianess);
1701 fr_srcR = iselDblExpr(env, e->Iex.Binop.arg2, IEndianess);
1702 addInstr(env, PPCInstr_FpCmp(r_ccPPC, fr_srcL, fr_srcR));
1703
1704 } else if (e->Iex.Binop.op == Iop_CmpD64) {
1705 fr_srcL = iselDfp64Expr(env, e->Iex.Binop.arg1, IEndianess);
1706 fr_srcR = iselDfp64Expr(env, e->Iex.Binop.arg2, IEndianess);
1707 addInstr(env, PPCInstr_Dfp64Cmp(r_ccPPC, fr_srcL, fr_srcR));
1708
1709 } else { // e->Iex.Binop.op == Iop_CmpD128
1710 iselDfp128Expr(&fr_srcL, &fr_srcL_lo, env, e->Iex.Binop.arg1,
1711 IEndianess);
1712 iselDfp128Expr(&fr_srcR, &fr_srcR_lo, env, e->Iex.Binop.arg2,
1713 IEndianess);
1714 addInstr(env, PPCInstr_Dfp128Cmp(r_ccPPC, fr_srcL, fr_srcL_lo,
1715 fr_srcR, fr_srcR_lo));
1716 }
1717
1718 /* Map compare result from PPC to IR,
1719 conforming to CmpF64 definition. */
1720 /*
1721 FP cmp result | PPC | IR
1722 --------------------------
1723 UN | 0x1 | 0x45
1724 EQ | 0x2 | 0x40
1725 GT | 0x4 | 0x00
1726 LT | 0x8 | 0x01
1727 */
1728
1729 // r_ccIR_b0 = r_ccPPC[0] | r_ccPPC[3]
1730 addInstr(env, PPCInstr_Shft(Pshft_SHR, True/*32bit shift*/,
1731 r_ccIR_b0, r_ccPPC,
1732 PPCRH_Imm(False,0x3)));
1733 addInstr(env, PPCInstr_Alu(Palu_OR, r_ccIR_b0,
1734 r_ccPPC, PPCRH_Reg(r_ccIR_b0)));
1735 addInstr(env, PPCInstr_Alu(Palu_AND, r_ccIR_b0,
1736 r_ccIR_b0, PPCRH_Imm(False,0x1)));
1737
1738 // r_ccIR_b2 = r_ccPPC[0]
1739 addInstr(env, PPCInstr_Shft(Pshft_SHL, True/*32bit shift*/,
1740 r_ccIR_b2, r_ccPPC,
1741 PPCRH_Imm(False,0x2)));
1742 addInstr(env, PPCInstr_Alu(Palu_AND, r_ccIR_b2,
1743 r_ccIR_b2, PPCRH_Imm(False,0x4)));
1744
1745 // r_ccIR_b6 = r_ccPPC[0] | r_ccPPC[1]
1746 addInstr(env, PPCInstr_Shft(Pshft_SHR, True/*32bit shift*/,
1747 r_ccIR_b6, r_ccPPC,
1748 PPCRH_Imm(False,0x1)));
1749 addInstr(env, PPCInstr_Alu(Palu_OR, r_ccIR_b6,
1750 r_ccPPC, PPCRH_Reg(r_ccIR_b6)));
1751 addInstr(env, PPCInstr_Shft(Pshft_SHL, True/*32bit shift*/,
1752 r_ccIR_b6, r_ccIR_b6,
1753 PPCRH_Imm(False,0x6)));
1754 addInstr(env, PPCInstr_Alu(Palu_AND, r_ccIR_b6,
1755 r_ccIR_b6, PPCRH_Imm(False,0x40)));
1756
1757 // r_ccIR = r_ccIR_b0 | r_ccIR_b2 | r_ccIR_b6
1758 addInstr(env, PPCInstr_Alu(Palu_OR, r_ccIR,
1759 r_ccIR_b0, PPCRH_Reg(r_ccIR_b2)));
1760 addInstr(env, PPCInstr_Alu(Palu_OR, r_ccIR,
1761 r_ccIR, PPCRH_Reg(r_ccIR_b6)));
1762 return r_ccIR;
1763 }
1764
1765 if ( e->Iex.Binop.op == Iop_F64toI32S ||
1766 e->Iex.Binop.op == Iop_F64toI32U ) {
1767 /* This works in both mode64 and mode32. */
1768 HReg r1 = StackFramePtr(env->mode64);
1769 PPCAMode* zero_r1 = PPCAMode_IR( 0, r1 );
1770 HReg fsrc = iselDblExpr(env, e->Iex.Binop.arg2, IEndianess);
1771 HReg ftmp = newVRegF(env);
1772 HReg idst = newVRegI(env);
1773
1774 /* Set host rounding mode */
1775 set_FPU_rounding_mode( env, e->Iex.Binop.arg1, IEndianess );
1776
1777 sub_from_sp( env, 16 );
1778 addInstr(env, PPCInstr_FpCftI(False/*F->I*/, True/*int32*/,
1779 e->Iex.Binop.op == Iop_F64toI32S ? True/*syned*/
1780 : False,
1781 True/*flt64*/,
1782 ftmp, fsrc));
1783 addInstr(env, PPCInstr_FpSTFIW(r1, ftmp));
1784 addInstr(env, PPCInstr_Load(4, idst, zero_r1, mode64));
1785
1786 /* in 64-bit mode we need to sign-widen idst. */
1787 if (mode64)
1788 addInstr(env, PPCInstr_Unary(Pun_EXTSW, idst, idst));
1789
1790 add_to_sp( env, 16 );
1791
1792 ///* Restore default FPU rounding. */
1793 //set_FPU_rounding_default( env );
1794 return idst;
1795 }
1796
1797 if (e->Iex.Binop.op == Iop_F64toI64S || e->Iex.Binop.op == Iop_F64toI64U ) {
1798 if (mode64) {
1799 HReg r1 = StackFramePtr(env->mode64);
1800 PPCAMode* zero_r1 = PPCAMode_IR( 0, r1 );
1801 HReg fsrc = iselDblExpr(env, e->Iex.Binop.arg2,
1802 IEndianess);
1803 HReg idst = newVRegI(env);
1804 HReg ftmp = newVRegF(env);
1805
1806 /* Set host rounding mode */
1807 set_FPU_rounding_mode( env, e->Iex.Binop.arg1, IEndianess );
1808
1809 sub_from_sp( env, 16 );
1810 addInstr(env, PPCInstr_FpCftI(False/*F->I*/, False/*int64*/,
1811 ( e->Iex.Binop.op == Iop_F64toI64S ) ? True
1812 : False,
1813 True, ftmp, fsrc));
1814 addInstr(env, PPCInstr_FpLdSt(False/*store*/, 8, ftmp, zero_r1));
1815 addInstr(env, PPCInstr_Load(8, idst, zero_r1, True/*mode64*/));
1816 add_to_sp( env, 16 );
1817
1818 ///* Restore default FPU rounding. */
1819 //set_FPU_rounding_default( env );
1820 return idst;
1821 }
1822 }
1823
1824 if (e->Iex.Binop.op == Iop_D64toI64S ) {
1825 HReg r1 = StackFramePtr(env->mode64);
1826 PPCAMode* zero_r1 = PPCAMode_IR( 0, r1 );
1827 HReg fr_src = iselDfp64Expr(env, e->Iex.Binop.arg2, IEndianess);
1828 HReg idst = newVRegI(env);
1829 HReg ftmp = newVRegF(env);
1830
1831 /* Set host rounding mode */
1832 set_FPU_DFP_rounding_mode( env, e->Iex.Binop.arg1, IEndianess );
1833 addInstr(env, PPCInstr_Dfp64Unary(Pfp_DCTFIX, ftmp, fr_src));
1834 sub_from_sp( env, 16 );
1835 addInstr(env, PPCInstr_FpLdSt(False/*store*/, 8, ftmp, zero_r1));
1836 addInstr(env, PPCInstr_Load(8, idst, zero_r1, mode64));
1837
1838 add_to_sp( env, 16 );
1839
1840 ///* Restore default FPU rounding. */
1841 //set_FPU_rounding_default( env );
1842 return idst;
1843 }
1844
1845 if (e->Iex.Binop.op == Iop_D128toI64S ) {
1846 PPCFpOp fpop = Pfp_DCTFIXQ;
1847 HReg r_srcHi = newVRegF(env);
1848 HReg r_srcLo = newVRegF(env);
1849 HReg idst = newVRegI(env);
1850 HReg ftmp = newVRegF(env);
1851 PPCAMode* zero_r1 = PPCAMode_IR( 0, StackFramePtr(env->mode64) );
1852
1853 set_FPU_DFP_rounding_mode( env, e->Iex.Binop.arg1, IEndianess );
1854 iselDfp128Expr(&r_srcHi, &r_srcLo, env, e->Iex.Binop.arg2,
1855 IEndianess);
1856 addInstr(env, PPCInstr_DfpD128toD64(fpop, ftmp, r_srcHi, r_srcLo));
1857
1858 // put the D64 result into an integer register
1859 sub_from_sp( env, 16 );
1860 addInstr(env, PPCInstr_FpLdSt(False/*store*/, 8, ftmp, zero_r1));
1861 addInstr(env, PPCInstr_Load(8, idst, zero_r1, True/*mode64*/));
1862 add_to_sp( env, 16 );
1863 return idst;
1864 }
1865 break;
1866 }
1867
1868 /* --------- UNARY OP --------- */
1869 case Iex_Unop: {
1870 IROp op_unop = e->Iex.Unop.op;
1871
1872 /* 1Uto8(32to1(expr32)) */
1873 DEFINE_PATTERN(p_32to1_then_1Uto8,
1874 unop(Iop_1Uto8,unop(Iop_32to1,bind(0))));
1875 if (matchIRExpr(&mi,p_32to1_then_1Uto8,e)) {
1876 const IRExpr* expr32 = mi.bindee[0];
1877 HReg r_dst = newVRegI(env);
1878 HReg r_src = iselWordExpr_R(env, expr32, IEndianess);
1879 addInstr(env, PPCInstr_Alu(Palu_AND, r_dst,
1880 r_src, PPCRH_Imm(False,1)));
1881 return r_dst;
1882 }
1883
1884 /* 16Uto32(LDbe:I16(expr32)) */
1885 {
1886 DECLARE_PATTERN(p_LDbe16_then_16Uto32);
1887 DEFINE_PATTERN(p_LDbe16_then_16Uto32,
1888 unop(Iop_16Uto32,
1889 IRExpr_Load(IEndianess,Ity_I16,bind(0))) );
1890 if (matchIRExpr(&mi,p_LDbe16_then_16Uto32,e)) {
1891 HReg r_dst = newVRegI(env);
1892 PPCAMode* amode
1893 = iselWordExpr_AMode( env, mi.bindee[0], Ity_I16/*xfer*/,
1894 IEndianess );
1895 addInstr(env, PPCInstr_Load(2,r_dst,amode, mode64));
1896 return r_dst;
1897 }
1898 }
1899
1900 switch (op_unop) {
1901 case Iop_8Uto16:
1902 case Iop_8Uto32:
1903 case Iop_8Uto64:
1904 case Iop_16Uto32:
1905 case Iop_16Uto64: {
1906 HReg r_dst = newVRegI(env);
1907 HReg r_src = iselWordExpr_R(env, e->Iex.Unop.arg, IEndianess);
1908 UShort mask = toUShort(op_unop==Iop_16Uto64 ? 0xFFFF :
1909 op_unop==Iop_16Uto32 ? 0xFFFF : 0xFF);
1910 addInstr(env, PPCInstr_Alu(Palu_AND,r_dst,r_src,
1911 PPCRH_Imm(False,mask)));
1912 return r_dst;
1913 }
1914 case Iop_32Uto64: {
1915 HReg r_dst = newVRegI(env);
1916 HReg r_src = iselWordExpr_R(env, e->Iex.Unop.arg, IEndianess);
1917 vassert(mode64);
1918 addInstr(env,
1919 PPCInstr_Shft(Pshft_SHL, False/*64bit shift*/,
1920 r_dst, r_src, PPCRH_Imm(False,32)));
1921 addInstr(env,
1922 PPCInstr_Shft(Pshft_SHR, False/*64bit shift*/,
1923 r_dst, r_dst, PPCRH_Imm(False,32)));
1924 return r_dst;
1925 }
1926 case Iop_8Sto16:
1927 case Iop_8Sto32:
1928 case Iop_16Sto32: {
1929 HReg r_dst = newVRegI(env);
1930 HReg r_src = iselWordExpr_R(env, e->Iex.Unop.arg, IEndianess);
1931 UShort amt = toUShort(op_unop==Iop_16Sto32 ? 16 : 24);
1932 addInstr(env,
1933 PPCInstr_Shft(Pshft_SHL, True/*32bit shift*/,
1934 r_dst, r_src, PPCRH_Imm(False,amt)));
1935 addInstr(env,
1936 PPCInstr_Shft(Pshft_SAR, True/*32bit shift*/,
1937 r_dst, r_dst, PPCRH_Imm(False,amt)));
1938 return r_dst;
1939 }
1940 case Iop_8Sto64:
1941 case Iop_16Sto64: {
1942 HReg r_dst = newVRegI(env);
1943 HReg r_src = iselWordExpr_R(env, e->Iex.Unop.arg, IEndianess);
1944 UShort amt = toUShort(op_unop==Iop_8Sto64 ? 56 : 48);
1945 vassert(mode64);
1946 addInstr(env,
1947 PPCInstr_Shft(Pshft_SHL, False/*64bit shift*/,
1948 r_dst, r_src, PPCRH_Imm(False,amt)));
1949 addInstr(env,
1950 PPCInstr_Shft(Pshft_SAR, False/*64bit shift*/,
1951 r_dst, r_dst, PPCRH_Imm(False,amt)));
1952 return r_dst;
1953 }
1954 case Iop_32Sto64: {
1955 HReg r_dst = newVRegI(env);
1956 HReg r_src = iselWordExpr_R(env, e->Iex.Unop.arg, IEndianess);
1957 vassert(mode64);
1958 /* According to the IBM docs, in 64 bit mode, srawi r,r,0
1959 sign extends the lower 32 bits into the upper 32 bits. */
1960 addInstr(env,
1961 PPCInstr_Shft(Pshft_SAR, True/*32bit shift*/,
1962 r_dst, r_src, PPCRH_Imm(False,0)));
1963 return r_dst;
1964 }
1965 case Iop_Not8:
1966 case Iop_Not16:
1967 case Iop_Not32:
1968 case Iop_Not64: {
1969 if (op_unop == Iop_Not64) vassert(mode64);
1970 HReg r_dst = newVRegI(env);
1971 HReg r_src = iselWordExpr_R(env, e->Iex.Unop.arg, IEndianess);
1972 addInstr(env, PPCInstr_Unary(Pun_NOT,r_dst,r_src));
1973 return r_dst;
1974 }
1975 case Iop_64HIto32: {
1976 if (!mode64) {
1977 HReg rHi, rLo;
1978 iselInt64Expr(&rHi,&rLo, env, e->Iex.Unop.arg, IEndianess);
1979 return rHi; /* and abandon rLo .. poor wee thing :-) */
1980 } else {
1981 HReg r_dst = newVRegI(env);
1982 HReg r_src = iselWordExpr_R(env, e->Iex.Unop.arg, IEndianess);
1983 addInstr(env,
1984 PPCInstr_Shft(Pshft_SHR, False/*64bit shift*/,
1985 r_dst, r_src, PPCRH_Imm(False,32)));
1986 return r_dst;
1987 }
1988 }
1989 case Iop_64to32: {
1990 if (!mode64) {
1991 HReg rHi, rLo;
1992 iselInt64Expr(&rHi,&rLo, env, e->Iex.Unop.arg, IEndianess);
1993 return rLo; /* similar stupid comment to the above ... */
1994 } else {
1995 /* This is a no-op. */
1996 return iselWordExpr_R(env, e->Iex.Unop.arg, IEndianess);
1997 }
1998 }
1999 case Iop_64to16: {
2000 if (mode64) { /* This is a no-op. */
2001 return iselWordExpr_R(env, e->Iex.Unop.arg, IEndianess);
2002 }
2003 break; /* evidently not used in 32-bit mode */
2004 }
2005 case Iop_16HIto8:
2006 case Iop_32HIto16: {
2007 HReg r_dst = newVRegI(env);
2008 HReg r_src = iselWordExpr_R(env, e->Iex.Unop.arg, IEndianess);
2009 UShort shift = toUShort(op_unop == Iop_16HIto8 ? 8 : 16);
2010 addInstr(env,
2011 PPCInstr_Shft(Pshft_SHR, True/*32bit shift*/,
2012 r_dst, r_src, PPCRH_Imm(False,shift)));
2013 return r_dst;
2014 }
2015 case Iop_128HIto64:
2016 if (mode64) {
2017 HReg rHi, rLo;
2018 iselInt128Expr(&rHi,&rLo, env, e->Iex.Unop.arg, IEndianess);
2019 return rHi; /* and abandon rLo .. poor wee thing :-) */
2020 }
2021 break;
2022 case Iop_128to64:
2023 if (mode64) {
2024 HReg rHi, rLo;
2025 iselInt128Expr(&rHi,&rLo, env, e->Iex.Unop.arg, IEndianess);
2026 return rLo; /* similar stupid comment to the above ... */
2027 }
2028 break;
2029 case Iop_1Uto64:
2030 case Iop_1Uto32:
2031 case Iop_1Uto8:
2032 if ((op_unop != Iop_1Uto64) || mode64) {
2033 HReg r_dst = newVRegI(env);
2034 PPCCondCode cond = iselCondCode(env, e->Iex.Unop.arg, IEndianess);
2035 addInstr(env, PPCInstr_Set(cond,r_dst));
2036 return r_dst;
2037 }
2038 break;
2039 case Iop_1Sto8:
2040 case Iop_1Sto16:
2041 case Iop_1Sto32: {
2042 /* could do better than this, but for now ... */
2043 HReg r_dst = newVRegI(env);
2044 PPCCondCode cond = iselCondCode(env, e->Iex.Unop.arg, IEndianess);
2045 addInstr(env, PPCInstr_Set(cond,r_dst));
2046 addInstr(env,
2047 PPCInstr_Shft(Pshft_SHL, True/*32bit shift*/,
2048 r_dst, r_dst, PPCRH_Imm(False,31)));
2049 addInstr(env,
2050 PPCInstr_Shft(Pshft_SAR, True/*32bit shift*/,
2051 r_dst, r_dst, PPCRH_Imm(False,31)));
2052 return r_dst;
2053 }
2054 case Iop_1Sto64:
2055 if (mode64) {
2056 /* could do better than this, but for now ... */
2057 HReg r_dst = newVRegI(env);
2058 PPCCondCode cond = iselCondCode(env, e->Iex.Unop.arg, IEndianess);
2059 addInstr(env, PPCInstr_Set(cond,r_dst));
2060 addInstr(env, PPCInstr_Shft(Pshft_SHL, False/*64bit shift*/,
2061 r_dst, r_dst, PPCRH_Imm(False,63)));
2062 addInstr(env, PPCInstr_Shft(Pshft_SAR, False/*64bit shift*/,
2063 r_dst, r_dst, PPCRH_Imm(False,63)));
2064 return r_dst;
2065 }
2066 break;
2067
2068 case Iop_Clz32: case Iop_ClzNat32:
2069 case Iop_Clz64: case Iop_ClzNat64: {
2070 // cntlz is available even in the most basic (earliest) ppc
2071 // variants, so it's safe to generate it unconditionally.
2072 HReg r_src, r_dst;
2073 PPCUnaryOp op_clz = (op_unop == Iop_Clz32 || op_unop == Iop_ClzNat32)
2074 ? Pun_CLZ32 : Pun_CLZ64;
2075 if ((op_unop == Iop_Clz64 || op_unop == Iop_ClzNat64) && !mode64)
2076 goto irreducible;
2077 /* Count leading zeroes. */
2078 r_dst = newVRegI(env);
2079 r_src = iselWordExpr_R(env, e->Iex.Unop.arg, IEndianess);
2080 addInstr(env, PPCInstr_Unary(op_clz,r_dst,r_src));
2081 return r_dst;
2082 }
2083
2084 //case Iop_Ctz32:
2085 case Iop_CtzNat32:
2086 //case Iop_Ctz64:
2087 case Iop_CtzNat64:
2088 {
2089 // Generate code using Clz, because we can't assume the host has
2090 // Ctz. In particular, part of the fix for bug 386945 involves
2091 // creating a Ctz in ir_opt.c from smaller fragments.
2092 PPCUnaryOp op_clz = Pun_CLZ64;
2093 Int WS = 64;
2094 if (op_unop == Iop_Ctz32 || op_unop == Iop_CtzNat32) {
2095 op_clz = Pun_CLZ32;
2096 WS = 32;
2097 }
2098 /* Compute ctz(arg) = wordsize - clz(~arg & (arg - 1)), thusly:
2099 t1 = arg - 1
2100 t2 = not arg
2101 t2 = t2 & t1
2102 t2 = clz t2
2103 t1 = WS
2104 t2 = t1 - t2
2105 // result in t2
2106 */
2107 HReg arg = iselWordExpr_R(env, e->Iex.Unop.arg, IEndianess);
2108 HReg t1 = newVRegI(env);
2109 HReg t2 = newVRegI(env);
2110 addInstr(env, PPCInstr_Alu(Palu_SUB, t1, arg, PPCRH_Imm(True, 1)));
2111 addInstr(env, PPCInstr_Unary(Pun_NOT, t2, arg));
2112 addInstr(env, PPCInstr_Alu(Palu_AND, t2, t2, PPCRH_Reg(t1)));
2113 addInstr(env, PPCInstr_Unary(op_clz, t2, t2));
2114 addInstr(env, PPCInstr_LI(t1, WS, False/*!64-bit imm*/));
2115 addInstr(env, PPCInstr_Alu(Palu_SUB, t2, t1, PPCRH_Reg(t2)));
2116 return t2;
2117 }
2118
2119 case Iop_PopCount64: {
2120 // popcnt{x,d} is only available in later arch revs (ISA 3.0,
2121 // maybe) so it's not really correct to emit it here without a caps
2122 // check for the host.
2123 if (mode64) {
2124 HReg r_dst = newVRegI(env);
2125 HReg r_src = iselWordExpr_R(env, e->Iex.Unop.arg, IEndianess);
2126 addInstr(env, PPCInstr_Unary(Pun_POP64, r_dst, r_src));
2127 return r_dst;
2128 }
2129 // We don't expect to be required to handle this in 32-bit mode.
2130 break;
2131 }
2132
2133 case Iop_PopCount32: {
2134 // Similar comment as for Ctz just above applies -- we really
2135 // should have a caps check here.
2136
2137 HReg r_dst = newVRegI(env);
2138 // This actually generates popcntw, which in 64 bit mode does a
2139 // 32-bit count individually for both low and high halves of the
2140 // word. Per the comment at the top of iselIntExpr_R, in the 64
2141 // bit mode case, the user of this result is required to ignore
2142 // the upper 32 bits of the result. In 32 bit mode this is all
2143 // moot. It is however unclear from the PowerISA 3.0 docs that
2144 // the instruction exists in 32 bit mode; however our own front
2145 // end (guest_ppc_toIR.c) accepts it, so I guess it does exist.
2146 HReg r_src = iselWordExpr_R(env, e->Iex.Unop.arg, IEndianess);
2147 addInstr(env, PPCInstr_Unary(Pun_POP32, r_dst, r_src));
2148 return r_dst;
2149 }
2150
2151 case Iop_Reverse8sIn32_x1: {
2152 // A bit of a mouthful, but simply .. 32-bit byte swap.
2153 // This is pretty rubbish code. We could do vastly better if
2154 // rotates, and better, rotate-inserts, were allowed. Note that
2155 // even on a 64 bit target, the right shifts must be done as 32-bit
2156 // so as to introduce zero bits in the right places. So it seems
2157 // simplest to do the whole sequence in 32-bit insns.
2158 /*
2159 r = <argument> // working temporary, initial byte order ABCD
2160 Mask = 00FF00FF
2161 nMask = not Mask
2162 tHi = and r, Mask
2163 tHi = shl tHi, 8
2164 tLo = and r, nMask
2165 tLo = shr tLo, 8
2166 r = or tHi, tLo // now r has order BADC
2167 and repeat for 16 bit chunks ..
2168 Mask = 0000FFFF
2169 nMask = not Mask
2170 tHi = and r, Mask
2171 tHi = shl tHi, 16
2172 tLo = and r, nMask
2173 tLo = shr tLo, 16
2174 r = or tHi, tLo // now r has order DCBA
2175 */
2176 HReg r_src = iselWordExpr_R(env, e->Iex.Unop.arg, IEndianess);
2177 HReg rr = newVRegI(env);
2178 HReg rMask = newVRegI(env);
2179 HReg rnMask = newVRegI(env);
2180 HReg rtHi = newVRegI(env);
2181 HReg rtLo = newVRegI(env);
2182 // Copy r_src since we need to modify it
2183 addInstr(env, mk_iMOVds_RR(rr, r_src));
2184 // Swap within 16-bit lanes
2185 addInstr(env, PPCInstr_LI(rMask, 0x00FF00FFULL,
2186 False/* !64bit imm*/));
2187 addInstr(env, PPCInstr_Unary(Pun_NOT, rnMask, rMask));
2188 addInstr(env, PPCInstr_Alu(Palu_AND, rtHi, rr, PPCRH_Reg(rMask)));
2189 addInstr(env, PPCInstr_Shft(Pshft_SHL, True/*32 bit shift*/,
2190 rtHi, rtHi,
2191 PPCRH_Imm(False/*!signed imm*/, 8)));
2192 addInstr(env, PPCInstr_Alu(Palu_AND, rtLo, rr, PPCRH_Reg(rnMask)));
2193 addInstr(env, PPCInstr_Shft(Pshft_SHR, True/*32 bit shift*/,
2194 rtLo, rtLo,
2195 PPCRH_Imm(False/*!signed imm*/, 8)));
2196 addInstr(env, PPCInstr_Alu(Palu_OR, rr, rtHi, PPCRH_Reg(rtLo)));
2197 // And now swap the two 16-bit chunks
2198 addInstr(env, PPCInstr_LI(rMask, 0x0000FFFFULL,
2199 False/* !64bit imm*/));
2200 addInstr(env, PPCInstr_Unary(Pun_NOT, rnMask, rMask));
2201 addInstr(env, PPCInstr_Alu(Palu_AND, rtHi, rr, PPCRH_Reg(rMask)));
2202 addInstr(env, PPCInstr_Shft(Pshft_SHL, True/*32 bit shift*/,
2203 rtHi, rtHi,
2204 PPCRH_Imm(False/*!signed imm*/, 16)));
2205 addInstr(env, PPCInstr_Alu(Palu_AND, rtLo, rr, PPCRH_Reg(rnMask)));
2206 addInstr(env, PPCInstr_Shft(Pshft_SHR, True/*32 bit shift*/,
2207 rtLo, rtLo,
2208 PPCRH_Imm(False/*!signed imm*/, 16)));
2209 addInstr(env, PPCInstr_Alu(Palu_OR, rr, rtHi, PPCRH_Reg(rtLo)));
2210 return rr;
2211 }
2212
2213 case Iop_Reverse8sIn64_x1: {
2214 /* See Iop_Reverse8sIn32_x1, but extended to 64bit.
2215 Can only be used in 64bit mode. */
2216 vassert (mode64);
2217
2218 HReg r_src = iselWordExpr_R(env, e->Iex.Unop.arg, IEndianess);
2219 HReg rr = newVRegI(env);
2220 HReg rMask = newVRegI(env);
2221 HReg rnMask = newVRegI(env);
2222 HReg rtHi = newVRegI(env);
2223 HReg rtLo = newVRegI(env);
2224
2225 // Copy r_src since we need to modify it
2226 addInstr(env, mk_iMOVds_RR(rr, r_src));
2227
2228 // r = (r & 0x00FF00FF00FF00FF) << 8 | (r & 0xFF00FF00FF00FF00) >> 8
2229 addInstr(env, PPCInstr_LI(rMask, 0x00FF00FF00FF00FFULL,
2230 True/* 64bit imm*/));
2231 addInstr(env, PPCInstr_Unary(Pun_NOT, rnMask, rMask));
2232 addInstr(env, PPCInstr_Alu(Palu_AND, rtHi, rr, PPCRH_Reg(rMask)));
2233 addInstr(env, PPCInstr_Shft(Pshft_SHL, False/*64 bit shift*/,
2234 rtHi, rtHi,
2235 PPCRH_Imm(False/*!signed imm*/, 8)));
2236 addInstr(env, PPCInstr_Alu(Palu_AND, rtLo, rr, PPCRH_Reg(rnMask)));
2237 addInstr(env, PPCInstr_Shft(Pshft_SHR, False/*64 bit shift*/,
2238 rtLo, rtLo,
2239 PPCRH_Imm(False/*!signed imm*/, 8)));
2240 addInstr(env, PPCInstr_Alu(Palu_OR, rr, rtHi, PPCRH_Reg(rtLo)));
2241
2242 // r = (r & 0x0000FFFF0000FFFF) << 16 | (r & 0xFFFF0000FFFF0000) >> 16
2243 addInstr(env, PPCInstr_LI(rMask, 0x0000FFFF0000FFFFULL,
2244 True/* !64bit imm*/));
2245 addInstr(env, PPCInstr_Unary(Pun_NOT, rnMask, rMask));
2246 addInstr(env, PPCInstr_Alu(Palu_AND, rtHi, rr, PPCRH_Reg(rMask)));
2247 addInstr(env, PPCInstr_Shft(Pshft_SHL, False/*64 bit shift*/,
2248 rtHi, rtHi,
2249 PPCRH_Imm(False/*!signed imm*/, 16)));
2250 addInstr(env, PPCInstr_Alu(Palu_AND, rtLo, rr, PPCRH_Reg(rnMask)));
2251 addInstr(env, PPCInstr_Shft(Pshft_SHR, False/*64 bit shift*/,
2252 rtLo, rtLo,
2253 PPCRH_Imm(False/*!signed imm*/, 16)));
2254 addInstr(env, PPCInstr_Alu(Palu_OR, rr, rtHi, PPCRH_Reg(rtLo)));
2255
2256 // r = (r & 0x00000000FFFFFFFF) << 32 | (r & 0xFFFFFFFF00000000) >> 32
2257 /* We don't need to mask anymore, just two more shifts and an or. */
2258 addInstr(env, mk_iMOVds_RR(rtLo, rr));
2259 addInstr(env, PPCInstr_Shft(Pshft_SHL, False/*64 bit shift*/,
2260 rtLo, rtLo,
2261 PPCRH_Imm(False/*!signed imm*/, 32)));
2262 addInstr(env, PPCInstr_Shft(Pshft_SHR, False/*64 bit shift*/,
2263 rr, rr,
2264 PPCRH_Imm(False/*!signed imm*/, 32)));
2265 addInstr(env, PPCInstr_Alu(Palu_OR, rr, rr, PPCRH_Reg(rtLo)));
2266
2267 return rr;
2268 }
2269
2270 case Iop_Left8:
2271 case Iop_Left16:
2272 case Iop_Left32:
2273 case Iop_Left64: {
2274 HReg r_src, r_dst;
2275 if (op_unop == Iop_Left64 && !mode64)
2276 goto irreducible;
2277 r_dst = newVRegI(env);
2278 r_src = iselWordExpr_R(env, e->Iex.Unop.arg, IEndianess);
2279 addInstr(env, PPCInstr_Unary(Pun_NEG,r_dst,r_src));
2280 addInstr(env, PPCInstr_Alu(Palu_OR, r_dst, r_dst, PPCRH_Reg(r_src)));
2281 return r_dst;
2282 }
2283
2284 case Iop_CmpwNEZ32: {
2285 HReg r_dst = newVRegI(env);
2286 HReg r_src = iselWordExpr_R(env, e->Iex.Unop.arg, IEndianess);
2287 addInstr(env, PPCInstr_Unary(Pun_NEG,r_dst,r_src));
2288 addInstr(env, PPCInstr_Alu(Palu_OR, r_dst, r_dst, PPCRH_Reg(r_src)));
2289 addInstr(env, PPCInstr_Shft(Pshft_SAR, True/*32bit shift*/,
2290 r_dst, r_dst, PPCRH_Imm(False, 31)));
2291 return r_dst;
2292 }
2293
2294 case Iop_CmpwNEZ64: {
2295 HReg r_dst = newVRegI(env);
2296 HReg r_src = iselWordExpr_R(env, e->Iex.Unop.arg, IEndianess);
2297 if (!mode64) goto irreducible;
2298 addInstr(env, PPCInstr_Unary(Pun_NEG,r_dst,r_src));
2299 addInstr(env, PPCInstr_Alu(Palu_OR, r_dst, r_dst, PPCRH_Reg(r_src)));
2300 addInstr(env, PPCInstr_Shft(Pshft_SAR, False/*64bit shift*/,
2301 r_dst, r_dst, PPCRH_Imm(False, 63)));
2302 return r_dst;
2303 }
2304
2305 case Iop_V128to32: {
2306 HReg r_aligned16;
2307 HReg dst = newVRegI(env);
2308 HReg vec = iselVecExpr(env, e->Iex.Unop.arg, IEndianess);
2309 PPCAMode *am_off0, *am_off_word0;
2310 sub_from_sp( env, 32 ); // Move SP down 32 bytes
2311
2312 // get a quadword aligned address within our stack space
2313 r_aligned16 = get_sp_aligned16( env );
2314 am_off0 = PPCAMode_IR( 0, r_aligned16 );
2315
2316 /* Note that the store below (done via PPCInstr_AvLdSt) uses
2317 * stvx, which stores the vector in proper LE format,
2318 * with byte zero (far right byte of the register in LE format)
2319 * stored at the lowest memory address. Therefore, to obtain
2320 * integer word zero, we need to use that lowest memory address
2321 * as the base for the load.
2322 */
2323 if (IEndianess == Iend_LE)
2324 am_off_word0 = am_off0;
2325 else
2326 am_off_word0 = PPCAMode_IR( 12,r_aligned16 );
2327
2328 // store vec, load low word to dst
2329 addInstr(env,
2330 PPCInstr_AvLdSt( False/*store*/, 16, vec, am_off0 ));
2331 addInstr(env,
2332 PPCInstr_Load( 4, dst, am_off_word0, mode64 ));
2333
2334 add_to_sp( env, 32 ); // Reset SP
2335 return dst;
2336 }
2337
2338 case Iop_V128to64:
2339 case Iop_V128HIto64:
2340 if (mode64) {
2341 HReg r_aligned16;
2342 HReg dst = newVRegI(env);
2343 HReg vec = iselVecExpr(env, e->Iex.Unop.arg, IEndianess);
2344 PPCAMode *am_off0, *am_off8, *am_off_arg;
2345 sub_from_sp( env, 32 ); // Move SP down 32 bytes
2346
2347 // get a quadword aligned address within our stack space
2348 r_aligned16 = get_sp_aligned16( env );
2349 am_off0 = PPCAMode_IR( 0, r_aligned16 );
2350 am_off8 = PPCAMode_IR( 8 ,r_aligned16 );
2351
2352 // store vec, load low word or high to dst
2353 addInstr(env,
2354 PPCInstr_AvLdSt( False/*store*/, 16, vec, am_off0 ));
2355 if (IEndianess == Iend_LE) {
2356 if (op_unop == Iop_V128HIto64)
2357 am_off_arg = am_off8;
2358 else
2359 am_off_arg = am_off0;
2360 } else {
2361 if (op_unop == Iop_V128HIto64)
2362 am_off_arg = am_off0;
2363 else
2364 am_off_arg = am_off8;
2365 }
2366 addInstr(env,
2367 PPCInstr_Load(
2368 8, dst,
2369 am_off_arg,
2370 mode64 ));
2371
2372 add_to_sp( env, 32 ); // Reset SP
2373 return dst;
2374 }
2375 break;
2376 case Iop_16to8:
2377 case Iop_32to8:
2378 case Iop_32to16:
2379 case Iop_64to8:
2380 /* These are no-ops. */
2381 return iselWordExpr_R(env, e->Iex.Unop.arg, IEndianess);
2382
2383 /* ReinterpF64asI64(e) */
2384 /* Given an IEEE754 double, produce an I64 with the same bit
2385 pattern. */
2386 case Iop_ReinterpF64asI64:
2387 if (mode64) {
2388 PPCAMode *am_addr;
2389 HReg fr_src = iselDblExpr(env, e->Iex.Unop.arg, IEndianess);
2390 HReg r_dst = newVRegI(env);
2391
2392 sub_from_sp( env, 16 ); // Move SP down 16 bytes
2393 am_addr = PPCAMode_IR( 0, StackFramePtr(mode64) );
2394
2395 // store as F64
2396 addInstr(env, PPCInstr_FpLdSt( False/*store*/, 8,
2397 fr_src, am_addr ));
2398 // load as Ity_I64
2399 addInstr(env, PPCInstr_Load( 8, r_dst, am_addr, mode64 ));
2400
2401 add_to_sp( env, 16 ); // Reset SP
2402 return r_dst;
2403 }
2404 break;
2405
2406 /* ReinterpF32asI32(e) */
2407 /* Given an IEEE754 float, produce an I32 with the same bit
2408 pattern. */
2409 case Iop_ReinterpF32asI32: {
2410 /* I believe this generates correct code for both 32- and
2411 64-bit hosts. */
2412 PPCAMode *am_addr;
2413 HReg fr_src = iselFltExpr(env, e->Iex.Unop.arg, IEndianess);
2414 HReg r_dst = newVRegI(env);
2415
2416 sub_from_sp( env, 16 ); // Move SP down 16 bytes
2417 am_addr = PPCAMode_IR( 0, StackFramePtr(mode64) );
2418
2419 // store as F32
2420 addInstr(env, PPCInstr_FpLdSt( False/*store*/, 4,
2421 fr_src, am_addr ));
2422 // load as Ity_I32
2423 addInstr(env, PPCInstr_Load( 4, r_dst, am_addr, mode64 ));
2424
2425 add_to_sp( env, 16 ); // Reset SP
2426 return r_dst;
2427 }
2428 break;
2429
2430 case Iop_ReinterpD64asI64:
2431 if (mode64) {
2432 PPCAMode *am_addr;
2433 HReg fr_src = iselDfp64Expr(env, e->Iex.Unop.arg, IEndianess);
2434 HReg r_dst = newVRegI(env);
2435
2436 sub_from_sp( env, 16 ); // Move SP down 16 bytes
2437 am_addr = PPCAMode_IR( 0, StackFramePtr(mode64) );
2438
2439 // store as D64
2440 addInstr(env, PPCInstr_FpLdSt( False/*store*/, 8,
2441 fr_src, am_addr ));
2442 // load as Ity_I64
2443 addInstr(env, PPCInstr_Load( 8, r_dst, am_addr, mode64 ));
2444 add_to_sp( env, 16 ); // Reset SP
2445 return r_dst;
2446 }
2447 break;
2448
2449 case Iop_BCDtoDPB: {
2450 /* the following is only valid in 64 bit mode */
2451 if (!mode64) break;
2452
2453 PPCCondCode cc;
2454 UInt argiregs;
2455 HReg argregs[1];
2456 HReg r_dst = newVRegI(env);
2457 Int argreg;
2458
2459 argiregs = 0;
2460 argreg = 0;
2461 argregs[0] = hregPPC_GPR3(mode64);
2462
2463 argiregs |= (1 << (argreg+3));
2464 addInstr(env, mk_iMOVds_RR( argregs[argreg++],
2465 iselWordExpr_R(env, e->Iex.Unop.arg,
2466 IEndianess) ) );
2467
2468 cc = mk_PPCCondCode( Pct_ALWAYS, Pcf_NONE );
2469 if (IEndianess == Iend_LE) {
2470 addInstr(env, PPCInstr_Call( cc, (Addr)h_calc_BCDtoDPB,
2471 argiregs,
2472 mk_RetLoc_simple(RLPri_Int)) );
2473 } else {
2474 HWord* fdescr;
2475 fdescr = (HWord*)h_calc_BCDtoDPB;
2476 addInstr(env, PPCInstr_Call( cc, (Addr64)(fdescr[0]),
2477 argiregs,
2478 mk_RetLoc_simple(RLPri_Int)) );
2479 }
2480
2481 addInstr(env, mk_iMOVds_RR(r_dst, argregs[0]));
2482 return r_dst;
2483 }
2484
2485 case Iop_DPBtoBCD: {
2486 /* the following is only valid in 64 bit mode */
2487 if (!mode64) break;
2488
2489 PPCCondCode cc;
2490 UInt argiregs;
2491 HReg argregs[1];
2492 HReg r_dst = newVRegI(env);
2493 Int argreg;
2494
2495 argiregs = 0;
2496 argreg = 0;
2497 argregs[0] = hregPPC_GPR3(mode64);
2498
2499 argiregs |= (1 << (argreg+3));
2500 addInstr(env, mk_iMOVds_RR( argregs[argreg++],
2501 iselWordExpr_R(env, e->Iex.Unop.arg,
2502 IEndianess) ) );
2503
2504 cc = mk_PPCCondCode( Pct_ALWAYS, Pcf_NONE );
2505
2506 if (IEndianess == Iend_LE) {
2507 addInstr(env, PPCInstr_Call( cc, (Addr)h_calc_DPBtoBCD,
2508 argiregs,
2509 mk_RetLoc_simple(RLPri_Int) ) );
2510 } else {
2511 HWord* fdescr;
2512 fdescr = (HWord*)h_calc_DPBtoBCD;
2513 addInstr(env, PPCInstr_Call( cc, (Addr64)(fdescr[0]),
2514 argiregs,
2515 mk_RetLoc_simple(RLPri_Int) ) );
2516 }
2517
2518 addInstr(env, mk_iMOVds_RR(r_dst, argregs[0]));
2519 return r_dst;
2520 }
2521 case Iop_F32toF16x4: {
2522 HReg vdst = newVRegV(env); /* V128 */
2523 HReg dst = newVRegI(env); /* I64*/
2524 HReg r0 = newVRegI(env); /* I16*/
2525 HReg r1 = newVRegI(env); /* I16*/
2526 HReg r2 = newVRegI(env); /* I16*/
2527 HReg r3 = newVRegI(env); /* I16*/
2528 HReg vsrc = iselVecExpr(env, e->Iex.Unop.arg, IEndianess);
2529 PPCAMode *am_off0, *am_off2, *am_off4, *am_off6, *am_off8;
2530 PPCAMode *am_off10, *am_off12, *am_off14;
2531 HReg r_aligned16;
2532
2533 sub_from_sp( env, 32 ); // Move SP down
2534
2535 /* issue instruction */
2536 addInstr(env, PPCInstr_AvUnary(Pav_F32toF16x4, vdst, vsrc));
2537
2538 /* Get a quadword aligned address within our stack space */
2539 r_aligned16 = get_sp_aligned16( env );
2540 am_off0 = PPCAMode_IR( 0, r_aligned16 );
2541 am_off2 = PPCAMode_IR( 2, r_aligned16 );
2542 am_off4 = PPCAMode_IR( 4, r_aligned16 );
2543 am_off6 = PPCAMode_IR( 6, r_aligned16 );
2544 am_off8 = PPCAMode_IR( 8, r_aligned16 );
2545 am_off10 = PPCAMode_IR( 10, r_aligned16 );
2546 am_off12 = PPCAMode_IR( 12, r_aligned16 );
2547 am_off14 = PPCAMode_IR( 14, r_aligned16 );
2548
2549 /* Store v128 result to stack. */
2550 addInstr(env, PPCInstr_AvLdSt(False/*store*/, 16, vdst, am_off0));
2551
2552 /* fetch four I16 from V128, store into contiguous I64 via stack, */
2553 if (IEndianess == Iend_LE) {
2554 addInstr(env, PPCInstr_Load( 2, r3, am_off12, mode64));
2555 addInstr(env, PPCInstr_Load( 2, r2, am_off8, mode64));
2556 addInstr(env, PPCInstr_Load( 2, r1, am_off4, mode64));
2557 addInstr(env, PPCInstr_Load( 2, r0, am_off0, mode64));
2558 } else {
2559 addInstr(env, PPCInstr_Load( 2, r0, am_off14, mode64));
2560 addInstr(env, PPCInstr_Load( 2, r1, am_off10, mode64));
2561 addInstr(env, PPCInstr_Load( 2, r2, am_off6, mode64));
2562 addInstr(env, PPCInstr_Load( 2, r3, am_off2, mode64));
2563 }
2564
2565 /* store in contiguous 64-bit values */
2566 addInstr(env, PPCInstr_Store( 2, am_off6, r3, mode64));
2567 addInstr(env, PPCInstr_Store( 2, am_off4, r2, mode64));
2568 addInstr(env, PPCInstr_Store( 2, am_off2, r1, mode64));
2569 addInstr(env, PPCInstr_Store( 2, am_off0, r0, mode64));
2570
2571 /* Fetch I64 */
2572 addInstr(env, PPCInstr_Load(8, dst, am_off0, mode64));
2573
2574 add_to_sp( env, 32 ); // Reset SP
2575 return dst;
2576 }
2577
2578 default:
2579 break;
2580 }
2581
2582 switch (e->Iex.Unop.op) {
2583 case Iop_ExtractExpD64: {
2584
2585 HReg fr_dst = newVRegI(env);
2586 HReg fr_src = iselDfp64Expr(env, e->Iex.Unop.arg, IEndianess);
2587 HReg tmp = newVRegF(env);
2588 PPCAMode* zero_r1 = PPCAMode_IR( 0, StackFramePtr(env->mode64) );
2589 addInstr(env, PPCInstr_Dfp64Unary(Pfp_DXEX, tmp, fr_src));
2590
2591 // put the D64 result into a integer register
2592 sub_from_sp( env, 16 );
2593 addInstr(env, PPCInstr_FpLdSt(False/*store*/, 8, tmp, zero_r1));
2594 addInstr(env, PPCInstr_Load(8, fr_dst, zero_r1, env->mode64));
2595 add_to_sp( env, 16 );
2596 return fr_dst;
2597 }
2598 case Iop_ExtractExpD128: {
2599 HReg fr_dst = newVRegI(env);
2600 HReg r_srcHi;
2601 HReg r_srcLo;
2602 HReg tmp = newVRegF(env);
2603 PPCAMode* zero_r1 = PPCAMode_IR( 0, StackFramePtr(env->mode64) );
2604
2605 iselDfp128Expr(&r_srcHi, &r_srcLo, env, e->Iex.Unop.arg,
2606 IEndianess);
2607 addInstr(env, PPCInstr_ExtractExpD128(Pfp_DXEXQ, tmp,
2608 r_srcHi, r_srcLo));
2609
2610 sub_from_sp( env, 16 );
2611 addInstr(env, PPCInstr_FpLdSt(False/*store*/, 8, tmp, zero_r1));
2612 addInstr(env, PPCInstr_Load(8, fr_dst, zero_r1, env->mode64));
2613 add_to_sp( env, 16 );
2614 return fr_dst;
2615 }
2616 default:
2617 break;
2618 }
2619
2620 break;
2621 }
2622
2623 /* --------- GET --------- */
2624 case Iex_Get: {
2625 if (ty == Ity_I8 || ty == Ity_I16 ||
2626 ty == Ity_I32 || ((ty == Ity_I64) && mode64)) {
2627 HReg r_dst = newVRegI(env);
2628 PPCAMode* am_addr = PPCAMode_IR( e->Iex.Get.offset,
2629 GuestStatePtr(mode64) );
2630 addInstr(env, PPCInstr_Load( toUChar(sizeofIRType(ty)),
2631 r_dst, am_addr, mode64 ));
2632 return r_dst;
2633 }
2634 break;
2635 }
2636
2637 case Iex_GetI: {
2638 PPCAMode* src_am
2639 = genGuestArrayOffset( env, e->Iex.GetI.descr,
2640 e->Iex.GetI.ix, e->Iex.GetI.bias,
2641 IEndianess );
2642 HReg r_dst = newVRegI(env);
2643 if (mode64 && ty == Ity_I64) {
2644 addInstr(env, PPCInstr_Load( toUChar(8),
2645 r_dst, src_am, mode64 ));
2646 return r_dst;
2647 }
2648 if ((!mode64) && ty == Ity_I32) {
2649 addInstr(env, PPCInstr_Load( toUChar(4),
2650 r_dst, src_am, mode64 ));
2651 return r_dst;
2652 }
2653 break;
2654 }
2655
2656 /* --------- CCALL --------- */
2657 case Iex_CCall: {
2658 vassert(ty == e->Iex.CCall.retty); /* well-formedness of IR */
2659
2660 /* be very restrictive for now. Only 32/64-bit ints allowed for
2661 args, and 32 bits or host machine word for return type. */
2662 if (!(ty == Ity_I32 || (mode64 && ty == Ity_I64)))
2663 goto irreducible;
2664
2665 /* Marshal args, do the call, clear stack. */
2666 UInt addToSp = 0;
2667 RetLoc rloc = mk_RetLoc_INVALID();
2668 doHelperCall( &addToSp, &rloc, env, NULL/*guard*/,
2669 e->Iex.CCall.cee, e->Iex.CCall.retty, e->Iex.CCall.args,
2670 IEndianess );
2671 vassert(is_sane_RetLoc(rloc));
2672 vassert(rloc.pri == RLPri_Int);
2673 vassert(addToSp == 0);
2674
2675 /* GPR3 now holds the destination address from Pin_Goto */
2676 HReg r_dst = newVRegI(env);
2677 addInstr(env, mk_iMOVds_RR(r_dst, hregPPC_GPR3(mode64)));
2678 return r_dst;
2679 }
2680
2681 /* --------- LITERAL --------- */
2682 /* 32/16/8-bit literals */
2683 case Iex_Const: {
2684 Long l;
2685 HReg r_dst = newVRegI(env);
2686 IRConst* con = e->Iex.Const.con;
2687 switch (con->tag) {
2688 case Ico_U64: if (!mode64) goto irreducible;
2689 l = (Long) con->Ico.U64; break;
2690 case Ico_U32: l = (Long)(Int) con->Ico.U32; break;
2691 case Ico_U16: l = (Long)(Int)(Short)con->Ico.U16; break;
2692 case Ico_U8: l = (Long)(Int)(Char )con->Ico.U8; break;
2693 default: vpanic("iselIntExpr_R.const(ppc)");
2694 }
2695 addInstr(env, PPCInstr_LI(r_dst, (ULong)l, mode64));
2696 return r_dst;
2697 }
2698
2699 /* --------- MULTIPLEX --------- */
2700 case Iex_ITE: { // VFD
2701 if ((ty == Ity_I8 || ty == Ity_I16 ||
2702 ty == Ity_I32 || ((ty == Ity_I64) && mode64)) &&
2703 typeOfIRExpr(env->type_env,e->Iex.ITE.cond) == Ity_I1) {
2704 PPCRI* r1 = iselWordExpr_RI(env, e->Iex.ITE.iftrue, IEndianess);
2705 HReg r0 = iselWordExpr_R(env, e->Iex.ITE.iffalse, IEndianess);
2706 HReg r_dst = newVRegI(env);
2707 addInstr(env, mk_iMOVds_RR(r_dst,r0));
2708 PPCCondCode cc = iselCondCode(env, e->Iex.ITE.cond, IEndianess);
2709 addInstr(env, PPCInstr_CMov(cc, r_dst, r1));
2710 return r_dst;
2711 }
2712 break;
2713 }
2714
2715 default:
2716 break;
2717 } /* switch (e->tag) */
2718
2719
2720 /* We get here if no pattern matched. */
2721 irreducible:
2722 ppIRExpr(e);
2723 vpanic("iselIntExpr_R(ppc): cannot reduce tree");
2724 }
2725
2726
2727 /*---------------------------------------------------------*/
2728 /*--- ISEL: Integer expression auxiliaries ---*/
2729 /*---------------------------------------------------------*/
2730
2731 /* --------------------- AMODEs --------------------- */
2732
2733 /* Return an AMode which computes the value of the specified
2734 expression, possibly also adding insns to the code list as a
2735 result. The expression may only be a word-size one.
2736 */
2737
uInt_fits_in_16_bits(UInt u)2738 static Bool uInt_fits_in_16_bits ( UInt u )
2739 {
2740 /* Is u the same as the sign-extend of its lower 16 bits? */
2741 UInt v = u & 0xFFFF;
2742
2743 v = (Int)(v << 16) >> 16; /* sign extend */
2744
2745 return u == v;
2746 }
2747
uLong_fits_in_16_bits(ULong u)2748 static Bool uLong_fits_in_16_bits ( ULong u )
2749 {
2750 /* Is u the same as the sign-extend of its lower 16 bits? */
2751 ULong v = u & 0xFFFFULL;
2752
2753 v = (Long)(v << 48) >> 48; /* sign extend */
2754
2755 return u == v;
2756 }
2757
uLong_is_4_aligned(ULong u)2758 static Bool uLong_is_4_aligned ( ULong u )
2759 {
2760 return toBool((u & 3ULL) == 0);
2761 }
2762
sane_AMode(ISelEnv * env,PPCAMode * am)2763 static Bool sane_AMode ( ISelEnv* env, PPCAMode* am )
2764 {
2765 Bool mode64 = env->mode64;
2766 switch (am->tag) {
2767 case Pam_IR:
2768 /* Using uInt_fits_in_16_bits in 64-bit mode seems a bit bogus,
2769 somehow, but I think it's OK. */
2770 return toBool( hregClass(am->Pam.IR.base) == HRcGPR(mode64) &&
2771 hregIsVirtual(am->Pam.IR.base) &&
2772 uInt_fits_in_16_bits(am->Pam.IR.index) );
2773 case Pam_RR:
2774 return toBool( hregClass(am->Pam.RR.base) == HRcGPR(mode64) &&
2775 hregIsVirtual(am->Pam.RR.base) &&
2776 hregClass(am->Pam.RR.index) == HRcGPR(mode64) &&
2777 hregIsVirtual(am->Pam.RR.index) );
2778 default:
2779 vpanic("sane_AMode: unknown ppc amode tag");
2780 }
2781 }
2782
2783 static
iselWordExpr_AMode(ISelEnv * env,const IRExpr * e,IRType xferTy,IREndness IEndianess)2784 PPCAMode* iselWordExpr_AMode ( ISelEnv* env, const IRExpr* e, IRType xferTy,
2785 IREndness IEndianess )
2786 {
2787 PPCAMode* am = iselWordExpr_AMode_wrk(env, e, xferTy, IEndianess);
2788 vassert(sane_AMode(env, am));
2789 return am;
2790 }
2791
2792 /* DO NOT CALL THIS DIRECTLY ! */
iselWordExpr_AMode_wrk(ISelEnv * env,const IRExpr * e,IRType xferTy,IREndness IEndianess)2793 static PPCAMode* iselWordExpr_AMode_wrk ( ISelEnv* env, const IRExpr* e,
2794 IRType xferTy, IREndness IEndianess )
2795 {
2796 IRType ty = typeOfIRExpr(env->type_env,e);
2797
2798 if (env->mode64) {
2799
2800 /* If the data load/store type is I32 or I64, this amode might
2801 be destined for use in ld/ldu/lwa/st/stu. In which case
2802 insist that if it comes out as an _IR, the immediate must
2803 have its bottom two bits be zero. This does assume that for
2804 any other type (I8/I16/I128/F32/F64/V128) the amode will not
2805 be parked in any such instruction. But that seems a
2806 reasonable assumption. */
2807 Bool aligned4imm = toBool(xferTy == Ity_I32 || xferTy == Ity_I64);
2808
2809 vassert(ty == Ity_I64);
2810
2811 /* Add64(expr,i), where i == sign-extend of (i & 0xFFFF) */
2812 if (e->tag == Iex_Binop
2813 && e->Iex.Binop.op == Iop_Add64
2814 && e->Iex.Binop.arg2->tag == Iex_Const
2815 && e->Iex.Binop.arg2->Iex.Const.con->tag == Ico_U64
2816 && (aligned4imm ? uLong_is_4_aligned(e->Iex.Binop.arg2
2817 ->Iex.Const.con->Ico.U64)
2818 : True)
2819 && uLong_fits_in_16_bits(e->Iex.Binop.arg2
2820 ->Iex.Const.con->Ico.U64)) {
2821 return PPCAMode_IR( (Int)e->Iex.Binop.arg2->Iex.Const.con->Ico.U64,
2822 iselWordExpr_R(env, e->Iex.Binop.arg1,
2823 IEndianess) );
2824 }
2825
2826 /* Add64(expr,expr) */
2827 if (e->tag == Iex_Binop
2828 && e->Iex.Binop.op == Iop_Add64) {
2829 HReg r_base = iselWordExpr_R(env, e->Iex.Binop.arg1, IEndianess);
2830 HReg r_idx = iselWordExpr_R(env, e->Iex.Binop.arg2, IEndianess);
2831 return PPCAMode_RR( r_idx, r_base );
2832 }
2833
2834 } else {
2835
2836 vassert(ty == Ity_I32);
2837
2838 /* Add32(expr,i), where i == sign-extend of (i & 0xFFFF) */
2839 if (e->tag == Iex_Binop
2840 && e->Iex.Binop.op == Iop_Add32
2841 && e->Iex.Binop.arg2->tag == Iex_Const
2842 && e->Iex.Binop.arg2->Iex.Const.con->tag == Ico_U32
2843 && uInt_fits_in_16_bits(e->Iex.Binop.arg2
2844 ->Iex.Const.con->Ico.U32)) {
2845 return PPCAMode_IR( (Int)e->Iex.Binop.arg2->Iex.Const.con->Ico.U32,
2846 iselWordExpr_R(env, e->Iex.Binop.arg1,
2847 IEndianess) );
2848 }
2849
2850 /* Add32(expr,expr) */
2851 if (e->tag == Iex_Binop
2852 && e->Iex.Binop.op == Iop_Add32) {
2853 HReg r_base = iselWordExpr_R(env, e->Iex.Binop.arg1, IEndianess);
2854 HReg r_idx = iselWordExpr_R(env, e->Iex.Binop.arg2, IEndianess);
2855 return PPCAMode_RR( r_idx, r_base );
2856 }
2857
2858 }
2859
2860 /* Doesn't match anything in particular. Generate it into
2861 a register and use that. */
2862 return PPCAMode_IR( 0, iselWordExpr_R(env,e,IEndianess) );
2863 }
2864
2865
2866 /* --------------------- RH --------------------- */
2867
2868 /* Compute an I8/I16/I32 (and I64, in 64-bit mode) into a RH
2869 (reg-or-halfword-immediate). It's important to specify whether the
2870 immediate is to be regarded as signed or not. If yes, this will
2871 never return -32768 as an immediate; this guaranteed that all
2872 signed immediates that are return can have their sign inverted if
2873 need be. */
2874
iselWordExpr_RH(ISelEnv * env,Bool syned,const IRExpr * e,IREndness IEndianess)2875 static PPCRH* iselWordExpr_RH ( ISelEnv* env, Bool syned, const IRExpr* e,
2876 IREndness IEndianess )
2877 {
2878 PPCRH* ri = iselWordExpr_RH_wrk(env, syned, e, IEndianess);
2879 /* sanity checks ... */
2880 switch (ri->tag) {
2881 case Prh_Imm:
2882 vassert(ri->Prh.Imm.syned == syned);
2883 if (syned)
2884 vassert(ri->Prh.Imm.imm16 != 0x8000);
2885 return ri;
2886 case Prh_Reg:
2887 vassert(hregClass(ri->Prh.Reg.reg) == HRcGPR(env->mode64));
2888 vassert(hregIsVirtual(ri->Prh.Reg.reg));
2889 return ri;
2890 default:
2891 vpanic("iselIntExpr_RH: unknown ppc RH tag");
2892 }
2893 }
2894
2895 /* DO NOT CALL THIS DIRECTLY ! */
iselWordExpr_RH_wrk(ISelEnv * env,Bool syned,const IRExpr * e,IREndness IEndianess)2896 static PPCRH* iselWordExpr_RH_wrk ( ISelEnv* env, Bool syned, const IRExpr* e,
2897 IREndness IEndianess )
2898 {
2899 ULong u;
2900 Long l;
2901 IRType ty = typeOfIRExpr(env->type_env,e);
2902 vassert(ty == Ity_I8 || ty == Ity_I16 ||
2903 ty == Ity_I32 || ((ty == Ity_I64) && env->mode64));
2904
2905 /* special case: immediate */
2906 if (e->tag == Iex_Const) {
2907 IRConst* con = e->Iex.Const.con;
2908 /* What value are we aiming to generate? */
2909 switch (con->tag) {
2910 /* Note: Not sign-extending - we carry 'syned' around */
2911 case Ico_U64: vassert(env->mode64);
2912 u = con->Ico.U64; break;
2913 case Ico_U32: u = 0xFFFFFFFF & con->Ico.U32; break;
2914 case Ico_U16: u = 0x0000FFFF & con->Ico.U16; break;
2915 case Ico_U8: u = 0x000000FF & con->Ico.U8; break;
2916 default: vpanic("iselIntExpr_RH.Iex_Const(ppch)");
2917 }
2918 l = (Long)u;
2919 /* Now figure out if it's representable. */
2920 if (!syned && u <= 65535) {
2921 return PPCRH_Imm(False/*unsigned*/, toUShort(u & 0xFFFF));
2922 }
2923 if (syned && l >= -32767 && l <= 32767) {
2924 return PPCRH_Imm(True/*signed*/, toUShort(u & 0xFFFF));
2925 }
2926 /* no luck; use the Slow Way. */
2927 }
2928
2929 /* default case: calculate into a register and return that */
2930 return PPCRH_Reg( iselWordExpr_R ( env, e, IEndianess ) );
2931 }
2932
2933
2934 /* --------------------- RIs --------------------- */
2935
2936 /* Calculate an expression into an PPCRI operand. As with
2937 iselIntExpr_R, the expression can have type 32, 16 or 8 bits, or,
2938 in 64-bit mode, 64 bits. */
2939
iselWordExpr_RI(ISelEnv * env,const IRExpr * e,IREndness IEndianess)2940 static PPCRI* iselWordExpr_RI ( ISelEnv* env, const IRExpr* e,
2941 IREndness IEndianess )
2942 {
2943 PPCRI* ri = iselWordExpr_RI_wrk(env, e, IEndianess);
2944 /* sanity checks ... */
2945 switch (ri->tag) {
2946 case Pri_Imm:
2947 return ri;
2948 case Pri_Reg:
2949 vassert(hregClass(ri->Pri.Reg) == HRcGPR(env->mode64));
2950 vassert(hregIsVirtual(ri->Pri.Reg));
2951 return ri;
2952 default:
2953 vpanic("iselIntExpr_RI: unknown ppc RI tag");
2954 }
2955 }
2956
2957 /* DO NOT CALL THIS DIRECTLY ! */
iselWordExpr_RI_wrk(ISelEnv * env,const IRExpr * e,IREndness IEndianess)2958 static PPCRI* iselWordExpr_RI_wrk ( ISelEnv* env, const IRExpr* e,
2959 IREndness IEndianess )
2960 {
2961 Long l;
2962 IRType ty = typeOfIRExpr(env->type_env,e);
2963 vassert(ty == Ity_I8 || ty == Ity_I16 ||
2964 ty == Ity_I32 || ((ty == Ity_I64) && env->mode64));
2965
2966 /* special case: immediate */
2967 if (e->tag == Iex_Const) {
2968 IRConst* con = e->Iex.Const.con;
2969 switch (con->tag) {
2970 case Ico_U64: vassert(env->mode64);
2971 l = (Long) con->Ico.U64; break;
2972 case Ico_U32: l = (Long)(Int) con->Ico.U32; break;
2973 case Ico_U16: l = (Long)(Int)(Short)con->Ico.U16; break;
2974 case Ico_U8: l = (Long)(Int)(Char )con->Ico.U8; break;
2975 default: vpanic("iselIntExpr_RI.Iex_Const(ppch)");
2976 }
2977 return PPCRI_Imm((ULong)l);
2978 }
2979
2980 /* default case: calculate into a register and return that */
2981 return PPCRI_Reg( iselWordExpr_R ( env, e, IEndianess ) );
2982 }
2983
2984
2985 /* --------------------- RH5u --------------------- */
2986
2987 /* Compute an I8 into a reg-or-5-bit-unsigned-immediate, the latter
2988 being an immediate in the range 1 .. 31 inclusive. Used for doing
2989 shift amounts. Only used in 32-bit mode. */
2990
iselWordExpr_RH5u(ISelEnv * env,const IRExpr * e,IREndness IEndianess)2991 static PPCRH* iselWordExpr_RH5u ( ISelEnv* env, const IRExpr* e,
2992 IREndness IEndianess )
2993 {
2994 PPCRH* ri;
2995 vassert(!env->mode64);
2996 ri = iselWordExpr_RH5u_wrk(env, e, IEndianess);
2997 /* sanity checks ... */
2998 switch (ri->tag) {
2999 case Prh_Imm:
3000 vassert(ri->Prh.Imm.imm16 >= 1 && ri->Prh.Imm.imm16 <= 31);
3001 vassert(!ri->Prh.Imm.syned);
3002 return ri;
3003 case Prh_Reg:
3004 vassert(hregClass(ri->Prh.Reg.reg) == HRcGPR(env->mode64));
3005 vassert(hregIsVirtual(ri->Prh.Reg.reg));
3006 return ri;
3007 default:
3008 vpanic("iselIntExpr_RH5u: unknown ppc RI tag");
3009 }
3010 }
3011
3012 /* DO NOT CALL THIS DIRECTLY ! */
iselWordExpr_RH5u_wrk(ISelEnv * env,const IRExpr * e,IREndness IEndianess)3013 static PPCRH* iselWordExpr_RH5u_wrk ( ISelEnv* env, const IRExpr* e,
3014 IREndness IEndianess )
3015 {
3016 IRType ty = typeOfIRExpr(env->type_env,e);
3017 vassert(ty == Ity_I8);
3018
3019 /* special case: immediate */
3020 if (e->tag == Iex_Const
3021 && e->Iex.Const.con->tag == Ico_U8
3022 && e->Iex.Const.con->Ico.U8 >= 1
3023 && e->Iex.Const.con->Ico.U8 <= 31) {
3024 return PPCRH_Imm(False/*unsigned*/, e->Iex.Const.con->Ico.U8);
3025 }
3026
3027 /* default case: calculate into a register and return that */
3028 return PPCRH_Reg( iselWordExpr_R ( env, e, IEndianess ) );
3029 }
3030
3031
3032 /* --------------------- RH6u --------------------- */
3033
3034 /* Compute an I8 into a reg-or-6-bit-unsigned-immediate, the latter
3035 being an immediate in the range 1 .. 63 inclusive. Used for doing
3036 shift amounts. Only used in 64-bit mode. */
3037
iselWordExpr_RH6u(ISelEnv * env,const IRExpr * e,IREndness IEndianess)3038 static PPCRH* iselWordExpr_RH6u ( ISelEnv* env, const IRExpr* e,
3039 IREndness IEndianess )
3040 {
3041 PPCRH* ri;
3042 vassert(env->mode64);
3043 ri = iselWordExpr_RH6u_wrk(env, e, IEndianess);
3044 /* sanity checks ... */
3045 switch (ri->tag) {
3046 case Prh_Imm:
3047 vassert(ri->Prh.Imm.imm16 >= 1 && ri->Prh.Imm.imm16 <= 63);
3048 vassert(!ri->Prh.Imm.syned);
3049 return ri;
3050 case Prh_Reg:
3051 vassert(hregClass(ri->Prh.Reg.reg) == HRcGPR(env->mode64));
3052 vassert(hregIsVirtual(ri->Prh.Reg.reg));
3053 return ri;
3054 default:
3055 vpanic("iselIntExpr_RH6u: unknown ppc64 RI tag");
3056 }
3057 }
3058
3059 /* DO NOT CALL THIS DIRECTLY ! */
iselWordExpr_RH6u_wrk(ISelEnv * env,const IRExpr * e,IREndness IEndianess)3060 static PPCRH* iselWordExpr_RH6u_wrk ( ISelEnv* env, const IRExpr* e,
3061 IREndness IEndianess )
3062 {
3063 IRType ty = typeOfIRExpr(env->type_env,e);
3064 vassert(ty == Ity_I8);
3065
3066 /* special case: immediate */
3067 if (e->tag == Iex_Const
3068 && e->Iex.Const.con->tag == Ico_U8
3069 && e->Iex.Const.con->Ico.U8 >= 1
3070 && e->Iex.Const.con->Ico.U8 <= 63) {
3071 return PPCRH_Imm(False/*unsigned*/, e->Iex.Const.con->Ico.U8);
3072 }
3073
3074 /* default case: calculate into a register and return that */
3075 return PPCRH_Reg( iselWordExpr_R ( env, e, IEndianess ) );
3076 }
3077
3078
3079 /* --------------------- CONDCODE --------------------- */
3080
3081 /* Generate code to evaluated a bit-typed expression, returning the
3082 condition code which would correspond when the expression would
3083 notionally have returned 1. */
3084
iselCondCode(ISelEnv * env,const IRExpr * e,IREndness IEndianess)3085 static PPCCondCode iselCondCode ( ISelEnv* env, const IRExpr* e,
3086 IREndness IEndianess )
3087 {
3088 /* Uh, there's nothing we can sanity check here, unfortunately. */
3089 return iselCondCode_wrk(env,e, IEndianess);
3090 }
3091
3092 /* DO NOT CALL THIS DIRECTLY ! */
iselCondCode_wrk(ISelEnv * env,const IRExpr * e,IREndness IEndianess)3093 static PPCCondCode iselCondCode_wrk ( ISelEnv* env, const IRExpr* e,
3094 IREndness IEndianess )
3095 {
3096 vassert(e);
3097 vassert(typeOfIRExpr(env->type_env,e) == Ity_I1);
3098
3099 /* Constant 1:Bit */
3100 if (e->tag == Iex_Const && e->Iex.Const.con->Ico.U1 == True) {
3101 // Make a compare that will always be true:
3102 HReg r_zero = newVRegI(env);
3103 addInstr(env, PPCInstr_LI(r_zero, 0, env->mode64));
3104 addInstr(env, PPCInstr_Cmp(False/*unsigned*/, True/*32bit cmp*/,
3105 7/*cr*/, r_zero, PPCRH_Reg(r_zero)));
3106 return mk_PPCCondCode( Pct_TRUE, Pcf_7EQ );
3107 }
3108
3109 /* Not1(...) */
3110 if (e->tag == Iex_Unop && e->Iex.Unop.op == Iop_Not1) {
3111 /* Generate code for the arg, and negate the test condition */
3112 PPCCondCode cond = iselCondCode(env, e->Iex.Unop.arg, IEndianess);
3113 cond.test = invertCondTest(cond.test);
3114 return cond;
3115 }
3116
3117 /* --- patterns rooted at: 32to1 or 64to1 --- */
3118
3119 /* 32to1, 64to1 */
3120 if (e->tag == Iex_Unop &&
3121 (e->Iex.Unop.op == Iop_32to1 || e->Iex.Unop.op == Iop_64to1)) {
3122 HReg src = iselWordExpr_R(env, e->Iex.Unop.arg, IEndianess);
3123 HReg tmp = newVRegI(env);
3124 /* could do better, probably -- andi. */
3125 addInstr(env, PPCInstr_Alu(Palu_AND, tmp,
3126 src, PPCRH_Imm(False,1)));
3127 addInstr(env, PPCInstr_Cmp(False/*unsigned*/, True/*32bit cmp*/,
3128 7/*cr*/, tmp, PPCRH_Imm(False,1)));
3129 return mk_PPCCondCode( Pct_TRUE, Pcf_7EQ );
3130 }
3131
3132 /* --- patterns rooted at: CmpNEZ8 --- */
3133
3134 /* CmpNEZ8(x) */
3135 /* Note this cloned as CmpNE8(x,0) below. */
3136 /* could do better -- andi. */
3137 if (e->tag == Iex_Unop
3138 && e->Iex.Unop.op == Iop_CmpNEZ8) {
3139 HReg arg = iselWordExpr_R(env, e->Iex.Unop.arg, IEndianess);
3140 HReg tmp = newVRegI(env);
3141 addInstr(env, PPCInstr_Alu(Palu_AND, tmp, arg,
3142 PPCRH_Imm(False,0xFF)));
3143 addInstr(env, PPCInstr_Cmp(False/*unsigned*/, True/*32bit cmp*/,
3144 7/*cr*/, tmp, PPCRH_Imm(False,0)));
3145 return mk_PPCCondCode( Pct_FALSE, Pcf_7EQ );
3146 }
3147
3148 /* --- patterns rooted at: CmpNEZ32 --- */
3149
3150 /* CmpNEZ32(x) */
3151 if (e->tag == Iex_Unop
3152 && e->Iex.Unop.op == Iop_CmpNEZ32) {
3153 HReg r1 = iselWordExpr_R(env, e->Iex.Unop.arg, IEndianess);
3154 addInstr(env, PPCInstr_Cmp(False/*unsigned*/, True/*32bit cmp*/,
3155 7/*cr*/, r1, PPCRH_Imm(False,0)));
3156 return mk_PPCCondCode( Pct_FALSE, Pcf_7EQ );
3157 }
3158
3159 /* --- patterns rooted at: Cmp*32* --- */
3160
3161 /* Cmp*32*(x,y) */
3162 if (e->tag == Iex_Binop
3163 && (e->Iex.Binop.op == Iop_CmpEQ32
3164 || e->Iex.Binop.op == Iop_CmpNE32
3165 || e->Iex.Binop.op == Iop_CmpLT32S
3166 || e->Iex.Binop.op == Iop_CmpLT32U
3167 || e->Iex.Binop.op == Iop_CmpLE32S
3168 || e->Iex.Binop.op == Iop_CmpLE32U)) {
3169 Bool syned = (e->Iex.Binop.op == Iop_CmpLT32S ||
3170 e->Iex.Binop.op == Iop_CmpLE32S);
3171 HReg r1 = iselWordExpr_R(env, e->Iex.Binop.arg1, IEndianess);
3172 PPCRH* ri2 = iselWordExpr_RH(env, syned, e->Iex.Binop.arg2, IEndianess);
3173 addInstr(env, PPCInstr_Cmp(syned, True/*32bit cmp*/,
3174 7/*cr*/, r1, ri2));
3175
3176 switch (e->Iex.Binop.op) {
3177 case Iop_CmpEQ32: return mk_PPCCondCode( Pct_TRUE, Pcf_7EQ );
3178 case Iop_CmpNE32: return mk_PPCCondCode( Pct_FALSE, Pcf_7EQ );
3179 case Iop_CmpLT32U: case Iop_CmpLT32S:
3180 return mk_PPCCondCode( Pct_TRUE, Pcf_7LT );
3181 case Iop_CmpLE32U: case Iop_CmpLE32S:
3182 return mk_PPCCondCode( Pct_FALSE, Pcf_7GT );
3183 default: vpanic("iselCondCode(ppc): CmpXX32");
3184 }
3185 }
3186
3187 /* --- patterns rooted at: CmpNEZ64 --- */
3188
3189 /* CmpNEZ64 */
3190 if (e->tag == Iex_Unop
3191 && e->Iex.Unop.op == Iop_CmpNEZ64) {
3192 if (!env->mode64) {
3193 HReg hi, lo;
3194 HReg tmp = newVRegI(env);
3195 iselInt64Expr( &hi, &lo, env, e->Iex.Unop.arg, IEndianess );
3196 addInstr(env, PPCInstr_Alu(Palu_OR, tmp, lo, PPCRH_Reg(hi)));
3197 addInstr(env, PPCInstr_Cmp(False/*sign*/, True/*32bit cmp*/,
3198 7/*cr*/, tmp,PPCRH_Imm(False,0)));
3199 return mk_PPCCondCode( Pct_FALSE, Pcf_7EQ );
3200 } else { // mode64
3201 HReg r_src = iselWordExpr_R(env, e->Iex.Unop.arg, IEndianess);
3202 addInstr(env, PPCInstr_Cmp(False/*sign*/, False/*64bit cmp*/,
3203 7/*cr*/, r_src,PPCRH_Imm(False,0)));
3204 return mk_PPCCondCode( Pct_FALSE, Pcf_7EQ );
3205 }
3206 }
3207
3208 /* --- patterns rooted at: Cmp*64* --- */
3209
3210 /* Cmp*64*(x,y) */
3211 if (e->tag == Iex_Binop
3212 && (e->Iex.Binop.op == Iop_CmpEQ64
3213 || e->Iex.Binop.op == Iop_CmpNE64
3214 || e->Iex.Binop.op == Iop_CmpLT64S
3215 || e->Iex.Binop.op == Iop_CmpLT64U
3216 || e->Iex.Binop.op == Iop_CmpLE64S
3217 || e->Iex.Binop.op == Iop_CmpLE64U)) {
3218 Bool syned = (e->Iex.Binop.op == Iop_CmpLT64S ||
3219 e->Iex.Binop.op == Iop_CmpLE64S);
3220 HReg r1 = iselWordExpr_R(env, e->Iex.Binop.arg1, IEndianess);
3221 PPCRH* ri2 = iselWordExpr_RH(env, syned, e->Iex.Binop.arg2, IEndianess);
3222 vassert(env->mode64);
3223 addInstr(env, PPCInstr_Cmp(syned, False/*64bit cmp*/,
3224 7/*cr*/, r1, ri2));
3225
3226 switch (e->Iex.Binop.op) {
3227 case Iop_CmpEQ64: return mk_PPCCondCode( Pct_TRUE, Pcf_7EQ );
3228 case Iop_CmpNE64: return mk_PPCCondCode( Pct_FALSE, Pcf_7EQ );
3229 case Iop_CmpLT64U: return mk_PPCCondCode( Pct_TRUE, Pcf_7LT );
3230 case Iop_CmpLE64U: return mk_PPCCondCode( Pct_FALSE, Pcf_7GT );
3231 default: vpanic("iselCondCode(ppc): CmpXX64");
3232 }
3233 }
3234
3235 /* --- patterns rooted at: CmpNE8 --- */
3236
3237 /* CmpNE8(x,0) */
3238 /* Note this is a direct copy of CmpNEZ8 above. */
3239 /* could do better -- andi. */
3240 if (e->tag == Iex_Binop
3241 && e->Iex.Binop.op == Iop_CmpNE8
3242 && isZeroU8(e->Iex.Binop.arg2)) {
3243 HReg arg = iselWordExpr_R(env, e->Iex.Binop.arg1, IEndianess);
3244 HReg tmp = newVRegI(env);
3245 addInstr(env, PPCInstr_Alu(Palu_AND, tmp, arg,
3246 PPCRH_Imm(False,0xFF)));
3247 addInstr(env, PPCInstr_Cmp(False/*unsigned*/, True/*32bit cmp*/,
3248 7/*cr*/, tmp, PPCRH_Imm(False,0)));
3249 return mk_PPCCondCode( Pct_FALSE, Pcf_7EQ );
3250 }
3251
3252 /* var */
3253 if (e->tag == Iex_RdTmp) {
3254 HReg r_src = lookupIRTemp(env, e->Iex.RdTmp.tmp);
3255 HReg src_masked = newVRegI(env);
3256 addInstr(env,
3257 PPCInstr_Alu(Palu_AND, src_masked,
3258 r_src, PPCRH_Imm(False,1)));
3259 addInstr(env,
3260 PPCInstr_Cmp(False/*unsigned*/, True/*32bit cmp*/,
3261 7/*cr*/, src_masked, PPCRH_Imm(False,1)));
3262 return mk_PPCCondCode( Pct_TRUE, Pcf_7EQ );
3263 }
3264
3265 vex_printf("iselCondCode(ppc): No such tag(%u)\n", e->tag);
3266 ppIRExpr(e);
3267 vpanic("iselCondCode(ppc)");
3268 }
3269
3270
3271 /*---------------------------------------------------------*/
3272 /*--- ISEL: Integer expressions (128 bit) ---*/
3273 /*---------------------------------------------------------*/
3274
3275 /* 64-bit mode ONLY: compute a 128-bit value into a register pair,
3276 which is returned as the first two parameters. As with
3277 iselWordExpr_R, these may be either real or virtual regs; in any
3278 case they must not be changed by subsequent code emitted by the
3279 caller. */
3280
iselInt128Expr(HReg * rHi,HReg * rLo,ISelEnv * env,const IRExpr * e,IREndness IEndianess)3281 static void iselInt128Expr ( HReg* rHi, HReg* rLo, ISelEnv* env,
3282 const IRExpr* e, IREndness IEndianess )
3283 {
3284 vassert(env->mode64);
3285 iselInt128Expr_wrk(rHi, rLo, env, e, IEndianess);
3286 # if 0
3287 vex_printf("\n"); ppIRExpr(e); vex_printf("\n");
3288 # endif
3289 vassert(hregClass(*rHi) == HRcGPR(env->mode64));
3290 vassert(hregIsVirtual(*rHi));
3291 vassert(hregClass(*rLo) == HRcGPR(env->mode64));
3292 vassert(hregIsVirtual(*rLo));
3293 }
3294
3295 /* DO NOT CALL THIS DIRECTLY ! */
iselInt128Expr_wrk(HReg * rHi,HReg * rLo,ISelEnv * env,const IRExpr * e,IREndness IEndianess)3296 static void iselInt128Expr_wrk ( HReg* rHi, HReg* rLo, ISelEnv* env,
3297 const IRExpr* e, IREndness IEndianess )
3298 {
3299 Bool mode64 = env->mode64;
3300
3301 vassert(e);
3302 vassert(typeOfIRExpr(env->type_env,e) == Ity_I128);
3303
3304 /* read 128-bit IRTemp */
3305 if (e->tag == Iex_RdTmp) {
3306 lookupIRTempPair( rHi, rLo, env, e->Iex.RdTmp.tmp);
3307 return;
3308 }
3309
3310 /* 128-bit GET */
3311 if (e->tag == Iex_Get) {
3312 PPCAMode* am_addr = PPCAMode_IR( e->Iex.Get.offset,
3313 GuestStatePtr(mode64) );
3314 PPCAMode* am_addr4 = advance4(env, am_addr);
3315 HReg tLo = newVRegI(env);
3316 HReg tHi = newVRegI(env);
3317
3318 addInstr(env, PPCInstr_Load( 8, tHi, am_addr, mode64));
3319 addInstr(env, PPCInstr_Load( 8, tLo, am_addr4, mode64));
3320 *rHi = tHi;
3321 *rLo = tLo;
3322 return;
3323 }
3324
3325 /* --------- BINARY ops --------- */
3326 if (e->tag == Iex_Binop) {
3327 switch (e->Iex.Binop.op) {
3328 /* 64 x 64 -> 128 multiply */
3329 case Iop_MullU64:
3330 case Iop_MullS64: {
3331 HReg tLo = newVRegI(env);
3332 HReg tHi = newVRegI(env);
3333 Bool syned = toBool(e->Iex.Binop.op == Iop_MullS64);
3334 HReg r_srcL = iselWordExpr_R(env, e->Iex.Binop.arg1, IEndianess);
3335 HReg r_srcR = iselWordExpr_R(env, e->Iex.Binop.arg2, IEndianess);
3336 addInstr(env, PPCInstr_MulL(False/*signedness irrelevant*/,
3337 False/*lo64*/, False/*64bit mul*/,
3338 tLo, r_srcL, r_srcR));
3339 addInstr(env, PPCInstr_MulL(syned,
3340 True/*hi64*/, False/*64bit mul*/,
3341 tHi, r_srcL, r_srcR));
3342 *rHi = tHi;
3343 *rLo = tLo;
3344 return;
3345 }
3346
3347 /* 64HLto128(e1,e2) */
3348 case Iop_64HLto128:
3349 *rHi = iselWordExpr_R(env, e->Iex.Binop.arg1, IEndianess);
3350 *rLo = iselWordExpr_R(env, e->Iex.Binop.arg2, IEndianess);
3351 return;
3352 default:
3353 break;
3354 }
3355 } /* if (e->tag == Iex_Binop) */
3356
3357
3358 /* --------- UNARY ops --------- */
3359 if (e->tag == Iex_Unop) {
3360 switch (e->Iex.Unop.op) {
3361 default:
3362 break;
3363 }
3364 } /* if (e->tag == Iex_Unop) */
3365
3366 vex_printf("iselInt128Expr(ppc64): No such tag(%u)\n", e->tag);
3367 ppIRExpr(e);
3368 vpanic("iselInt128Expr(ppc64)");
3369 }
3370
3371
3372 /*---------------------------------------------------------*/
3373 /*--- ISEL: Integer expressions (64 bit) ---*/
3374 /*---------------------------------------------------------*/
3375
3376 /* 32-bit mode ONLY: compute a 128-bit value into a register quad */
iselInt128Expr_to_32x4(HReg * rHi,HReg * rMedHi,HReg * rMedLo,HReg * rLo,ISelEnv * env,const IRExpr * e,IREndness IEndianess)3377 static void iselInt128Expr_to_32x4 ( HReg* rHi, HReg* rMedHi, HReg* rMedLo,
3378 HReg* rLo, ISelEnv* env, const IRExpr* e,
3379 IREndness IEndianess )
3380 {
3381 vassert(!env->mode64);
3382 iselInt128Expr_to_32x4_wrk(rHi, rMedHi, rMedLo, rLo, env, e, IEndianess);
3383 # if 0
3384 vex_printf("\n"); ppIRExpr(e); vex_printf("\n");
3385 # endif
3386 vassert(hregClass(*rHi) == HRcInt32);
3387 vassert(hregIsVirtual(*rHi));
3388 vassert(hregClass(*rMedHi) == HRcInt32);
3389 vassert(hregIsVirtual(*rMedHi));
3390 vassert(hregClass(*rMedLo) == HRcInt32);
3391 vassert(hregIsVirtual(*rMedLo));
3392 vassert(hregClass(*rLo) == HRcInt32);
3393 vassert(hregIsVirtual(*rLo));
3394 }
3395
iselInt128Expr_to_32x4_wrk(HReg * rHi,HReg * rMedHi,HReg * rMedLo,HReg * rLo,ISelEnv * env,const IRExpr * e,IREndness IEndianess)3396 static void iselInt128Expr_to_32x4_wrk ( HReg* rHi, HReg* rMedHi,
3397 HReg* rMedLo, HReg* rLo,
3398 ISelEnv* env, const IRExpr* e,
3399 IREndness IEndianess )
3400 {
3401 vassert(e);
3402 vassert(typeOfIRExpr(env->type_env,e) == Ity_I128);
3403
3404 /* read 128-bit IRTemp */
3405 if (e->tag == Iex_RdTmp) {
3406 lookupIRTempQuad( rHi, rMedHi, rMedLo, rLo, env, e->Iex.RdTmp.tmp);
3407 return;
3408 }
3409
3410 if (e->tag == Iex_Binop) {
3411
3412 IROp op_binop = e->Iex.Binop.op;
3413 switch (op_binop) {
3414 case Iop_64HLto128:
3415 iselInt64Expr(rHi, rMedHi, env, e->Iex.Binop.arg1, IEndianess);
3416 iselInt64Expr(rMedLo, rLo, env, e->Iex.Binop.arg2, IEndianess);
3417 return;
3418 default:
3419 vex_printf("iselInt128Expr_to_32x4_wrk: Binop case 0x%x not found\n",
3420 op_binop);
3421 break;
3422 }
3423 }
3424
3425 vex_printf("iselInt128Expr_to_32x4_wrk: e->tag 0x%x not found\n", e->tag);
3426 return;
3427 }
3428
3429 /* 32-bit mode ONLY: compute a 64-bit value into a register pair,
3430 which is returned as the first two parameters. As with
3431 iselIntExpr_R, these may be either real or virtual regs; in any
3432 case they must not be changed by subsequent code emitted by the
3433 caller. */
3434
iselInt64Expr(HReg * rHi,HReg * rLo,ISelEnv * env,const IRExpr * e,IREndness IEndianess)3435 static void iselInt64Expr ( HReg* rHi, HReg* rLo,
3436 ISelEnv* env, const IRExpr* e,
3437 IREndness IEndianess )
3438 {
3439 vassert(!env->mode64);
3440 iselInt64Expr_wrk(rHi, rLo, env, e, IEndianess);
3441 # if 0
3442 vex_printf("\n"); ppIRExpr(e); vex_printf("\n");
3443 # endif
3444 vassert(hregClass(*rHi) == HRcInt32);
3445 vassert(hregIsVirtual(*rHi));
3446 vassert(hregClass(*rLo) == HRcInt32);
3447 vassert(hregIsVirtual(*rLo));
3448 }
3449
3450 /* DO NOT CALL THIS DIRECTLY ! */
iselInt64Expr_wrk(HReg * rHi,HReg * rLo,ISelEnv * env,const IRExpr * e,IREndness IEndianess)3451 static void iselInt64Expr_wrk ( HReg* rHi, HReg* rLo,
3452 ISelEnv* env, const IRExpr* e,
3453 IREndness IEndianess )
3454 {
3455 vassert(e);
3456 vassert(typeOfIRExpr(env->type_env,e) == Ity_I64);
3457
3458 /* 64-bit load */
3459 if (e->tag == Iex_Load && e->Iex.Load.end == IEndianess) {
3460 HReg tLo = newVRegI(env);
3461 HReg tHi = newVRegI(env);
3462 HReg r_addr = iselWordExpr_R(env, e->Iex.Load.addr, IEndianess);
3463 vassert(!env->mode64);
3464 addInstr(env, PPCInstr_Load( 4/*byte-load*/,
3465 tHi, PPCAMode_IR( 0, r_addr ),
3466 False/*32-bit insn please*/) );
3467 addInstr(env, PPCInstr_Load( 4/*byte-load*/,
3468 tLo, PPCAMode_IR( 4, r_addr ),
3469 False/*32-bit insn please*/) );
3470 *rHi = tHi;
3471 *rLo = tLo;
3472 return;
3473 }
3474
3475 /* 64-bit literal */
3476 if (e->tag == Iex_Const) {
3477 ULong w64 = e->Iex.Const.con->Ico.U64;
3478 UInt wHi = ((UInt)(w64 >> 32)) & 0xFFFFFFFF;
3479 UInt wLo = ((UInt)w64) & 0xFFFFFFFF;
3480 HReg tLo = newVRegI(env);
3481 HReg tHi = newVRegI(env);
3482 vassert(e->Iex.Const.con->tag == Ico_U64);
3483 addInstr(env, PPCInstr_LI(tHi, (Long)(Int)wHi, False/*mode32*/));
3484 addInstr(env, PPCInstr_LI(tLo, (Long)(Int)wLo, False/*mode32*/));
3485 *rHi = tHi;
3486 *rLo = tLo;
3487 return;
3488 }
3489
3490 /* read 64-bit IRTemp */
3491 if (e->tag == Iex_RdTmp) {
3492 lookupIRTempPair( rHi, rLo, env, e->Iex.RdTmp.tmp);
3493 return;
3494 }
3495
3496 /* 64-bit GET */
3497 if (e->tag == Iex_Get) {
3498 PPCAMode* am_addr = PPCAMode_IR( e->Iex.Get.offset,
3499 GuestStatePtr(False/*mode32*/) );
3500 PPCAMode* am_addr4 = advance4(env, am_addr);
3501 HReg tLo = newVRegI(env);
3502 HReg tHi = newVRegI(env);
3503 addInstr(env, PPCInstr_Load( 4, tHi, am_addr, False/*mode32*/ ));
3504 addInstr(env, PPCInstr_Load( 4, tLo, am_addr4, False/*mode32*/ ));
3505 *rHi = tHi;
3506 *rLo = tLo;
3507 return;
3508 }
3509
3510 /* --------- CCALL --------- */
3511 if(e->tag == Iex_CCall) {
3512 IRType ty = typeOfIRExpr(env->type_env,e);
3513 Bool mode64 = env->mode64;
3514
3515 vassert(ty == e->Iex.CCall.retty); /* well-formedness of IR */
3516
3517 /* be very restrictive for now. Only 32-bit ints allowed for
3518 args, and 32 bits or host machine word for return type. */
3519 vassert(!(ty == Ity_I32 || (mode64 && ty == Ity_I64)));
3520
3521 /* Marshal args, do the call, clear stack. */
3522 UInt addToSp = 0;
3523 RetLoc rloc = mk_RetLoc_INVALID();
3524 doHelperCall( &addToSp, &rloc, env, NULL/*guard*/,
3525 e->Iex.CCall.cee, e->Iex.CCall.retty, e->Iex.CCall.args,
3526 IEndianess );
3527 vassert(is_sane_RetLoc(rloc));
3528
3529 vassert(rloc.pri == RLPri_2Int);
3530 vassert(addToSp == 0);
3531
3532 /* GPR3 now holds the destination address from Pin_Goto */
3533 HReg r_dst = newVRegI(env);
3534 addInstr(env, mk_iMOVds_RR(r_dst, hregPPC_GPR3(mode64)));
3535 *rHi = r_dst;
3536 *rLo = r_dst;
3537 return;
3538 }
3539
3540 /* 64-bit ITE */
3541 if (e->tag == Iex_ITE) { // VFD
3542 HReg e0Lo, e0Hi, eXLo, eXHi;
3543 iselInt64Expr(&eXHi, &eXLo, env, e->Iex.ITE.iftrue, IEndianess);
3544 iselInt64Expr(&e0Hi, &e0Lo, env, e->Iex.ITE.iffalse, IEndianess);
3545 HReg tLo = newVRegI(env);
3546 HReg tHi = newVRegI(env);
3547 addInstr(env, mk_iMOVds_RR(tHi,e0Hi));
3548 addInstr(env, mk_iMOVds_RR(tLo,e0Lo));
3549 PPCCondCode cc = iselCondCode(env, e->Iex.ITE.cond, IEndianess);
3550 addInstr(env, PPCInstr_CMov(cc,tHi,PPCRI_Reg(eXHi)));
3551 addInstr(env, PPCInstr_CMov(cc,tLo,PPCRI_Reg(eXLo)));
3552 *rHi = tHi;
3553 *rLo = tLo;
3554 return;
3555 }
3556
3557 /* --------- BINARY ops --------- */
3558 if (e->tag == Iex_Binop) {
3559 IROp op_binop = e->Iex.Binop.op;
3560 switch (op_binop) {
3561 /* 32 x 32 -> 64 multiply */
3562 case Iop_MullU32:
3563 case Iop_MullS32: {
3564 HReg tLo = newVRegI(env);
3565 HReg tHi = newVRegI(env);
3566 Bool syned = toBool(op_binop == Iop_MullS32);
3567 HReg r_srcL = iselWordExpr_R(env, e->Iex.Binop.arg1,
3568 IEndianess);
3569 HReg r_srcR = iselWordExpr_R(env, e->Iex.Binop.arg2,
3570 IEndianess);
3571 addInstr(env, PPCInstr_MulL(False/*signedness irrelevant*/,
3572 False/*lo32*/, True/*32bit mul*/,
3573 tLo, r_srcL, r_srcR));
3574 addInstr(env, PPCInstr_MulL(syned,
3575 True/*hi32*/, True/*32bit mul*/,
3576 tHi, r_srcL, r_srcR));
3577 *rHi = tHi;
3578 *rLo = tLo;
3579 return;
3580 }
3581
3582 /* Or64/And64/Xor64 */
3583 case Iop_Or64:
3584 case Iop_And64:
3585 case Iop_Xor64: {
3586 HReg xLo, xHi, yLo, yHi;
3587 HReg tLo = newVRegI(env);
3588 HReg tHi = newVRegI(env);
3589 PPCAluOp op = (op_binop == Iop_Or64) ? Palu_OR :
3590 (op_binop == Iop_And64) ? Palu_AND : Palu_XOR;
3591 iselInt64Expr(&xHi, &xLo, env, e->Iex.Binop.arg1, IEndianess);
3592 iselInt64Expr(&yHi, &yLo, env, e->Iex.Binop.arg2, IEndianess);
3593 addInstr(env, PPCInstr_Alu(op, tHi, xHi, PPCRH_Reg(yHi)));
3594 addInstr(env, PPCInstr_Alu(op, tLo, xLo, PPCRH_Reg(yLo)));
3595 *rHi = tHi;
3596 *rLo = tLo;
3597 return;
3598 }
3599
3600 /* Add64 */
3601 case Iop_Add64: {
3602 HReg xLo, xHi, yLo, yHi;
3603 HReg tLo = newVRegI(env);
3604 HReg tHi = newVRegI(env);
3605 iselInt64Expr(&xHi, &xLo, env, e->Iex.Binop.arg1, IEndianess);
3606 iselInt64Expr(&yHi, &yLo, env, e->Iex.Binop.arg2, IEndianess);
3607 addInstr(env, PPCInstr_AddSubC( True/*add*/, True /*set carry*/,
3608 tLo, xLo, yLo));
3609 addInstr(env, PPCInstr_AddSubC( True/*add*/, False/*read carry*/,
3610 tHi, xHi, yHi));
3611 *rHi = tHi;
3612 *rLo = tLo;
3613 return;
3614 }
3615
3616 /* 32HLto64(e1,e2) */
3617 case Iop_32HLto64:
3618 *rHi = iselWordExpr_R(env, e->Iex.Binop.arg1, IEndianess);
3619 *rLo = iselWordExpr_R(env, e->Iex.Binop.arg2, IEndianess);
3620 return;
3621
3622 /* F64toI64[S|U] */
3623 case Iop_F64toI64S: case Iop_F64toI64U: {
3624 HReg tLo = newVRegI(env);
3625 HReg tHi = newVRegI(env);
3626 HReg r1 = StackFramePtr(env->mode64);
3627 PPCAMode* zero_r1 = PPCAMode_IR( 0, r1 );
3628 PPCAMode* four_r1 = PPCAMode_IR( 4, r1 );
3629 HReg fsrc = iselDblExpr(env, e->Iex.Binop.arg2,
3630 IEndianess);
3631 HReg ftmp = newVRegF(env);
3632
3633 vassert(!env->mode64);
3634 /* Set host rounding mode */
3635 set_FPU_rounding_mode( env, e->Iex.Binop.arg1, IEndianess );
3636
3637 sub_from_sp( env, 16 );
3638 addInstr(env, PPCInstr_FpCftI(False/*F->I*/, False/*int64*/,
3639 (op_binop == Iop_F64toI64S) ? True : False,
3640 True, ftmp, fsrc));
3641 addInstr(env, PPCInstr_FpLdSt(False/*store*/, 8, ftmp, zero_r1));
3642 addInstr(env, PPCInstr_Load(4, tHi, zero_r1, False/*mode32*/));
3643 addInstr(env, PPCInstr_Load(4, tLo, four_r1, False/*mode32*/));
3644 add_to_sp( env, 16 );
3645
3646 ///* Restore default FPU rounding. */
3647 //set_FPU_rounding_default( env );
3648 *rHi = tHi;
3649 *rLo = tLo;
3650 return;
3651 }
3652 case Iop_D64toI64S: {
3653 HReg tLo = newVRegI(env);
3654 HReg tHi = newVRegI(env);
3655 HReg r1 = StackFramePtr(env->mode64);
3656 PPCAMode* zero_r1 = PPCAMode_IR( 0, r1 );
3657 PPCAMode* four_r1 = PPCAMode_IR( 4, r1 );
3658 HReg fr_src = iselDfp64Expr(env, e->Iex.Binop.arg2, IEndianess);
3659 HReg tmp = newVRegF(env);
3660
3661 vassert(!env->mode64);
3662 set_FPU_DFP_rounding_mode( env, e->Iex.Binop.arg1, IEndianess );
3663 addInstr(env, PPCInstr_Dfp64Unary(Pfp_DCTFIX, tmp, fr_src));
3664
3665 sub_from_sp( env, 16 );
3666 addInstr(env, PPCInstr_FpLdSt(False/*store*/, 8, tmp, zero_r1));
3667 addInstr(env, PPCInstr_Load(4, tHi, zero_r1, False/*mode32*/));
3668 addInstr(env, PPCInstr_Load(4, tLo, four_r1, False/*mode32*/));
3669 add_to_sp( env, 16 );
3670 *rHi = tHi;
3671 *rLo = tLo;
3672 return;
3673 }
3674 case Iop_D128toI64S: {
3675 PPCFpOp fpop = Pfp_DCTFIXQ;
3676 HReg r_srcHi = newVRegF(env);
3677 HReg r_srcLo = newVRegF(env);
3678 HReg tLo = newVRegI(env);
3679 HReg tHi = newVRegI(env);
3680 HReg ftmp = newVRegF(env);
3681 PPCAMode* zero_r1 = PPCAMode_IR( 0, StackFramePtr(env->mode64) );
3682 PPCAMode* four_r1 = PPCAMode_IR( 4, StackFramePtr(env->mode64) );
3683
3684 set_FPU_DFP_rounding_mode( env, e->Iex.Binop.arg1, IEndianess );
3685 iselDfp128Expr(&r_srcHi, &r_srcLo, env, e->Iex.Binop.arg2,
3686 IEndianess);
3687 addInstr(env, PPCInstr_DfpD128toD64(fpop, ftmp, r_srcHi, r_srcLo));
3688
3689 // put the D64 result into an integer register pair
3690 sub_from_sp( env, 16 );
3691 addInstr(env, PPCInstr_FpLdSt(False/*store*/, 8, ftmp, zero_r1));
3692 addInstr(env, PPCInstr_Load(4, tHi, zero_r1, False/*mode32*/));
3693 addInstr(env, PPCInstr_Load(4, tLo, four_r1, False/*mode32*/));
3694 add_to_sp( env, 16 );
3695 *rHi = tHi;
3696 *rLo = tLo;
3697 return;
3698 }
3699 default:
3700 break;
3701 }
3702 } /* if (e->tag == Iex_Binop) */
3703
3704
3705 /* --------- UNARY ops --------- */
3706 if (e->tag == Iex_Unop) {
3707 switch (e->Iex.Unop.op) {
3708
3709 /* CmpwNEZ64(e) */
3710 case Iop_CmpwNEZ64: {
3711 HReg argHi, argLo;
3712 HReg tmp1 = newVRegI(env);
3713 HReg tmp2 = newVRegI(env);
3714 iselInt64Expr(&argHi, &argLo, env, e->Iex.Unop.arg, IEndianess);
3715 /* tmp1 = argHi | argLo */
3716 addInstr(env, PPCInstr_Alu(Palu_OR, tmp1, argHi, PPCRH_Reg(argLo)));
3717 /* tmp2 = (tmp1 | -tmp1) >>s 31 */
3718 addInstr(env, PPCInstr_Unary(Pun_NEG,tmp2,tmp1));
3719 addInstr(env, PPCInstr_Alu(Palu_OR, tmp2, tmp2, PPCRH_Reg(tmp1)));
3720 addInstr(env, PPCInstr_Shft(Pshft_SAR, True/*32bit shift*/,
3721 tmp2, tmp2, PPCRH_Imm(False, 31)));
3722 *rHi = tmp2;
3723 *rLo = tmp2; /* yes, really tmp2 */
3724 return;
3725 }
3726
3727 /* Left64 */
3728 case Iop_Left64: {
3729 HReg argHi, argLo;
3730 HReg zero32 = newVRegI(env);
3731 HReg resHi = newVRegI(env);
3732 HReg resLo = newVRegI(env);
3733 iselInt64Expr(&argHi, &argLo, env, e->Iex.Unop.arg, IEndianess);
3734 vassert(env->mode64 == False);
3735 addInstr(env, PPCInstr_LI(zero32, 0, env->mode64));
3736 /* resHi:resLo = - argHi:argLo */
3737 addInstr(env, PPCInstr_AddSubC( False/*sub*/, True/*set carry*/,
3738 resLo, zero32, argLo ));
3739 addInstr(env, PPCInstr_AddSubC( False/*sub*/, False/*read carry*/,
3740 resHi, zero32, argHi ));
3741 /* resHi:resLo |= srcHi:srcLo */
3742 addInstr(env, PPCInstr_Alu(Palu_OR, resLo, resLo, PPCRH_Reg(argLo)));
3743 addInstr(env, PPCInstr_Alu(Palu_OR, resHi, resHi, PPCRH_Reg(argHi)));
3744 *rHi = resHi;
3745 *rLo = resLo;
3746 return;
3747 }
3748
3749 /* 32Sto64(e) */
3750 case Iop_32Sto64: {
3751 HReg tHi = newVRegI(env);
3752 HReg src = iselWordExpr_R(env, e->Iex.Unop.arg, IEndianess);
3753 addInstr(env, PPCInstr_Shft(Pshft_SAR, True/*32bit shift*/,
3754 tHi, src, PPCRH_Imm(False,31)));
3755 *rHi = tHi;
3756 *rLo = src;
3757 return;
3758 }
3759 case Iop_ExtractExpD64: {
3760 HReg tmp = newVRegF(env);
3761 HReg fr_src = iselDfp64Expr(env, e->Iex.Unop.arg, IEndianess);
3762 HReg tLo = newVRegI(env);
3763 HReg tHi = newVRegI(env);
3764 PPCAMode* zero_r1 = PPCAMode_IR( 0, StackFramePtr(env->mode64) );
3765 PPCAMode* four_r1 = PPCAMode_IR( 4, StackFramePtr(env->mode64) );
3766
3767 addInstr(env, PPCInstr_Dfp64Unary(Pfp_DXEX, tmp, fr_src));
3768
3769 // put the D64 result into a integer register pair
3770 sub_from_sp( env, 16 );
3771 addInstr(env, PPCInstr_FpLdSt(False/*store*/, 8, tmp, zero_r1));
3772 addInstr(env, PPCInstr_Load(4, tHi, zero_r1, False/*mode32*/));
3773 addInstr(env, PPCInstr_Load(4, tLo, four_r1, False/*mode32*/));
3774 add_to_sp( env, 16 );
3775 *rHi = tHi;
3776 *rLo = tLo;
3777 return;
3778 }
3779 case Iop_ExtractExpD128: {
3780 HReg r_srcHi;
3781 HReg r_srcLo;
3782 HReg tmp = newVRegF(env);
3783 HReg tLo = newVRegI(env);
3784 HReg tHi = newVRegI(env);
3785 PPCAMode* zero_r1 = PPCAMode_IR( 0, StackFramePtr(env->mode64) );
3786 PPCAMode* four_r1 = PPCAMode_IR( 4, StackFramePtr(env->mode64) );
3787
3788 iselDfp128Expr(&r_srcHi, &r_srcLo, env, e->Iex.Unop.arg, IEndianess);
3789 addInstr(env, PPCInstr_ExtractExpD128(Pfp_DXEXQ, tmp,
3790 r_srcHi, r_srcLo));
3791
3792 // put the D64 result into a integer register pair
3793 sub_from_sp( env, 16 );
3794 addInstr(env, PPCInstr_FpLdSt(False/*store*/, 8, tmp, zero_r1));
3795 addInstr(env, PPCInstr_Load(4, tHi, zero_r1, False/*mode32*/));
3796 addInstr(env, PPCInstr_Load(4, tLo, four_r1, False/*mode32*/));
3797 add_to_sp( env, 16 );
3798 *rHi = tHi;
3799 *rLo = tLo;
3800 return;
3801 }
3802
3803 /* 32Uto64(e) */
3804 case Iop_32Uto64: {
3805 HReg tHi = newVRegI(env);
3806 HReg tLo = iselWordExpr_R(env, e->Iex.Unop.arg, IEndianess);
3807 addInstr(env, PPCInstr_LI(tHi, 0, False/*mode32*/));
3808 *rHi = tHi;
3809 *rLo = tLo;
3810 return;
3811 }
3812
3813 case Iop_128to64: {
3814 /* Narrow, return the low 64-bit half as a 32-bit
3815 * register pair */
3816 HReg r_Hi = INVALID_HREG;
3817 HReg r_MedHi = INVALID_HREG;
3818 HReg r_MedLo = INVALID_HREG;
3819 HReg r_Lo = INVALID_HREG;
3820
3821 iselInt128Expr_to_32x4(&r_Hi, &r_MedHi, &r_MedLo, &r_Lo,
3822 env, e->Iex.Unop.arg, IEndianess);
3823 *rHi = r_MedLo;
3824 *rLo = r_Lo;
3825 return;
3826 }
3827
3828 case Iop_128HIto64: {
3829 /* Narrow, return the high 64-bit half as a 32-bit
3830 * register pair */
3831 HReg r_Hi = INVALID_HREG;
3832 HReg r_MedHi = INVALID_HREG;
3833 HReg r_MedLo = INVALID_HREG;
3834 HReg r_Lo = INVALID_HREG;
3835
3836 iselInt128Expr_to_32x4(&r_Hi, &r_MedHi, &r_MedLo, &r_Lo,
3837 env, e->Iex.Unop.arg, IEndianess);
3838 *rHi = r_Hi;
3839 *rLo = r_MedHi;
3840 return;
3841 }
3842
3843 /* V128{HI}to64 */
3844 case Iop_V128HIto64:
3845 case Iop_V128to64: {
3846 HReg r_aligned16;
3847 Int off = e->Iex.Unop.op==Iop_V128HIto64 ? 0 : 8;
3848 HReg tLo = newVRegI(env);
3849 HReg tHi = newVRegI(env);
3850 HReg vec = iselVecExpr(env, e->Iex.Unop.arg, IEndianess);
3851 PPCAMode *am_off0, *am_offLO, *am_offHI;
3852 sub_from_sp( env, 32 ); // Move SP down 32 bytes
3853
3854 // get a quadword aligned address within our stack space
3855 r_aligned16 = get_sp_aligned16( env );
3856 am_off0 = PPCAMode_IR( 0, r_aligned16 );
3857 am_offHI = PPCAMode_IR( off, r_aligned16 );
3858 am_offLO = PPCAMode_IR( off+4, r_aligned16 );
3859
3860 // store as Vec128
3861 addInstr(env,
3862 PPCInstr_AvLdSt( False/*store*/, 16, vec, am_off0 ));
3863
3864 // load hi,lo words (of hi/lo half of vec) as Ity_I32's
3865 addInstr(env,
3866 PPCInstr_Load( 4, tHi, am_offHI, False/*mode32*/ ));
3867 addInstr(env,
3868 PPCInstr_Load( 4, tLo, am_offLO, False/*mode32*/ ));
3869
3870 add_to_sp( env, 32 ); // Reset SP
3871 *rHi = tHi;
3872 *rLo = tLo;
3873 return;
3874 }
3875
3876 /* could do better than this, but for now ... */
3877 case Iop_1Sto64: {
3878 HReg tLo = newVRegI(env);
3879 HReg tHi = newVRegI(env);
3880 PPCCondCode cond = iselCondCode(env, e->Iex.Unop.arg, IEndianess);
3881 addInstr(env, PPCInstr_Set(cond,tLo));
3882 addInstr(env, PPCInstr_Shft(Pshft_SHL, True/*32bit shift*/,
3883 tLo, tLo, PPCRH_Imm(False,31)));
3884 addInstr(env, PPCInstr_Shft(Pshft_SAR, True/*32bit shift*/,
3885 tLo, tLo, PPCRH_Imm(False,31)));
3886 addInstr(env, mk_iMOVds_RR(tHi, tLo));
3887 *rHi = tHi;
3888 *rLo = tLo;
3889 return;
3890 }
3891
3892 case Iop_Not64: {
3893 HReg xLo, xHi;
3894 HReg tmpLo = newVRegI(env);
3895 HReg tmpHi = newVRegI(env);
3896 iselInt64Expr(&xHi, &xLo, env, e->Iex.Unop.arg, IEndianess);
3897 addInstr(env, PPCInstr_Unary(Pun_NOT,tmpLo,xLo));
3898 addInstr(env, PPCInstr_Unary(Pun_NOT,tmpHi,xHi));
3899 *rHi = tmpHi;
3900 *rLo = tmpLo;
3901 return;
3902 }
3903
3904 /* ReinterpF64asI64(e) */
3905 /* Given an IEEE754 double, produce an I64 with the same bit
3906 pattern. */
3907 case Iop_ReinterpF64asI64: {
3908 PPCAMode *am_addr0, *am_addr1;
3909 HReg fr_src = iselDblExpr(env, e->Iex.Unop.arg, IEndianess);
3910 HReg r_dstLo = newVRegI(env);
3911 HReg r_dstHi = newVRegI(env);
3912
3913 sub_from_sp( env, 16 ); // Move SP down 16 bytes
3914 am_addr0 = PPCAMode_IR( 0, StackFramePtr(False/*mode32*/) );
3915 am_addr1 = PPCAMode_IR( 4, StackFramePtr(False/*mode32*/) );
3916
3917 // store as F64
3918 addInstr(env, PPCInstr_FpLdSt( False/*store*/, 8,
3919 fr_src, am_addr0 ));
3920
3921 // load hi,lo as Ity_I32's
3922 addInstr(env, PPCInstr_Load( 4, r_dstHi,
3923 am_addr0, False/*mode32*/ ));
3924 addInstr(env, PPCInstr_Load( 4, r_dstLo,
3925 am_addr1, False/*mode32*/ ));
3926 *rHi = r_dstHi;
3927 *rLo = r_dstLo;
3928
3929 add_to_sp( env, 16 ); // Reset SP
3930 return;
3931 }
3932
3933 case Iop_ReinterpD64asI64: {
3934 HReg fr_src = iselDfp64Expr(env, e->Iex.Unop.arg, IEndianess);
3935 PPCAMode *am_addr0, *am_addr1;
3936 HReg r_dstLo = newVRegI(env);
3937 HReg r_dstHi = newVRegI(env);
3938
3939
3940 sub_from_sp( env, 16 ); // Move SP down 16 bytes
3941 am_addr0 = PPCAMode_IR( 0, StackFramePtr(False/*mode32*/) );
3942 am_addr1 = PPCAMode_IR( 4, StackFramePtr(False/*mode32*/) );
3943
3944 // store as D64
3945 addInstr(env, PPCInstr_FpLdSt( False/*store*/, 8,
3946 fr_src, am_addr0 ));
3947
3948 // load hi,lo as Ity_I32's
3949 addInstr(env, PPCInstr_Load( 4, r_dstHi,
3950 am_addr0, False/*mode32*/ ));
3951 addInstr(env, PPCInstr_Load( 4, r_dstLo,
3952 am_addr1, False/*mode32*/ ));
3953 *rHi = r_dstHi;
3954 *rLo = r_dstLo;
3955
3956 add_to_sp( env, 16 ); // Reset SP
3957
3958 return;
3959 }
3960
3961 case Iop_BCDtoDPB: {
3962 PPCCondCode cc;
3963 UInt argiregs;
3964 HReg argregs[2];
3965 Int argreg;
3966 HReg tLo = newVRegI(env);
3967 HReg tHi = newVRegI(env);
3968 HReg tmpHi;
3969 HReg tmpLo;
3970 Bool mode64 = env->mode64;
3971
3972 argregs[0] = hregPPC_GPR3(mode64);
3973 argregs[1] = hregPPC_GPR4(mode64);
3974
3975 argiregs = 0;
3976 argreg = 0;
3977
3978 iselInt64Expr( &tmpHi, &tmpLo, env, e->Iex.Unop.arg, IEndianess );
3979
3980 argiregs |= ( 1 << (argreg+3 ) );
3981 addInstr( env, mk_iMOVds_RR( argregs[argreg++], tmpHi ) );
3982
3983 argiregs |= ( 1 << (argreg+3 ) );
3984 addInstr( env, mk_iMOVds_RR( argregs[argreg], tmpLo ) );
3985
3986 cc = mk_PPCCondCode( Pct_ALWAYS, Pcf_NONE );
3987
3988 if (IEndianess == Iend_LE) {
3989 addInstr( env, PPCInstr_Call( cc, (Addr)h_calc_BCDtoDPB,
3990 argiregs,
3991 mk_RetLoc_simple(RLPri_2Int) ) );
3992 } else {
3993 Addr64 target;
3994 target = mode64 ? (Addr)h_calc_BCDtoDPB :
3995 toUInt( (Addr)h_calc_BCDtoDPB );
3996 addInstr( env, PPCInstr_Call( cc, target,
3997 argiregs,
3998 mk_RetLoc_simple(RLPri_2Int) ) );
3999 }
4000
4001 addInstr( env, mk_iMOVds_RR( tHi, argregs[argreg-1] ) );
4002 addInstr( env, mk_iMOVds_RR( tLo, argregs[argreg] ) );
4003
4004 *rHi = tHi;
4005 *rLo = tLo;
4006 return;
4007 }
4008
4009 case Iop_DPBtoBCD: {
4010 PPCCondCode cc;
4011 UInt argiregs;
4012 HReg argregs[2];
4013 Int argreg;
4014 HReg tLo = newVRegI(env);
4015 HReg tHi = newVRegI(env);
4016 HReg tmpHi;
4017 HReg tmpLo;
4018 Bool mode64 = env->mode64;
4019
4020 argregs[0] = hregPPC_GPR3(mode64);
4021 argregs[1] = hregPPC_GPR4(mode64);
4022
4023 argiregs = 0;
4024 argreg = 0;
4025
4026 iselInt64Expr(&tmpHi, &tmpLo, env, e->Iex.Unop.arg, IEndianess);
4027
4028 argiregs |= (1 << (argreg+3));
4029 addInstr(env, mk_iMOVds_RR( argregs[argreg++], tmpHi ));
4030
4031 argiregs |= (1 << (argreg+3));
4032 addInstr(env, mk_iMOVds_RR( argregs[argreg], tmpLo));
4033
4034 cc = mk_PPCCondCode( Pct_ALWAYS, Pcf_NONE );
4035
4036 if (IEndianess == Iend_LE) {
4037 addInstr(env, PPCInstr_Call( cc, (Addr)h_calc_DPBtoBCD,
4038 argiregs,
4039 mk_RetLoc_simple(RLPri_2Int) ) );
4040 } else {
4041 Addr64 target;
4042 target = mode64 ? (Addr)h_calc_DPBtoBCD :
4043 toUInt( (Addr)h_calc_DPBtoBCD );
4044 addInstr(env, PPCInstr_Call( cc, target, argiregs,
4045 mk_RetLoc_simple(RLPri_2Int) ) );
4046 }
4047
4048 addInstr(env, mk_iMOVds_RR(tHi, argregs[argreg-1]));
4049 addInstr(env, mk_iMOVds_RR(tLo, argregs[argreg]));
4050
4051 *rHi = tHi;
4052 *rLo = tLo;
4053 return;
4054 }
4055
4056 default:
4057 break;
4058 }
4059 } /* if (e->tag == Iex_Unop) */
4060
4061 vex_printf("iselInt64Expr(ppc): No such tag(%u)\n", e->tag);
4062 ppIRExpr(e);
4063 vpanic("iselInt64Expr(ppc)");
4064 }
4065
4066
4067 /*---------------------------------------------------------*/
4068 /*--- ISEL: Floating point expressions (32 bit) ---*/
4069 /*---------------------------------------------------------*/
4070
4071 /* Nothing interesting here; really just wrappers for
4072 64-bit stuff. */
4073
iselFltExpr(ISelEnv * env,const IRExpr * e,IREndness IEndianess)4074 static HReg iselFltExpr ( ISelEnv* env, const IRExpr* e, IREndness IEndianess )
4075 {
4076 HReg r = iselFltExpr_wrk( env, e, IEndianess );
4077 # if 0
4078 vex_printf("\n"); ppIRExpr(e); vex_printf("\n");
4079 # endif
4080 vassert(hregClass(r) == HRcFlt64); /* yes, really Flt64 */
4081 vassert(hregIsVirtual(r));
4082 return r;
4083 }
4084
4085 /* DO NOT CALL THIS DIRECTLY */
iselFltExpr_wrk(ISelEnv * env,const IRExpr * e,IREndness IEndianess)4086 static HReg iselFltExpr_wrk ( ISelEnv* env, const IRExpr* e,
4087 IREndness IEndianess )
4088 {
4089 Bool mode64 = env->mode64;
4090
4091 IRType ty = typeOfIRExpr(env->type_env,e);
4092 vassert(ty == Ity_F32);
4093
4094 if (e->tag == Iex_RdTmp) {
4095 return lookupIRTemp(env, e->Iex.RdTmp.tmp);
4096 }
4097
4098 if (e->tag == Iex_Load && e->Iex.Load.end == IEndianess) {
4099 PPCAMode* am_addr;
4100 HReg r_dst = newVRegF(env);
4101 vassert(e->Iex.Load.ty == Ity_F32);
4102 am_addr = iselWordExpr_AMode(env, e->Iex.Load.addr, Ity_F32/*xfer*/,
4103 IEndianess);
4104 addInstr(env, PPCInstr_FpLdSt(True/*load*/, 4, r_dst, am_addr));
4105 return r_dst;
4106 }
4107
4108 if (e->tag == Iex_Get) {
4109 HReg r_dst = newVRegF(env);
4110 PPCAMode* am_addr = PPCAMode_IR( e->Iex.Get.offset,
4111 GuestStatePtr(env->mode64) );
4112 addInstr(env, PPCInstr_FpLdSt( True/*load*/, 4, r_dst, am_addr ));
4113 return r_dst;
4114 }
4115
4116 if (e->tag == Iex_Unop && e->Iex.Unop.op == Iop_TruncF64asF32) {
4117 /* This is quite subtle. The only way to do the relevant
4118 truncation is to do a single-precision store and then a
4119 double precision load to get it back into a register. The
4120 problem is, if the data is then written to memory a second
4121 time, as in
4122
4123 STbe(...) = TruncF64asF32(...)
4124
4125 then will the second truncation further alter the value? The
4126 answer is no: flds (as generated here) followed by fsts
4127 (generated for the STbe) is the identity function on 32-bit
4128 floats, so we are safe.
4129
4130 Another upshot of this is that if iselStmt can see the
4131 entirety of
4132
4133 STbe(...) = TruncF64asF32(arg)
4134
4135 then it can short circuit having to deal with TruncF64asF32
4136 individually; instead just compute arg into a 64-bit FP
4137 register and do 'fsts' (since that itself does the
4138 truncation).
4139
4140 We generate pretty poor code here (should be ok both for
4141 32-bit and 64-bit mode); but it is expected that for the most
4142 part the latter optimisation will apply and hence this code
4143 will not often be used.
4144 */
4145 HReg fsrc = iselDblExpr(env, e->Iex.Unop.arg, IEndianess);
4146 HReg fdst = newVRegF(env);
4147 PPCAMode* zero_r1 = PPCAMode_IR( 0, StackFramePtr(env->mode64) );
4148
4149 sub_from_sp( env, 16 );
4150 // store as F32, hence truncating
4151 addInstr(env, PPCInstr_FpLdSt( False/*store*/, 4,
4152 fsrc, zero_r1 ));
4153 // and reload. Good huh?! (sigh)
4154 addInstr(env, PPCInstr_FpLdSt( True/*load*/, 4,
4155 fdst, zero_r1 ));
4156 add_to_sp( env, 16 );
4157 return fdst;
4158 }
4159
4160 if (e->tag == Iex_Binop && e->Iex.Binop.op == Iop_I64UtoF32) {
4161 if (mode64) {
4162 HReg fdst = newVRegF(env);
4163 HReg isrc = iselWordExpr_R(env, e->Iex.Binop.arg2, IEndianess);
4164 HReg r1 = StackFramePtr(env->mode64);
4165 PPCAMode* zero_r1 = PPCAMode_IR( 0, r1 );
4166
4167 /* Set host rounding mode */
4168 set_FPU_rounding_mode( env, e->Iex.Binop.arg1, IEndianess );
4169
4170 sub_from_sp( env, 16 );
4171
4172 addInstr(env, PPCInstr_Store(8, zero_r1, isrc, True/*mode64*/));
4173 addInstr(env, PPCInstr_FpLdSt(True/*load*/, 8, fdst, zero_r1));
4174 addInstr(env, PPCInstr_FpCftI(True/*I->F*/, False/*int64*/,
4175 False, False,
4176 fdst, fdst));
4177
4178 add_to_sp( env, 16 );
4179
4180 ///* Restore default FPU rounding. */
4181 //set_FPU_rounding_default( env );
4182 return fdst;
4183 } else {
4184 /* 32-bit mode */
4185 HReg fdst = newVRegF(env);
4186 HReg isrcHi, isrcLo;
4187 HReg r1 = StackFramePtr(env->mode64);
4188 PPCAMode* zero_r1 = PPCAMode_IR( 0, r1 );
4189 PPCAMode* four_r1 = PPCAMode_IR( 4, r1 );
4190
4191 iselInt64Expr(&isrcHi, &isrcLo, env, e->Iex.Binop.arg2, IEndianess);
4192
4193 /* Set host rounding mode */
4194 set_FPU_rounding_mode( env, e->Iex.Binop.arg1, IEndianess );
4195
4196 sub_from_sp( env, 16 );
4197
4198 addInstr(env, PPCInstr_Store(4, zero_r1, isrcHi, False/*mode32*/));
4199 addInstr(env, PPCInstr_Store(4, four_r1, isrcLo, False/*mode32*/));
4200 addInstr(env, PPCInstr_FpLdSt(True/*load*/, 8, fdst, zero_r1));
4201 addInstr(env, PPCInstr_FpCftI(True/*I->F*/, False/*int64*/,
4202 False, False,
4203 fdst, fdst));
4204
4205 add_to_sp( env, 16 );
4206
4207 ///* Restore default FPU rounding. */
4208 //set_FPU_rounding_default( env );
4209 return fdst;
4210 }
4211
4212 }
4213
4214 vex_printf("iselFltExpr(ppc): No such tag(%u)\n", e->tag);
4215 ppIRExpr(e);
4216 vpanic("iselFltExpr_wrk(ppc)");
4217 }
4218
4219
4220 /*---------------------------------------------------------*/
4221 /*--- ISEL: Floating point expressions (64 bit) ---*/
4222 /*---------------------------------------------------------*/
4223
4224 /* Compute a 64-bit floating point value into a register, the identity
4225 of which is returned. As with iselIntExpr_R, the reg may be either
4226 real or virtual; in any case it must not be changed by subsequent
4227 code emitted by the caller. */
4228
4229 /* IEEE 754 formats. From http://www.freesoft.org/CIE/RFC/1832/32.htm:
4230
4231 Type S (1 bit) E (11 bits) F (52 bits)
4232 ---- --------- ----------- -----------
4233 signalling NaN u 2047 (max) .0uuuuu---u
4234 (with at least
4235 one 1 bit)
4236 quiet NaN u 2047 (max) .1uuuuu---u
4237
4238 negative infinity 1 2047 (max) .000000---0
4239
4240 positive infinity 0 2047 (max) .000000---0
4241
4242 negative zero 1 0 .000000---0
4243
4244 positive zero 0 0 .000000---0
4245 */
4246
iselDblExpr(ISelEnv * env,const IRExpr * e,IREndness IEndianess)4247 static HReg iselDblExpr ( ISelEnv* env, const IRExpr* e, IREndness IEndianess )
4248 {
4249 HReg r = iselDblExpr_wrk( env, e, IEndianess );
4250 # if 0
4251 vex_printf("\n"); ppIRExpr(e); vex_printf("\n");
4252 # endif
4253 vassert(hregClass(r) == HRcFlt64);
4254 vassert(hregIsVirtual(r));
4255 return r;
4256 }
4257
4258 /* DO NOT CALL THIS DIRECTLY */
iselDblExpr_wrk(ISelEnv * env,const IRExpr * e,IREndness IEndianess)4259 static HReg iselDblExpr_wrk ( ISelEnv* env, const IRExpr* e,
4260 IREndness IEndianess )
4261 {
4262 Bool mode64 = env->mode64;
4263 IRType ty = typeOfIRExpr(env->type_env,e);
4264 vassert(e);
4265 vassert(ty == Ity_F64);
4266
4267 if (e->tag == Iex_RdTmp) {
4268 return lookupIRTemp(env, e->Iex.RdTmp.tmp);
4269 }
4270
4271 /* --------- LITERAL --------- */
4272 if (e->tag == Iex_Const) {
4273 union { UInt u32x2[2]; ULong u64; Double f64; } u;
4274 vassert(sizeof(u) == 8);
4275 vassert(sizeof(u.u64) == 8);
4276 vassert(sizeof(u.f64) == 8);
4277 vassert(sizeof(u.u32x2) == 8);
4278
4279 if (e->Iex.Const.con->tag == Ico_F64) {
4280 u.f64 = e->Iex.Const.con->Ico.F64;
4281 }
4282 else if (e->Iex.Const.con->tag == Ico_F64i) {
4283 u.u64 = e->Iex.Const.con->Ico.F64i;
4284 }
4285 else
4286 vpanic("iselDblExpr(ppc): const");
4287
4288 if (!mode64) {
4289 HReg r_srcHi = newVRegI(env);
4290 HReg r_srcLo = newVRegI(env);
4291 addInstr(env, PPCInstr_LI(r_srcHi, u.u32x2[0], mode64));
4292 addInstr(env, PPCInstr_LI(r_srcLo, u.u32x2[1], mode64));
4293 return mk_LoadRR32toFPR( env, r_srcHi, r_srcLo );
4294 } else { // mode64
4295 HReg r_src = newVRegI(env);
4296 addInstr(env, PPCInstr_LI(r_src, u.u64, mode64));
4297 return mk_LoadR64toFPR( env, r_src ); // 1*I64 -> F64
4298 }
4299 }
4300
4301 /* --------- LOAD --------- */
4302 if (e->tag == Iex_Load && e->Iex.Load.end == IEndianess) {
4303 HReg r_dst = newVRegF(env);
4304 PPCAMode* am_addr;
4305 vassert(e->Iex.Load.ty == Ity_F64);
4306 am_addr = iselWordExpr_AMode(env, e->Iex.Load.addr, Ity_F64/*xfer*/,
4307 IEndianess);
4308 addInstr(env, PPCInstr_FpLdSt(True/*load*/, 8, r_dst, am_addr));
4309 return r_dst;
4310 }
4311
4312 /* --------- GET --------- */
4313 if (e->tag == Iex_Get) {
4314 HReg r_dst = newVRegF(env);
4315 PPCAMode* am_addr = PPCAMode_IR( e->Iex.Get.offset,
4316 GuestStatePtr(mode64) );
4317 addInstr(env, PPCInstr_FpLdSt( True/*load*/, 8, r_dst, am_addr ));
4318 return r_dst;
4319 }
4320
4321 /* --------- OPS --------- */
4322 if (e->tag == Iex_Qop) {
4323 PPCFpOp fpop = Pfp_INVALID;
4324 switch (e->Iex.Qop.details->op) {
4325 case Iop_MAddF64: fpop = Pfp_MADDD; break;
4326 case Iop_MAddF64r32: fpop = Pfp_MADDS; break;
4327 case Iop_MSubF64: fpop = Pfp_MSUBD; break;
4328 case Iop_MSubF64r32: fpop = Pfp_MSUBS; break;
4329 default: break;
4330 }
4331 if (fpop != Pfp_INVALID) {
4332 HReg r_dst = newVRegF(env);
4333 HReg r_srcML = iselDblExpr(env, e->Iex.Qop.details->arg2,
4334 IEndianess);
4335 HReg r_srcMR = iselDblExpr(env, e->Iex.Qop.details->arg3,
4336 IEndianess);
4337 HReg r_srcAcc = iselDblExpr(env, e->Iex.Qop.details->arg4,
4338 IEndianess);
4339 set_FPU_rounding_mode( env, e->Iex.Qop.details->arg1, IEndianess );
4340 addInstr(env, PPCInstr_FpMulAcc(fpop, r_dst,
4341 r_srcML, r_srcMR, r_srcAcc));
4342 return r_dst;
4343 }
4344 }
4345
4346 if (e->tag == Iex_Triop) {
4347 IRTriop *triop = e->Iex.Triop.details;
4348 PPCFpOp fpop = Pfp_INVALID;
4349 switch (triop->op) {
4350 case Iop_AddF64: fpop = Pfp_ADDD; break;
4351 case Iop_SubF64: fpop = Pfp_SUBD; break;
4352 case Iop_MulF64: fpop = Pfp_MULD; break;
4353 case Iop_DivF64: fpop = Pfp_DIVD; break;
4354 case Iop_AddF64r32: fpop = Pfp_ADDS; break;
4355 case Iop_SubF64r32: fpop = Pfp_SUBS; break;
4356 case Iop_MulF64r32: fpop = Pfp_MULS; break;
4357 case Iop_DivF64r32: fpop = Pfp_DIVS; break;
4358 default: break;
4359 }
4360 if (fpop != Pfp_INVALID) {
4361 HReg r_dst = newVRegF(env);
4362 HReg r_srcL = iselDblExpr(env, triop->arg2, IEndianess);
4363 HReg r_srcR = iselDblExpr(env, triop->arg3, IEndianess);
4364 set_FPU_rounding_mode( env, triop->arg1, IEndianess );
4365 addInstr(env, PPCInstr_FpBinary(fpop, r_dst, r_srcL, r_srcR));
4366 return r_dst;
4367 }
4368 }
4369
4370 if (e->tag == Iex_Binop) {
4371 PPCFpOp fpop = Pfp_INVALID;
4372 switch (e->Iex.Binop.op) {
4373 case Iop_SqrtF64: fpop = Pfp_SQRT; break;
4374 default: break;
4375 }
4376 if (fpop == Pfp_SQRT) {
4377 HReg fr_dst = newVRegF(env);
4378 HReg fr_src = iselDblExpr(env, e->Iex.Binop.arg2, IEndianess);
4379 set_FPU_rounding_mode( env, e->Iex.Binop.arg1, IEndianess );
4380 addInstr(env, PPCInstr_FpUnary(fpop, fr_dst, fr_src));
4381 return fr_dst;
4382 }
4383 }
4384
4385 if (e->tag == Iex_Binop) {
4386
4387 if (e->Iex.Binop.op == Iop_F128toF64) {
4388 HReg fr_dst = newVRegF(env);
4389 HReg fr_src = iselFp128Expr(env, e->Iex.Binop.arg2, IEndianess);
4390 HReg tmp = newVRegV(env);
4391 PPCAMode* zero_r1 = PPCAMode_IR( 0, StackFramePtr(env->mode64) );
4392 PPCAMode* eight_r1 = PPCAMode_IR( 8, StackFramePtr(env->mode64) );
4393 PPCFpOp fpop = Pfp_INVALID;
4394
4395 if (FPU_rounding_mode_isOdd(e->Iex.Binop.arg1)) {
4396 /* use rounding mode specified by RN. Issue inst with R0 = 0 */
4397 fpop = Pfp_FPQTODRNDODD;
4398 } else {
4399 set_FPU_rounding_mode( env, e->Iex.Binop.arg1, IEndianess );
4400 fpop = Pfp_FPQTOD;
4401 }
4402
4403 addInstr(env, PPCInstr_Fp128Unary(fpop, tmp, fr_src));
4404
4405 /* result is in a 128-bit vector register, move to 64-bit reg to
4406 * match the Iop specification. The result will get moved back
4407 * to a 128-bit register and stored once the value is returned.
4408 */
4409 sub_from_sp( env, 16 );
4410 addInstr(env, PPCInstr_AvLdSt(False/*store*/, 16, tmp, zero_r1));
4411 if (IEndianess == Iend_LE)
4412 addInstr(env, PPCInstr_FpLdSt(True/*load*/, 8, fr_dst, eight_r1));
4413 else
4414 /* High 64-bits stored at lower address */
4415 addInstr(env, PPCInstr_FpLdSt(True/*load*/, 8, fr_dst, zero_r1));
4416
4417 add_to_sp( env, 16 );
4418
4419 return fr_dst;
4420 }
4421
4422 if (e->Iex.Binop.op == Iop_RoundF64toF32) {
4423 HReg r_dst = newVRegF(env);
4424 HReg r_src = iselDblExpr(env, e->Iex.Binop.arg2, IEndianess);
4425 set_FPU_rounding_mode( env, e->Iex.Binop.arg1, IEndianess );
4426 addInstr(env, PPCInstr_FpRSP(r_dst, r_src));
4427 //set_FPU_rounding_default( env );
4428 return r_dst;
4429 }
4430
4431 if (e->Iex.Binop.op == Iop_I64StoF64 || e->Iex.Binop.op == Iop_I64UtoF64) {
4432 if (mode64) {
4433 HReg fdst = newVRegF(env);
4434 HReg isrc = iselWordExpr_R(env, e->Iex.Binop.arg2, IEndianess);
4435 HReg r1 = StackFramePtr(env->mode64);
4436 PPCAMode* zero_r1 = PPCAMode_IR( 0, r1 );
4437
4438 /* Set host rounding mode */
4439 set_FPU_rounding_mode( env, e->Iex.Binop.arg1, IEndianess );
4440
4441 sub_from_sp( env, 16 );
4442
4443 addInstr(env, PPCInstr_Store(8, zero_r1, isrc, True/*mode64*/));
4444 addInstr(env, PPCInstr_FpLdSt(True/*load*/, 8, fdst, zero_r1));
4445 addInstr(env, PPCInstr_FpCftI(True/*I->F*/, False/*int64*/,
4446 e->Iex.Binop.op == Iop_I64StoF64,
4447 True/*fdst is 64 bit*/,
4448 fdst, fdst));
4449
4450 add_to_sp( env, 16 );
4451
4452 ///* Restore default FPU rounding. */
4453 //set_FPU_rounding_default( env );
4454 return fdst;
4455 } else {
4456 /* 32-bit mode */
4457 HReg fdst = newVRegF(env);
4458 HReg isrcHi, isrcLo;
4459 HReg r1 = StackFramePtr(env->mode64);
4460 PPCAMode* zero_r1 = PPCAMode_IR( 0, r1 );
4461 PPCAMode* four_r1 = PPCAMode_IR( 4, r1 );
4462
4463 iselInt64Expr(&isrcHi, &isrcLo, env, e->Iex.Binop.arg2,
4464 IEndianess);
4465
4466 /* Set host rounding mode */
4467 set_FPU_rounding_mode( env, e->Iex.Binop.arg1, IEndianess );
4468
4469 sub_from_sp( env, 16 );
4470
4471 addInstr(env, PPCInstr_Store(4, zero_r1, isrcHi, False/*mode32*/));
4472 addInstr(env, PPCInstr_Store(4, four_r1, isrcLo, False/*mode32*/));
4473 addInstr(env, PPCInstr_FpLdSt(True/*load*/, 8, fdst, zero_r1));
4474 addInstr(env, PPCInstr_FpCftI(True/*I->F*/, False/*int64*/,
4475 e->Iex.Binop.op == Iop_I64StoF64,
4476 True/*fdst is 64 bit*/,
4477 fdst, fdst));
4478
4479 add_to_sp( env, 16 );
4480
4481 ///* Restore default FPU rounding. */
4482 //set_FPU_rounding_default( env );
4483 return fdst;
4484 }
4485 }
4486
4487 }
4488
4489 if (e->tag == Iex_Unop) {
4490 PPCFpOp fpop = Pfp_INVALID;
4491 switch (e->Iex.Unop.op) {
4492 case Iop_NegF64: fpop = Pfp_NEG; break;
4493 case Iop_AbsF64: fpop = Pfp_ABS; break;
4494 case Iop_RSqrtEst5GoodF64: fpop = Pfp_RSQRTE; break;
4495 case Iop_RoundF64toF64_NegINF: fpop = Pfp_FRIM; break;
4496 case Iop_RoundF64toF64_PosINF: fpop = Pfp_FRIP; break;
4497 case Iop_RoundF64toF64_NEAREST: fpop = Pfp_FRIN; break;
4498 case Iop_RoundF64toF64_ZERO: fpop = Pfp_FRIZ; break;
4499 default: break;
4500 }
4501 if (fpop != Pfp_INVALID) {
4502 HReg fr_dst = newVRegF(env);
4503 HReg fr_src = iselDblExpr(env, e->Iex.Unop.arg, IEndianess);
4504 addInstr(env, PPCInstr_FpUnary(fpop, fr_dst, fr_src));
4505 return fr_dst;
4506 }
4507 }
4508
4509 if (e->tag == Iex_Unop) {
4510 switch (e->Iex.Unop.op) {
4511 case Iop_F128HItoF64:
4512 case Iop_F128LOtoF64:
4513 {
4514 /* put upper/lower 64-bits of F128 into an F64. */
4515 HReg r_aligned16;
4516 HReg fdst = newVRegF(env);
4517 HReg fsrc = iselFp128Expr(env, e->Iex.Unop.arg, IEndianess);
4518 PPCAMode *am_off0, *am_off8, *am_off_arg;
4519 sub_from_sp( env, 32 ); // Move SP down 32 bytes
4520
4521 // get a quadword aligned address within our stack space
4522 r_aligned16 = get_sp_aligned16( env );
4523 am_off0 = PPCAMode_IR( 0, r_aligned16 );
4524 am_off8 = PPCAMode_IR( 8 ,r_aligned16 );
4525
4526 /* store 128-bit floating point value to memory, load low word
4527 * or high to 64-bit destination floating point register
4528 */
4529 addInstr(env, PPCInstr_AvLdSt(False/*store*/, 16, fsrc, am_off0));
4530 if (IEndianess == Iend_LE) {
4531 if (e->Iex.Binop.op == Iop_F128HItoF64)
4532 am_off_arg = am_off8;
4533 else
4534 am_off_arg = am_off0;
4535 } else {
4536 if (e->Iex.Binop.op == Iop_F128HItoF64)
4537 am_off_arg = am_off0;
4538 else
4539 am_off_arg = am_off8;
4540 }
4541 addInstr(env,
4542 PPCInstr_FpLdSt( True /*load*/,
4543 8, fdst,
4544 am_off_arg ));
4545 add_to_sp( env, 32 ); // Reset SP
4546 return fdst;
4547 }
4548 case Iop_ReinterpI64asF64: {
4549 /* Given an I64, produce an IEEE754 double with the same
4550 bit pattern. */
4551 if (!mode64) {
4552 HReg r_srcHi, r_srcLo;
4553 iselInt64Expr( &r_srcHi, &r_srcLo, env, e->Iex.Unop.arg,
4554 IEndianess);
4555 return mk_LoadRR32toFPR( env, r_srcHi, r_srcLo );
4556 } else {
4557 HReg r_src = iselWordExpr_R(env, e->Iex.Unop.arg, IEndianess);
4558 return mk_LoadR64toFPR( env, r_src );
4559 }
4560 }
4561
4562 case Iop_F32toF64: {
4563 if (e->Iex.Unop.arg->tag == Iex_Unop &&
4564 e->Iex.Unop.arg->Iex.Unop.op == Iop_ReinterpI32asF32 ) {
4565 e = e->Iex.Unop.arg;
4566
4567 HReg src = iselWordExpr_R(env, e->Iex.Unop.arg, IEndianess);
4568 HReg fr_dst = newVRegF(env);
4569 PPCAMode *am_addr;
4570
4571 sub_from_sp( env, 16 ); // Move SP down 16 bytes
4572 am_addr = PPCAMode_IR( 0, StackFramePtr(env->mode64) );
4573
4574 // store src as Ity_I32's
4575 addInstr(env, PPCInstr_Store( 4, am_addr, src, env->mode64 ));
4576
4577 // load single precision float, but the end results loads into a
4578 // 64-bit FP register -- i.e., F64.
4579 addInstr(env, PPCInstr_FpLdSt(True/*load*/, 4, fr_dst, am_addr));
4580
4581 add_to_sp( env, 16 ); // Reset SP
4582 return fr_dst;
4583 }
4584
4585
4586 /* this is a no-op */
4587 HReg res = iselFltExpr(env, e->Iex.Unop.arg, IEndianess);
4588 return res;
4589 }
4590 default:
4591 break;
4592 }
4593 }
4594
4595 /* --------- MULTIPLEX --------- */
4596 if (e->tag == Iex_ITE) { // VFD
4597 if (ty == Ity_F64
4598 && typeOfIRExpr(env->type_env,e->Iex.ITE.cond) == Ity_I1) {
4599 HReg fr1 = iselDblExpr(env, e->Iex.ITE.iftrue, IEndianess);
4600 HReg fr0 = iselDblExpr(env, e->Iex.ITE.iffalse, IEndianess);
4601 HReg fr_dst = newVRegF(env);
4602 addInstr(env, PPCInstr_FpUnary( Pfp_MOV, fr_dst, fr0 ));
4603 PPCCondCode cc = iselCondCode(env, e->Iex.ITE.cond, IEndianess);
4604 addInstr(env, PPCInstr_FpCMov( cc, fr_dst, fr1 ));
4605 return fr_dst;
4606 }
4607 }
4608
4609 vex_printf("iselDblExpr(ppc): No such tag(%u)\n", e->tag);
4610 ppIRExpr(e);
4611 vpanic("iselDblExpr_wrk(ppc)");
4612 }
4613
iselDfp32Expr(ISelEnv * env,const IRExpr * e,IREndness IEndianess)4614 static HReg iselDfp32Expr(ISelEnv* env, const IRExpr* e, IREndness IEndianess)
4615 {
4616 HReg r = iselDfp32Expr_wrk( env, e, IEndianess );
4617 vassert(hregClass(r) == HRcFlt64);
4618 vassert( hregIsVirtual(r) );
4619 return r;
4620 }
4621
4622 /* DO NOT CALL THIS DIRECTLY */
iselDfp32Expr_wrk(ISelEnv * env,const IRExpr * e,IREndness IEndianess)4623 static HReg iselDfp32Expr_wrk(ISelEnv* env, const IRExpr* e,
4624 IREndness IEndianess)
4625 {
4626 Bool mode64 = env->mode64;
4627 IRType ty = typeOfIRExpr( env->type_env, e );
4628
4629 vassert( e );
4630 vassert( ty == Ity_D32 );
4631
4632 /* --------- GET --------- */
4633 if (e->tag == Iex_Get) {
4634 HReg r_dst = newVRegF( env );
4635 PPCAMode* am_addr = PPCAMode_IR( e->Iex.Get.offset,
4636 GuestStatePtr(mode64) );
4637 addInstr( env, PPCInstr_FpLdSt( True/*load*/, 8, r_dst, am_addr ) );
4638 return r_dst;
4639 }
4640
4641 /* --------- LOAD --------- */
4642 if (e->tag == Iex_Load && e->Iex.Load.end == IEndianess) {
4643 PPCAMode* am_addr;
4644 HReg r_dst = newVRegF(env);
4645 vassert(e->Iex.Load.ty == Ity_D32);
4646 am_addr = iselWordExpr_AMode(env, e->Iex.Load.addr, Ity_D32/*xfer*/,
4647 IEndianess);
4648 addInstr(env, PPCInstr_FpLdSt(True/*load*/, 4, r_dst, am_addr));
4649 return r_dst;
4650 }
4651
4652 /* --------- OPS --------- */
4653 if (e->tag == Iex_Binop) {
4654 if (e->Iex.Binop.op == Iop_D64toD32) {
4655 HReg fr_dst = newVRegF(env);
4656 HReg fr_src = iselDfp64Expr(env, e->Iex.Binop.arg2, IEndianess);
4657 set_FPU_DFP_rounding_mode( env, e->Iex.Binop.arg1, IEndianess );
4658 addInstr(env, PPCInstr_Dfp64Unary(Pfp_DRSP, fr_dst, fr_src));
4659 return fr_dst;
4660 }
4661 }
4662
4663 ppIRExpr( e );
4664 vpanic( "iselDfp32Expr_wrk(ppc)" );
4665 }
4666
iselFp128Expr(ISelEnv * env,const IRExpr * e,IREndness IEndianess)4667 static HReg iselFp128Expr( ISelEnv* env, const IRExpr* e, IREndness IEndianess )
4668 {
4669 HReg r = iselFp128Expr_wrk( env, e, IEndianess );
4670 vassert(hregClass(r) == HRcVec128);
4671 vassert(hregIsVirtual(r));
4672 return r;
4673 }
4674
4675 /* DO NOT CALL THIS DIRECTLY */
iselFp128Expr_wrk(ISelEnv * env,const IRExpr * e,IREndness IEndianess)4676 static HReg iselFp128Expr_wrk( ISelEnv* env, const IRExpr* e,
4677 IREndness IEndianess)
4678 {
4679 Bool mode64 = env->mode64;
4680 PPCFpOp fpop = Pfp_INVALID;
4681 IRType ty = typeOfIRExpr(env->type_env,e);
4682
4683 vassert(e);
4684 vassert( ty == Ity_F128 );
4685
4686 /* read 128-bit IRTemp */
4687 if (e->tag == Iex_RdTmp) {
4688 return lookupIRTemp(env, e->Iex.RdTmp.tmp);
4689 }
4690
4691 if (e->tag == Iex_Get) {
4692 /* Guest state vectors are 16byte aligned,
4693 so don't need to worry here */
4694 HReg dst = newVRegV(env);
4695
4696 addInstr(env,
4697 PPCInstr_AvLdSt( True/*load*/, 16, dst,
4698 PPCAMode_IR( e->Iex.Get.offset,
4699 GuestStatePtr(mode64) )));
4700 return dst;
4701 }
4702
4703 if (e->tag == Iex_Unop) {
4704 switch (e->Iex.Unop.op) {
4705 case Iop_TruncF128toI64S:
4706 fpop = Pfp_TRUNCFPQTOISD; goto do_Un_F128;
4707 case Iop_TruncF128toI32S:
4708 fpop = Pfp_TRUNCFPQTOISW; goto do_Un_F128;
4709 case Iop_TruncF128toI64U:
4710 fpop = Pfp_TRUNCFPQTOIUD; goto do_Un_F128;
4711 case Iop_TruncF128toI32U:
4712 fpop = Pfp_TRUNCFPQTOIUW; goto do_Un_F128;
4713
4714 do_Un_F128: {
4715 HReg r_dst = newVRegV(env);
4716 HReg r_src = iselFp128Expr(env, e->Iex.Unop.arg, IEndianess);
4717 addInstr(env, PPCInstr_Fp128Unary(fpop, r_dst, r_src));
4718 return r_dst;
4719 }
4720
4721 case Iop_F64toF128: {
4722 fpop = Pfp_FPDTOQ;
4723 HReg r_dst = newVRegV(env);
4724 HReg r_src = iselDblExpr(env, e->Iex.Unop.arg, IEndianess);
4725 HReg v128tmp = newVRegV(env);
4726 PPCAMode* zero_r1 = PPCAMode_IR( 0, StackFramePtr(env->mode64) );
4727
4728 /* value is in 64-bit float reg, need to move to 128-bit vector reg */
4729 sub_from_sp( env, 16 );
4730 addInstr(env, PPCInstr_FpLdSt(False/*store*/, 8, r_src, zero_r1));
4731 addInstr(env, PPCInstr_AvLdSt(True/*load*/, 16, v128tmp, zero_r1));
4732 add_to_sp( env, 16 );
4733
4734 addInstr(env, PPCInstr_Fp128Unary(fpop, r_dst, v128tmp));
4735 return r_dst;
4736 }
4737
4738 case Iop_I64StoF128:
4739 fpop = Pfp_IDSTOQ; goto do_Un_int_F128;
4740 case Iop_I64UtoF128:
4741 fpop = Pfp_IDUTOQ; goto do_Un_int_F128;
4742
4743 do_Un_int_F128: {
4744 HReg r_dst = newVRegV(env);
4745 HReg tmp = newVRegV(env);
4746 HReg r_src = iselWordExpr_R(env, e->Iex.Unop.arg, IEndianess);
4747 PPCAMode *am_offhi, *am_offlo;
4748 HReg r_aligned16;
4749
4750 /* source is in a 64-bit integer reg, move to 128-bit float reg
4751 * do this via the stack (easy, convenient, etc).
4752 */
4753 sub_from_sp( env, 32 ); // Move SP down
4754
4755 /* Get a quadword aligned address within our stack space */
4756 r_aligned16 = get_sp_aligned16( env );
4757
4758 am_offlo = PPCAMode_IR( 0, r_aligned16 );
4759 am_offhi = PPCAMode_IR( 8, r_aligned16 );
4760
4761 /* Inst only uses the upper 64-bit of the source */
4762 addInstr(env, PPCInstr_Load(8, r_src, am_offhi, mode64));
4763
4764 /* Fetch result back from stack. */
4765 addInstr(env, PPCInstr_AvLdSt(True/*load*/, 16, tmp, am_offlo));
4766
4767 add_to_sp( env, 32 ); // Reset SP
4768
4769 addInstr(env, PPCInstr_Fp128Unary(fpop, r_dst, tmp));
4770 return r_dst;
4771 }
4772
4773 default:
4774 break;
4775 } /* switch (e->Iex.Unop.op) */
4776 } /* if (e->tag == Iex_Unop) */
4777
4778 if (e->tag == Iex_Binop) {
4779 switch (e->Iex.Binop.op) {
4780
4781 case Iop_F64HLtoF128:
4782 {
4783 HReg dst = newVRegV(env);
4784 HReg r_src_hi = iselDblExpr(env, e->Iex.Binop.arg1, IEndianess);
4785 HReg r_src_lo = iselDblExpr(env, e->Iex.Binop.arg2, IEndianess);
4786 PPCAMode *am_offhi, *am_offlo;
4787 HReg r_aligned16;
4788
4789 /* do this via the stack (easy, convenient, etc) */
4790 sub_from_sp( env, 16 ); // Move SP down
4791
4792 /* Get a quadword aligned address within our stack space */
4793 r_aligned16 = get_sp_aligned16( env );
4794
4795 am_offlo = PPCAMode_IR( 0, r_aligned16 );
4796 am_offhi = PPCAMode_IR( 8, r_aligned16 );
4797
4798 addInstr(env, PPCInstr_FpLdSt(False/*store*/, 8,
4799 r_src_lo, am_offlo));
4800 addInstr(env, PPCInstr_FpLdSt(False/*store*/, 8,
4801 r_src_hi, am_offhi));
4802
4803 /* Fetch result back from stack. */
4804 addInstr(env, PPCInstr_AvLdSt(True/*load*/, 16,
4805 dst, am_offlo));
4806
4807 add_to_sp( env, 16 ); // Reset SP
4808 return dst;
4809 }
4810 case Iop_F128toI128S:
4811 {
4812 HReg dst = newVRegV(env);
4813 HReg r_src = iselFp128Expr(env, e->Iex.Binop.arg2, IEndianess);
4814 PPCRI* rm = iselWordExpr_RI(env, e->Iex.Binop.arg1, IEndianess);
4815 /* Note: rm is a set of three bit fields that specify the
4816 * rounding mode and which of the two instructions to issue.
4817 */
4818 addInstr(env, PPCInstr_AvBinaryInt(Pav_F128toI128S, dst,
4819 r_src, rm));
4820 return dst;
4821 }
4822 case Iop_RndF128:
4823 {
4824 HReg dst = newVRegV(env);
4825 HReg r_src = iselFp128Expr(env, e->Iex.Binop.arg2, IEndianess);
4826 PPCRI* rm = iselWordExpr_RI(env, e->Iex.Binop.arg1, IEndianess);
4827 /* Note: rm is a set of three bit fields that specify the
4828 * rounding mode and which of the two instructions to issue.
4829 */
4830 addInstr(env, PPCInstr_AvBinaryInt(Pav_ROUNDFPQ, dst,
4831 r_src, rm));
4832 return dst;
4833 }
4834 case Iop_SqrtF128:
4835 if (FPU_rounding_mode_isOdd(e->Iex.Binop.arg1)) {
4836 /* use rounding mode specified by RN. Issue inst with R0 = 0 */
4837 fpop = Pfp_FPSQRTQRNDODD;
4838 goto do_Bin_F128;
4839 } else {
4840 set_FPU_rounding_mode( env, e->Iex.Binop.arg1, IEndianess );
4841 fpop = Pfp_FPSQRTQ;
4842 goto do_Bin_F128;
4843 }
4844 case Iop_F128toF32:
4845 if (FPU_rounding_mode_isOdd(e->Iex.Binop.arg1)) {
4846 /* use rounding mode specified by RN. Issue inst with R0 = 0 */
4847 fpop = Pfp_FPQTOWRNDODD;
4848 goto do_Bin_F128;
4849 } else {
4850 set_FPU_rounding_mode( env, e->Iex.Binop.arg1, IEndianess );
4851 fpop = Pfp_FPQTOW;
4852 goto do_Bin_F128;
4853 }
4854 do_Bin_F128: {
4855 HReg r_dst = newVRegV(env);
4856 HReg r_src = iselFp128Expr(env, e->Iex.Binop.arg2, IEndianess);
4857 addInstr(env, PPCInstr_Fp128Unary(fpop, r_dst, r_src));
4858 return r_dst;
4859 }
4860
4861 default:
4862 break;
4863 } /* switch (e->Iex.Binop.op) */
4864 } /* if (e->tag == Iex_Binop) */
4865
4866 if (e->tag == Iex_Triop) {
4867 IRTriop *triop = e->Iex.Triop.details;
4868
4869 switch (triop->op) {
4870 case Iop_AddF128:
4871 if (FPU_rounding_mode_isOdd(triop->arg1)) {
4872 /* use rounding mode specified by RN. Issue inst with R0 = 0 */
4873 fpop = Pfp_FPADDQRNDODD; goto do_Tri_F128;
4874 } else {
4875 set_FPU_rounding_mode( env, triop->arg1, IEndianess );
4876 fpop = Pfp_FPADDQ; goto do_Tri_F128;
4877 }
4878 case Iop_SubF128:
4879 if (FPU_rounding_mode_isOdd(triop->arg1)) {
4880 /* use rounding mode specified by RN. Issue inst with R0 = 0 */
4881 fpop = Pfp_FPSUBQRNDODD; goto do_Tri_F128;
4882 } else {
4883 set_FPU_rounding_mode( env, triop->arg1, IEndianess );
4884 fpop = Pfp_FPSUBQ; goto do_Tri_F128;
4885 }
4886 case Iop_MulF128:
4887 if (FPU_rounding_mode_isOdd(triop->arg1)) {
4888 /* use rounding mode specified by RN. Issue inst with R0 = 0 */
4889 fpop = Pfp_FPMULQRNDODD; goto do_Tri_F128;
4890 } else {
4891 set_FPU_rounding_mode( env, triop->arg1, IEndianess );
4892 fpop = Pfp_FPMULQ; goto do_Tri_F128;
4893 }
4894 case Iop_DivF128:
4895 if (FPU_rounding_mode_isOdd(triop->arg1)) {
4896 /* use rounding mode specified by RN. Issue inst with R0 = 0 */
4897 fpop = Pfp_FPDIVQRNDODD; goto do_Tri_F128;
4898 } else {
4899 set_FPU_rounding_mode( env, triop->arg1, IEndianess );
4900 fpop = Pfp_FPDIVQ; goto do_Tri_F128;
4901 }
4902 case Iop_MAddF128:
4903 if (FPU_rounding_mode_isOdd(triop->arg1)) {
4904 /* use rounding mode specified by RN. Issue inst with R0 = 0 */
4905 fpop = Pfp_FPMULADDQRNDODD; goto do_Tri_F128;
4906 } else {
4907 set_FPU_rounding_mode( env, triop->arg1, IEndianess );
4908 fpop = Pfp_FPMULADDQ; goto do_Tri_F128;
4909 }
4910
4911 do_Tri_F128: {
4912 HReg r_dst = newVRegV(env);
4913 HReg r_srcL = iselFp128Expr(env, triop->arg2, IEndianess);
4914 HReg r_srcR = iselFp128Expr(env, triop->arg3, IEndianess);
4915
4916 addInstr(env, PPCInstr_Fp128Binary(fpop, r_dst, r_srcL, r_srcR));
4917 return r_dst;
4918 }
4919
4920 default:
4921 break;
4922 } /* switch (e->Iex.Triop.op) */
4923
4924 } /* if (e->tag == Iex_Trinop) */
4925
4926 if (e->tag == Iex_Qop) {
4927 IRQop *qop = e->Iex.Qop.details;
4928
4929 switch (qop->op) {
4930 case Iop_MAddF128:
4931 if (FPU_rounding_mode_isOdd(qop->arg1)) {
4932 /* use rounding mode specified by RN. Issue inst with R0 = 0 */
4933 fpop = Pfp_FPMULADDQRNDODD; goto do_Quad_F128;
4934 } else {
4935 set_FPU_rounding_mode( env, qop->arg1, IEndianess );
4936 fpop = Pfp_FPMULADDQ; goto do_Quad_F128;
4937 }
4938 case Iop_MSubF128:
4939 if (FPU_rounding_mode_isOdd(qop->arg1)) {
4940 /* use rounding mode specified by RN. Issue inst with R0 = 0 */
4941 fpop = Pfp_FPMULSUBQRNDODD; goto do_Quad_F128;
4942 } else {
4943 set_FPU_rounding_mode( env, qop->arg1, IEndianess );
4944 fpop = Pfp_FPMULSUBQ; goto do_Quad_F128;
4945 }
4946 case Iop_NegMAddF128:
4947 if (FPU_rounding_mode_isOdd(qop->arg1)) {
4948 /* use rounding mode specified by RN. Issue inst with R0 = 0 */
4949 fpop = Pfp_FPNEGMULADDQRNDODD; goto do_Quad_F128;
4950 } else {
4951 set_FPU_rounding_mode( env, qop->arg1, IEndianess );
4952 fpop = Pfp_FPNEGMULADDQ; goto do_Quad_F128;
4953 }
4954 case Iop_NegMSubF128:
4955 if (FPU_rounding_mode_isOdd(qop->arg1)) {
4956 /* use rounding mode specified by RN. Issue inst with R0 = 0 */
4957 fpop = Pfp_FPNEGMULSUBQRNDODD; goto do_Quad_F128;
4958 } else {
4959 set_FPU_rounding_mode( env, qop->arg1, IEndianess );
4960 fpop = Pfp_FPNEGMULSUBQ; goto do_Quad_F128;
4961 }
4962
4963 do_Quad_F128: {
4964 HReg r_dst = iselFp128Expr(env, qop->arg3,
4965 IEndianess);
4966 HReg r_srcL = iselFp128Expr(env, qop->arg2,
4967 IEndianess);
4968 HReg r_srcR = iselFp128Expr(env, qop->arg4,
4969 IEndianess);
4970
4971 addInstr(env, PPCInstr_Fp128Trinary(fpop, r_dst, r_srcL, r_srcR));
4972 return r_dst;
4973 }
4974
4975 default:
4976 break;
4977 }
4978 } /* if (e->tag == Iex_Qop) */
4979
4980 ppIRExpr( e );
4981 vpanic( "iselFp128Expr(ppc64)" );
4982 }
4983
iselDfp64Expr(ISelEnv * env,const IRExpr * e,IREndness IEndianess)4984 static HReg iselDfp64Expr(ISelEnv* env, const IRExpr* e, IREndness IEndianess)
4985 {
4986 HReg r = iselDfp64Expr_wrk( env, e, IEndianess );
4987 vassert(hregClass(r) == HRcFlt64);
4988 vassert( hregIsVirtual(r) );
4989 return r;
4990 }
4991
4992 /* DO NOT CALL THIS DIRECTLY */
iselDfp64Expr_wrk(ISelEnv * env,const IRExpr * e,IREndness IEndianess)4993 static HReg iselDfp64Expr_wrk(ISelEnv* env, const IRExpr* e,
4994 IREndness IEndianess)
4995 {
4996 Bool mode64 = env->mode64;
4997 IRType ty = typeOfIRExpr( env->type_env, e );
4998 HReg r_dstHi, r_dstLo;
4999
5000 vassert( e );
5001 vassert( ty == Ity_D64 );
5002
5003 if (e->tag == Iex_RdTmp) {
5004 return lookupIRTemp( env, e->Iex.RdTmp.tmp );
5005 }
5006
5007 /* --------- GET --------- */
5008 if (e->tag == Iex_Get) {
5009 HReg r_dst = newVRegF( env );
5010 PPCAMode* am_addr = PPCAMode_IR( e->Iex.Get.offset,
5011 GuestStatePtr(mode64) );
5012 addInstr( env, PPCInstr_FpLdSt( True/*load*/, 8, r_dst, am_addr ) );
5013 return r_dst;
5014 }
5015
5016 if (e->tag == Iex_Load && e->Iex.Load.end == IEndianess) {
5017 PPCAMode* am_addr;
5018 HReg r_dst = newVRegF(env);
5019 vassert(e->Iex.Load.ty == Ity_D64);
5020 am_addr = iselWordExpr_AMode(env, e->Iex.Load.addr, Ity_D64/*xfer*/,
5021 IEndianess);
5022 addInstr(env, PPCInstr_FpLdSt(True/*load*/, 8, r_dst, am_addr));
5023 return r_dst;
5024 }
5025
5026 /* --------- OPS --------- */
5027 if (e->tag == Iex_Qop) {
5028 HReg r_dst = newVRegF( env );
5029 return r_dst;
5030 }
5031
5032 if (e->tag == Iex_Unop) {
5033 HReg fr_dst = newVRegF(env);
5034 switch (e->Iex.Unop.op) {
5035 case Iop_ReinterpI64asD64: {
5036 /* Given an I64, produce an IEEE754 DFP with the same
5037 bit pattern. */
5038 if (!mode64) {
5039 HReg r_srcHi, r_srcLo;
5040 iselInt64Expr( &r_srcHi, &r_srcLo, env, e->Iex.Unop.arg,
5041 IEndianess);
5042 return mk_LoadRR32toFPR( env, r_srcHi, r_srcLo );
5043 } else {
5044 HReg r_src = iselWordExpr_R(env, e->Iex.Unop.arg, IEndianess);
5045 return mk_LoadR64toFPR( env, r_src );
5046 }
5047 }
5048 case Iop_D32toD64: {
5049 HReg fr_src = iselDfp32Expr(env, e->Iex.Unop.arg, IEndianess);
5050 addInstr(env, PPCInstr_Dfp64Unary(Pfp_DCTDP, fr_dst, fr_src));
5051 return fr_dst;
5052 }
5053 case Iop_D128HItoD64:
5054 iselDfp128Expr( &r_dstHi, &r_dstLo, env, e->Iex.Unop.arg,
5055 IEndianess );
5056 return r_dstHi;
5057 case Iop_D128LOtoD64:
5058 iselDfp128Expr( &r_dstHi, &r_dstLo, env, e->Iex.Unop.arg,
5059 IEndianess );
5060 return r_dstLo;
5061 case Iop_InsertExpD64: {
5062 HReg fr_srcL = iselDblExpr(env, e->Iex.Binop.arg1, IEndianess);
5063 HReg fr_srcR = iselDblExpr(env, e->Iex.Binop.arg2, IEndianess);
5064
5065 addInstr(env, PPCInstr_Dfp64Binary(Pfp_DIEX, fr_dst, fr_srcL,
5066 fr_srcR));
5067 return fr_dst;
5068 }
5069 default:
5070 vex_printf( "ERROR: iselDfp64Expr_wrk, UNKNOWN unop case %d\n",
5071 (Int)e->Iex.Unop.op );
5072 }
5073 }
5074
5075 if (e->tag == Iex_Binop) {
5076 PPCFpOp fpop = Pfp_INVALID;
5077 HReg fr_dst = newVRegF(env);
5078
5079 switch (e->Iex.Binop.op) {
5080 case Iop_D128toD64: fpop = Pfp_DRDPQ; break;
5081 case Iop_D64toD32: fpop = Pfp_DRSP; break;
5082 case Iop_I64StoD64: fpop = Pfp_DCFFIX; break;
5083 case Iop_RoundD64toInt: fpop = Pfp_DRINTN; break;
5084 default: break;
5085 }
5086 if (fpop == Pfp_DRDPQ) {
5087 HReg r_srcHi = newVRegF(env);
5088 HReg r_srcLo = newVRegF(env);
5089
5090 set_FPU_DFP_rounding_mode( env, e->Iex.Binop.arg1, IEndianess );
5091 iselDfp128Expr(&r_srcHi, &r_srcLo, env, e->Iex.Binop.arg2,
5092 IEndianess);
5093 addInstr(env, PPCInstr_DfpD128toD64(fpop, fr_dst, r_srcHi, r_srcLo));
5094 return fr_dst;
5095
5096 } else if (fpop == Pfp_DRINTN) {
5097 HReg fr_src = newVRegF(env);
5098 PPCRI* r_rmc = iselWordExpr_RI(env, e->Iex.Binop.arg1, IEndianess);
5099
5100 /* NOTE, this IOP takes a DFP value and rounds to the
5101 * neares floating point integer value, i.e. fractional part
5102 * is zero. The result is a decimal floating point number.
5103 * the INT in the name is a bit misleading.
5104 */
5105 fr_src = iselDfp64Expr(env, e->Iex.Binop.arg2, IEndianess);
5106 addInstr(env, PPCInstr_DfpRound(fr_dst, fr_src, r_rmc));
5107 return fr_dst;
5108
5109 } else if (fpop == Pfp_DRSP) {
5110 HReg fr_src = iselDfp64Expr(env, e->Iex.Binop.arg2, IEndianess);
5111 set_FPU_DFP_rounding_mode( env, e->Iex.Binop.arg1, IEndianess );
5112 addInstr(env, PPCInstr_Dfp64Unary(fpop, fr_dst, fr_src));
5113 return fr_dst;
5114
5115 } else if (fpop == Pfp_DCFFIX) {
5116 HReg fr_src = newVRegF(env);
5117 PPCAMode* zero_r1 = PPCAMode_IR( 0, StackFramePtr(env->mode64) );
5118
5119 set_FPU_DFP_rounding_mode( env, e->Iex.Binop.arg1, IEndianess );
5120 sub_from_sp( env, 16 );
5121
5122 // put the I64 value into a floating point register
5123 if (mode64) {
5124 HReg tmp = iselWordExpr_R(env, e->Iex.Binop.arg2, IEndianess);
5125
5126 addInstr(env, PPCInstr_Store(8, zero_r1, tmp, True/*mode64*/));
5127 } else {
5128 HReg tmpHi, tmpLo;
5129 PPCAMode* four_r1 = PPCAMode_IR( 4, StackFramePtr(env->mode64) );
5130
5131 iselInt64Expr(&tmpHi, &tmpLo, env, e->Iex.Binop.arg2,
5132 IEndianess);
5133 addInstr(env, PPCInstr_Store(4, zero_r1, tmpHi, False/*mode32*/));
5134 addInstr(env, PPCInstr_Store(4, four_r1, tmpLo, False/*mode32*/));
5135 }
5136
5137 addInstr(env, PPCInstr_FpLdSt(True/*load*/, 8, fr_src, zero_r1));
5138 addInstr(env, PPCInstr_Dfp64Unary(fpop, fr_dst, fr_src));
5139 add_to_sp( env, 16 );
5140 return fr_dst;
5141 }
5142
5143 switch (e->Iex.Binop.op) {
5144 /* shift instructions D64, I32 -> D64 */
5145 case Iop_ShlD64: fpop = Pfp_DSCLI; break;
5146 case Iop_ShrD64: fpop = Pfp_DSCRI; break;
5147 default: break;
5148 }
5149 if (fpop != Pfp_INVALID) {
5150 HReg fr_src = iselDfp64Expr(env, e->Iex.Binop.arg1, IEndianess);
5151 PPCRI* shift = iselWordExpr_RI(env, e->Iex.Binop.arg2, IEndianess);
5152
5153 /* shift value must be an immediate value */
5154 vassert(shift->tag == Pri_Imm);
5155
5156 addInstr(env, PPCInstr_DfpShift(fpop, fr_dst, fr_src, shift));
5157 return fr_dst;
5158 }
5159
5160 switch (e->Iex.Binop.op) {
5161 case Iop_InsertExpD64:
5162 fpop = Pfp_DIEX;
5163 break;
5164 default: break;
5165 }
5166 if (fpop != Pfp_INVALID) {
5167 HReg fr_srcL = newVRegF(env);
5168 HReg fr_srcR = iselDfp64Expr(env, e->Iex.Binop.arg2, IEndianess);
5169 PPCAMode* zero_r1 = PPCAMode_IR( 0, StackFramePtr(env->mode64) );
5170 sub_from_sp( env, 16 );
5171
5172 if (env->mode64) {
5173 // put the I64 value into a floating point reg
5174 HReg tmp = iselWordExpr_R(env, e->Iex.Binop.arg1, IEndianess);
5175
5176 addInstr(env, PPCInstr_Store(8, zero_r1, tmp, True/*mode64*/));
5177 } else {
5178 // put the I64 register pair into a floating point reg
5179 HReg tmpHi;
5180 HReg tmpLo;
5181 PPCAMode* four_r1 = PPCAMode_IR( 4, StackFramePtr(env->mode64) );
5182
5183 iselInt64Expr(&tmpHi, &tmpLo, env, e->Iex.Binop.arg1,
5184 IEndianess);
5185 addInstr(env, PPCInstr_Store(4, zero_r1, tmpHi, False/*!mode64*/));
5186 addInstr(env, PPCInstr_Store(4, four_r1, tmpLo, False/*!mode64*/));
5187 }
5188 addInstr(env, PPCInstr_FpLdSt(True/*load*/, 8, fr_srcL, zero_r1));
5189 addInstr(env, PPCInstr_Dfp64Binary(fpop, fr_dst, fr_srcL,
5190 fr_srcR));
5191 add_to_sp( env, 16 );
5192 return fr_dst;
5193 }
5194 }
5195
5196 if (e->tag == Iex_Triop) {
5197 IRTriop *triop = e->Iex.Triop.details;
5198 PPCFpOp fpop = Pfp_INVALID;
5199
5200 switch (triop->op) {
5201 case Iop_AddD64:
5202 fpop = Pfp_DFPADD;
5203 break;
5204 case Iop_SubD64:
5205 fpop = Pfp_DFPSUB;
5206 break;
5207 case Iop_MulD64:
5208 fpop = Pfp_DFPMUL;
5209 break;
5210 case Iop_DivD64:
5211 fpop = Pfp_DFPDIV;
5212 break;
5213 default:
5214 break;
5215 }
5216 if (fpop != Pfp_INVALID) {
5217 HReg r_dst = newVRegF( env );
5218 HReg r_srcL = iselDfp64Expr( env, triop->arg2, IEndianess );
5219 HReg r_srcR = iselDfp64Expr( env, triop->arg3, IEndianess );
5220
5221 set_FPU_DFP_rounding_mode( env, triop->arg1, IEndianess );
5222 addInstr( env, PPCInstr_Dfp64Binary( fpop, r_dst, r_srcL, r_srcR ) );
5223 return r_dst;
5224 }
5225
5226 switch (triop->op) {
5227 case Iop_QuantizeD64: fpop = Pfp_DQUA; break;
5228 case Iop_SignificanceRoundD64: fpop = Pfp_RRDTR; break;
5229 default: break;
5230 }
5231 if (fpop == Pfp_DQUA) {
5232 HReg r_dst = newVRegF(env);
5233 HReg r_srcL = iselDfp64Expr(env, triop->arg2, IEndianess);
5234 HReg r_srcR = iselDfp64Expr(env, triop->arg3, IEndianess);
5235 PPCRI* rmc = iselWordExpr_RI(env, triop->arg1, IEndianess);
5236 addInstr(env, PPCInstr_DfpQuantize(fpop, r_dst, r_srcL, r_srcR,
5237 rmc));
5238 return r_dst;
5239
5240 } else if (fpop == Pfp_RRDTR) {
5241 HReg r_dst = newVRegF(env);
5242 HReg r_srcL = newVRegF(env);
5243 HReg r_srcR = iselDfp64Expr(env, triop->arg3, IEndianess);
5244 PPCRI* rmc = iselWordExpr_RI(env, triop->arg1, IEndianess);
5245 PPCAMode* zero_r1 = PPCAMode_IR( 0, StackFramePtr(env->mode64) );
5246 HReg i8_val = iselWordExpr_R(env, triop->arg2, IEndianess);
5247
5248 /* Move I8 to float register to issue instruction */
5249 sub_from_sp( env, 16 );
5250 if (mode64)
5251 addInstr(env, PPCInstr_Store(8, zero_r1, i8_val, True/*mode64*/));
5252 else
5253 addInstr(env, PPCInstr_Store(4, zero_r1, i8_val, False/*mode32*/));
5254
5255 addInstr(env, PPCInstr_FpLdSt(True/*load*/, 8, r_srcL, zero_r1));
5256 add_to_sp( env, 16 );
5257
5258 // will set TE and RMC when issuing instruction
5259 addInstr(env, PPCInstr_DfpQuantize(fpop, r_dst, r_srcL, r_srcR, rmc));
5260 return r_dst;
5261 }
5262 }
5263
5264 ppIRExpr( e );
5265 vpanic( "iselDfp64Expr_wrk(ppc)" );
5266 }
5267
iselDfp128Expr(HReg * rHi,HReg * rLo,ISelEnv * env,const IRExpr * e,IREndness IEndianess)5268 static void iselDfp128Expr(HReg* rHi, HReg* rLo, ISelEnv* env, const IRExpr* e,
5269 IREndness IEndianess)
5270 {
5271 iselDfp128Expr_wrk( rHi, rLo, env, e, IEndianess );
5272 vassert( hregIsVirtual(*rHi) );
5273 vassert( hregIsVirtual(*rLo) );
5274 }
5275
5276 /* DO NOT CALL THIS DIRECTLY */
iselDfp128Expr_wrk(HReg * rHi,HReg * rLo,ISelEnv * env,const IRExpr * e,IREndness IEndianess)5277 static void iselDfp128Expr_wrk(HReg* rHi, HReg *rLo, ISelEnv* env,
5278 const IRExpr* e, IREndness IEndianess)
5279 {
5280 vassert( e );
5281 vassert( typeOfIRExpr(env->type_env,e) == Ity_D128 );
5282
5283 /* read 128-bit IRTemp */
5284 if (e->tag == Iex_RdTmp) {
5285 lookupIRTempPair( rHi, rLo, env, e->Iex.RdTmp.tmp );
5286 return;
5287 }
5288
5289 if (e->tag == Iex_Unop) {
5290 HReg r_dstHi = newVRegF(env);
5291 HReg r_dstLo = newVRegF(env);
5292
5293 if (e->Iex.Unop.op == Iop_I64StoD128) {
5294 HReg fr_src = newVRegF(env);
5295 PPCAMode* zero_r1 = PPCAMode_IR( 0, StackFramePtr(env->mode64) );
5296
5297 // put the I64 value into a floating point reg
5298 if (env->mode64) {
5299 HReg tmp = iselWordExpr_R(env, e->Iex.Unop.arg, IEndianess);
5300 addInstr(env, PPCInstr_Store(8, zero_r1, tmp, True/*mode64*/));
5301 } else {
5302 HReg tmpHi, tmpLo;
5303 PPCAMode* four_r1 = PPCAMode_IR( 4, StackFramePtr(env->mode64) );
5304
5305 iselInt64Expr(&tmpHi, &tmpLo, env, e->Iex.Unop.arg,
5306 IEndianess);
5307 addInstr(env, PPCInstr_Store(4, zero_r1, tmpHi, False/*mode32*/));
5308 addInstr(env, PPCInstr_Store(4, four_r1, tmpLo, False/*mode32*/));
5309 }
5310
5311 addInstr(env, PPCInstr_FpLdSt(True/*load*/, 8, fr_src, zero_r1));
5312 addInstr(env, PPCInstr_DfpI64StoD128(Pfp_DCFFIXQ, r_dstHi, r_dstLo,
5313 fr_src));
5314 }
5315
5316 if (e->Iex.Unop.op == Iop_D64toD128) {
5317 HReg r_src = iselDfp64Expr(env, e->Iex.Unop.arg, IEndianess);
5318
5319 /* Source is 64bit, result is 128 bit. High 64bit source arg,
5320 * is ignored by the instruction. Set high arg to r_src just
5321 * to meet the vassert tests.
5322 */
5323 addInstr(env, PPCInstr_Dfp128Unary(Pfp_DCTQPQ, r_dstHi, r_dstLo,
5324 r_src, r_src));
5325 }
5326 *rHi = r_dstHi;
5327 *rLo = r_dstLo;
5328 return;
5329 }
5330
5331 /* --------- OPS --------- */
5332 if (e->tag == Iex_Binop) {
5333 HReg r_srcHi;
5334 HReg r_srcLo;
5335
5336 switch (e->Iex.Binop.op) {
5337 case Iop_D64HLtoD128:
5338 r_srcHi = iselDfp64Expr( env, e->Iex.Binop.arg1, IEndianess );
5339 r_srcLo = iselDfp64Expr( env, e->Iex.Binop.arg2, IEndianess );
5340 *rHi = r_srcHi;
5341 *rLo = r_srcLo;
5342 return;
5343 break;
5344 case Iop_D128toD64: {
5345 PPCFpOp fpop = Pfp_DRDPQ;
5346 HReg fr_dst = newVRegF(env);
5347
5348 set_FPU_DFP_rounding_mode( env, e->Iex.Binop.arg1, IEndianess );
5349 iselDfp128Expr(&r_srcHi, &r_srcLo, env, e->Iex.Binop.arg2,
5350 IEndianess);
5351 addInstr(env, PPCInstr_DfpD128toD64(fpop, fr_dst, r_srcHi, r_srcLo));
5352
5353 /* Need to meet the interface spec but the result is
5354 * just 64-bits so send the result back in both halfs.
5355 */
5356 *rHi = fr_dst;
5357 *rLo = fr_dst;
5358 return;
5359 }
5360 case Iop_ShlD128:
5361 case Iop_ShrD128: {
5362 HReg fr_dst_hi = newVRegF(env);
5363 HReg fr_dst_lo = newVRegF(env);
5364 PPCRI* shift = iselWordExpr_RI(env, e->Iex.Binop.arg2, IEndianess);
5365 PPCFpOp fpop = Pfp_DSCLIQ; /* fix later if necessary */
5366
5367 iselDfp128Expr(&r_srcHi, &r_srcLo, env, e->Iex.Binop.arg1,
5368 IEndianess);
5369
5370 if (e->Iex.Binop.op == Iop_ShrD128)
5371 fpop = Pfp_DSCRIQ;
5372
5373 addInstr(env, PPCInstr_DfpShift128(fpop, fr_dst_hi, fr_dst_lo,
5374 r_srcHi, r_srcLo, shift));
5375
5376 *rHi = fr_dst_hi;
5377 *rLo = fr_dst_lo;
5378 return;
5379 }
5380 case Iop_RoundD128toInt: {
5381 HReg r_dstHi = newVRegF(env);
5382 HReg r_dstLo = newVRegF(env);
5383 PPCRI* r_rmc = iselWordExpr_RI(env, e->Iex.Binop.arg1, IEndianess);
5384
5385 // will set R and RMC when issuing instruction
5386 iselDfp128Expr(&r_srcHi, &r_srcLo, env, e->Iex.Binop.arg2,
5387 IEndianess);
5388
5389 addInstr(env, PPCInstr_DfpRound128(r_dstHi, r_dstLo,
5390 r_srcHi, r_srcLo, r_rmc));
5391 *rHi = r_dstHi;
5392 *rLo = r_dstLo;
5393 return;
5394 }
5395 case Iop_InsertExpD128: {
5396 HReg r_dstHi = newVRegF(env);
5397 HReg r_dstLo = newVRegF(env);
5398 HReg r_srcL = newVRegF(env);
5399 PPCAMode* zero_r1 = PPCAMode_IR( 0, StackFramePtr(env->mode64) );
5400 r_srcHi = newVRegF(env);
5401 r_srcLo = newVRegF(env);
5402
5403 iselDfp128Expr(&r_srcHi, &r_srcLo, env, e->Iex.Binop.arg2,
5404 IEndianess);
5405
5406 /* Move I64 to float register to issue instruction */
5407 if (env->mode64) {
5408 HReg tmp = iselWordExpr_R(env, e->Iex.Binop.arg1, IEndianess);
5409 addInstr(env, PPCInstr_Store(8, zero_r1, tmp, True/*mode64*/));
5410 } else {
5411 HReg tmpHi, tmpLo;
5412 PPCAMode* four_r1 = PPCAMode_IR( 4, StackFramePtr(env->mode64) );
5413
5414 iselInt64Expr(&tmpHi, &tmpLo, env, e->Iex.Unop.arg,
5415 IEndianess);
5416 addInstr(env, PPCInstr_Store(4, zero_r1, tmpHi, False/*mode32*/));
5417 addInstr(env, PPCInstr_Store(4, four_r1, tmpLo, False/*mode32*/));
5418 }
5419
5420 addInstr(env, PPCInstr_FpLdSt(True/*load*/, 8, r_srcL, zero_r1));
5421 addInstr(env, PPCInstr_InsertExpD128(Pfp_DIEXQ,
5422 r_dstHi, r_dstLo,
5423 r_srcL, r_srcHi, r_srcLo));
5424 *rHi = r_dstHi;
5425 *rLo = r_dstLo;
5426 return;
5427 }
5428 default:
5429 vex_printf( "ERROR: iselDfp128Expr_wrk, UNKNOWN binop case %d\n",
5430 (Int)e->Iex.Binop.op );
5431 break;
5432 }
5433 }
5434
5435 if (e->tag == Iex_Triop) {
5436 IRTriop *triop = e->Iex.Triop.details;
5437 PPCFpOp fpop = Pfp_INVALID;
5438 HReg r_dstHi = newVRegF(env);
5439 HReg r_dstLo = newVRegF(env);
5440
5441 switch (triop->op) {
5442 case Iop_AddD128:
5443 fpop = Pfp_DFPADDQ;
5444 break;
5445 case Iop_SubD128:
5446 fpop = Pfp_DFPSUBQ;
5447 break;
5448 case Iop_MulD128:
5449 fpop = Pfp_DFPMULQ;
5450 break;
5451 case Iop_DivD128:
5452 fpop = Pfp_DFPDIVQ;
5453 break;
5454 default:
5455 break;
5456 }
5457
5458 if (fpop != Pfp_INVALID) {
5459 HReg r_srcRHi = newVRegV( env );
5460 HReg r_srcRLo = newVRegV( env );
5461
5462 /* dst will be used to pass in the left operand and get the result. */
5463 iselDfp128Expr( &r_dstHi, &r_dstLo, env, triop->arg2, IEndianess );
5464 iselDfp128Expr( &r_srcRHi, &r_srcRLo, env, triop->arg3, IEndianess );
5465 set_FPU_DFP_rounding_mode( env, triop->arg1, IEndianess );
5466 addInstr( env,
5467 PPCInstr_Dfp128Binary( fpop, r_dstHi, r_dstLo,
5468 r_srcRHi, r_srcRLo ) );
5469 *rHi = r_dstHi;
5470 *rLo = r_dstLo;
5471 return;
5472 }
5473 switch (triop->op) {
5474 case Iop_QuantizeD128: fpop = Pfp_DQUAQ; break;
5475 case Iop_SignificanceRoundD128: fpop = Pfp_DRRNDQ; break;
5476 default: break;
5477 }
5478 if (fpop == Pfp_DQUAQ) {
5479 HReg r_srcHi = newVRegF(env);
5480 HReg r_srcLo = newVRegF(env);
5481 PPCRI* rmc = iselWordExpr_RI(env, triop->arg1, IEndianess);
5482
5483 /* dst will be used to pass in the left operand and get the result */
5484 iselDfp128Expr(&r_dstHi, &r_dstLo, env, triop->arg2, IEndianess);
5485 iselDfp128Expr(&r_srcHi, &r_srcLo, env, triop->arg3, IEndianess);
5486
5487 // will set RMC when issuing instruction
5488 addInstr(env, PPCInstr_DfpQuantize128(fpop, r_dstHi, r_dstLo,
5489 r_srcHi, r_srcLo, rmc));
5490 *rHi = r_dstHi;
5491 *rLo = r_dstLo;
5492 return;
5493
5494 } else if (fpop == Pfp_DRRNDQ) {
5495 HReg r_srcHi = newVRegF(env);
5496 HReg r_srcLo = newVRegF(env);
5497 PPCRI* rmc = iselWordExpr_RI(env, triop->arg1, IEndianess);
5498 PPCAMode* zero_r1 = PPCAMode_IR( 0, StackFramePtr(env->mode64) );
5499 PPCAMode* four_r1 = PPCAMode_IR( 4, StackFramePtr(env->mode64) );
5500 HReg i8_val = iselWordExpr_R(env, triop->arg2, IEndianess);
5501 HReg r_zero = newVRegI( env );
5502
5503 iselDfp128Expr(&r_srcHi, &r_srcLo, env, triop->arg3, IEndianess);
5504
5505 /* dst will be used to pass in the left operand and get the result */
5506 /* Move I8 to float register to issue instruction. Note, the
5507 * instruction only looks at the bottom 6 bits so we really don't
5508 * have to clear the upper bits since the iselWordExpr_R sets the
5509 * bottom 8-bits.
5510 */
5511 sub_from_sp( env, 16 );
5512
5513 if (env->mode64)
5514 addInstr(env, PPCInstr_Store(4, four_r1, i8_val, True/*mode64*/));
5515 else
5516 addInstr(env, PPCInstr_Store(4, four_r1, i8_val, False/*mode32*/));
5517
5518 /* Have to write to the upper bits to ensure they have been
5519 * initialized. The instruction ignores all but the lower 6-bits.
5520 */
5521 addInstr( env, PPCInstr_LI( r_zero, 0, env->mode64 ) );
5522 addInstr(env, PPCInstr_FpLdSt(True/*load*/, 8, r_dstHi, zero_r1));
5523 addInstr(env, PPCInstr_FpLdSt(True/*load*/, 8, r_dstLo, zero_r1));
5524
5525 add_to_sp( env, 16 );
5526
5527 // will set RMC when issuing instruction
5528 addInstr(env, PPCInstr_DfpQuantize128(fpop, r_dstHi, r_dstLo,
5529 r_srcHi, r_srcLo, rmc));
5530 *rHi = r_dstHi;
5531 *rLo = r_dstLo;
5532 return;
5533 }
5534 }
5535
5536 ppIRExpr( e );
5537 vpanic( "iselDfp128Expr(ppc64)" );
5538 }
5539
5540
5541 /*---------------------------------------------------------*/
5542 /*--- ISEL: SIMD (Vector) expressions, 128 bit. ---*/
5543 /*---------------------------------------------------------*/
5544
iselVecExpr(ISelEnv * env,const IRExpr * e,IREndness IEndianess)5545 static HReg iselVecExpr ( ISelEnv* env, const IRExpr* e, IREndness IEndianess )
5546 {
5547 HReg r = iselVecExpr_wrk( env, e, IEndianess );
5548 # if 0
5549 vex_printf("\n"); ppIRExpr(e); vex_printf("\n");
5550 # endif
5551 vassert(hregClass(r) == HRcVec128);
5552 vassert(hregIsVirtual(r));
5553 return r;
5554 }
5555
5556 /* DO NOT CALL THIS DIRECTLY */
iselVecExpr_wrk(ISelEnv * env,const IRExpr * e,IREndness IEndianess)5557 static HReg iselVecExpr_wrk ( ISelEnv* env, const IRExpr* e,
5558 IREndness IEndianess )
5559 {
5560 Bool mode64 = env->mode64;
5561 PPCAvOp op = Pav_INVALID;
5562 PPCAvFpOp fpop = Pavfp_INVALID;
5563 IRType ty = typeOfIRExpr(env->type_env,e);
5564 vassert(e);
5565 vassert(ty == Ity_V128);
5566
5567 if (e->tag == Iex_RdTmp) {
5568 return lookupIRTemp(env, e->Iex.RdTmp.tmp);
5569 }
5570
5571 if (e->tag == Iex_Get) {
5572 /* Guest state vectors are 16byte aligned,
5573 so don't need to worry here */
5574 HReg dst = newVRegV(env);
5575 addInstr(env,
5576 PPCInstr_AvLdSt( True/*load*/, 16, dst,
5577 PPCAMode_IR( e->Iex.Get.offset,
5578 GuestStatePtr(mode64) )));
5579 return dst;
5580 }
5581
5582 if (e->tag == Iex_Load && e->Iex.Load.end == IEndianess) {
5583 /* Need to be able to do V128 unaligned loads. The BE unaligned load
5584 * can be accomplised using the following code sequece from the ISA.
5585 * It uses the lvx instruction that does two aligned loads and then
5586 * permute the data to store the required data as if it had been an
5587 * unaligned load.
5588 *
5589 * lvx Vhi,0,Rb # load MSQ, using the unaligned address in Rb
5590 * lvsl Vp, 0,Rb # Set permute control vector
5591 * addi Rb,Rb,15 # Address of LSQ
5592 * lvx Vlo,0,Rb # load LSQ
5593 * vperm Vt,Vhi,Vlo,Vp # align the data as requested
5594 */
5595
5596 HReg Vhi = newVRegV(env);
5597 HReg Vlo = newVRegV(env);
5598 HReg Vp = newVRegV(env);
5599 HReg v_dst = newVRegV(env);
5600 HReg rB;
5601 HReg rB_plus_15 = newVRegI(env);
5602
5603 vassert(e->Iex.Load.ty == Ity_V128);
5604 rB = iselWordExpr_R( env, e->Iex.Load.addr, IEndianess );
5605
5606 // lvx Vhi, 0, Rb
5607 addInstr(env, PPCInstr_AvLdSt( True/*load*/, 16, Vhi,
5608 PPCAMode_IR(0, rB)) );
5609
5610 if (IEndianess == Iend_LE)
5611 // lvsr Vp, 0, Rb
5612 addInstr(env, PPCInstr_AvSh( False/*right shift*/, Vp,
5613 PPCAMode_IR(0, rB)) );
5614 else
5615 // lvsl Vp, 0, Rb
5616 addInstr(env, PPCInstr_AvSh( True/*left shift*/, Vp,
5617 PPCAMode_IR(0, rB)) );
5618
5619 // addi Rb_plus_15, Rb, 15
5620 addInstr(env, PPCInstr_Alu( Palu_ADD, rB_plus_15,
5621 rB, PPCRH_Imm(True, toUShort(15))) );
5622
5623 // lvx Vlo, 0, Rb_plus_15
5624 addInstr(env, PPCInstr_AvLdSt( True/*load*/, 16, Vlo,
5625 PPCAMode_IR(0, rB_plus_15)) );
5626
5627 if (IEndianess == Iend_LE)
5628 // vperm Vt, Vhi, Vlo, Vp
5629 addInstr(env, PPCInstr_AvPerm( v_dst, Vlo, Vhi, Vp ));
5630 else
5631 // vperm Vt, Vhi, Vlo, Vp
5632 addInstr(env, PPCInstr_AvPerm( v_dst, Vhi, Vlo, Vp ));
5633
5634 return v_dst;
5635 }
5636
5637 if (e->tag == Iex_Unop) {
5638 switch (e->Iex.Unop.op) {
5639
5640 case Iop_F16toF64x2:
5641 {
5642 HReg dst = newVRegV(env);
5643 HReg arg = iselVecExpr(env, e->Iex.Unop.arg, IEndianess);
5644 /* Note: PPC only coverts the 16-bt value in the upper word
5645 * to a 64-bit value stored in the upper word. The
5646 * contents of the lower word is undefined.
5647 */
5648 addInstr(env, PPCInstr_AvUnary(Pav_F16toF64x2, dst, arg));
5649 return dst;
5650 }
5651
5652 case Iop_F64toF16x2:
5653 {
5654 HReg dst = newVRegV(env);
5655 HReg arg = iselVecExpr(env, e->Iex.Unop.arg, IEndianess);
5656 /* Note: PPC only coverts the 64-bt value in the upper 64-bit of V128
5657 * to a 16-bit value stored in the upper 64-bits of the result
5658 * V128. The contents of the lower 64-bits is undefined.
5659 */
5660 addInstr(env, PPCInstr_AvUnary(Pav_F64toF16x2, dst, arg));
5661 return dst;
5662 }
5663
5664 case Iop_F16toF32x4:
5665 {
5666 HReg src = newVRegV(env);
5667 HReg dst = newVRegV(env);
5668 HReg arg = iselWordExpr_R(env, e->Iex.Unop.arg, IEndianess);
5669 PPCAMode *am_off0, *am_off8;
5670 HReg r_aligned16;
5671
5672 vassert(mode64);
5673 /* need to put I64 src into upper 64-bits of vector register,
5674 use stack */
5675 sub_from_sp( env, 32 ); // Move SP down
5676
5677 /* Get a quadword aligned address within our stack space */
5678 r_aligned16 = get_sp_aligned16( env );
5679 am_off0 = PPCAMode_IR( 0, r_aligned16 );
5680 am_off8 = PPCAMode_IR( 8, r_aligned16 );
5681
5682 /* Store I64 to stack */
5683
5684 if (IEndianess == Iend_LE) {
5685 addInstr(env, PPCInstr_Store( 8, am_off8, arg, mode64 ));
5686 } else {
5687 addInstr(env, PPCInstr_Store( 8, am_off0, arg, mode64 ));
5688 }
5689
5690 /* Fetch new v128 src back from stack. */
5691 addInstr(env, PPCInstr_AvLdSt(True/*ld*/, 16, src, am_off0));
5692
5693 /* issue instruction */
5694 addInstr(env, PPCInstr_AvUnary(Pav_F16toF32x4, dst, src));
5695 add_to_sp( env, 32 ); // Reset SP
5696
5697 return dst;
5698 }
5699
5700 case Iop_F32toF16x4:
5701 {
5702 HReg dst = newVRegI(env);
5703 HReg tmp = newVRegV(env);
5704 HReg arg = iselVecExpr(env, e->Iex.Unop.arg, IEndianess);
5705 PPCAMode *am_off0, *am_off8;
5706 HReg r_aligned16;
5707
5708 /* Instruction returns a V128, the Iop_F32toF16x4 needs to return
5709 * I64. Move the upper 64-bits from the instruction to an I64 via
5710 * the stack and return it.
5711 */
5712 sub_from_sp( env, 32 ); // Move SP down
5713
5714 addInstr(env, PPCInstr_AvUnary(Pav_F32toF16x4, tmp, arg));
5715
5716 /* Get a quadword aligned address within our stack space */
5717 r_aligned16 = get_sp_aligned16( env );
5718 am_off0 = PPCAMode_IR( 0, r_aligned16 );
5719 am_off8 = PPCAMode_IR( 8, r_aligned16 );
5720
5721 /* Store v128 tmp to stack. */
5722 addInstr(env, PPCInstr_AvLdSt(False/*store*/, 16, tmp, am_off0));
5723
5724 /* Fetch I64 from stack */
5725 if (IEndianess == Iend_LE) {
5726 addInstr(env, PPCInstr_Load( 8, dst, am_off8, mode64 ));
5727 } else {
5728 addInstr(env, PPCInstr_Load( 8, dst, am_off0, mode64 ));
5729 }
5730
5731 add_to_sp( env, 32 ); // Reset SP
5732 return dst;
5733 }
5734
5735 case Iop_NotV128: {
5736 HReg arg = iselVecExpr(env, e->Iex.Unop.arg, IEndianess);
5737 HReg dst = newVRegV(env);
5738 addInstr(env, PPCInstr_AvUnary(Pav_NOT, dst, arg));
5739 return dst;
5740 }
5741
5742 case Iop_CmpNEZ8x16: {
5743 HReg arg = iselVecExpr(env, e->Iex.Unop.arg, IEndianess);
5744 HReg zero = newVRegV(env);
5745 HReg dst = newVRegV(env);
5746 addInstr(env, PPCInstr_AvBinary(Pav_XOR, zero, zero, zero));
5747 addInstr(env, PPCInstr_AvBin8x16(Pav_CMPEQU, dst, arg, zero));
5748 addInstr(env, PPCInstr_AvUnary(Pav_NOT, dst, dst));
5749 return dst;
5750 }
5751
5752 case Iop_CmpNEZ16x8: {
5753 HReg arg = iselVecExpr(env, e->Iex.Unop.arg, IEndianess);
5754 HReg zero = newVRegV(env);
5755 HReg dst = newVRegV(env);
5756 addInstr(env, PPCInstr_AvBinary(Pav_XOR, zero, zero, zero));
5757 addInstr(env, PPCInstr_AvBin16x8(Pav_CMPEQU, dst, arg, zero));
5758 addInstr(env, PPCInstr_AvUnary(Pav_NOT, dst, dst));
5759 return dst;
5760 }
5761
5762 case Iop_CmpNEZ32x4: {
5763 HReg arg = iselVecExpr(env, e->Iex.Unop.arg, IEndianess);
5764 HReg zero = newVRegV(env);
5765 HReg dst = newVRegV(env);
5766 addInstr(env, PPCInstr_AvBinary(Pav_XOR, zero, zero, zero));
5767 addInstr(env, PPCInstr_AvBin32x4(Pav_CMPEQU, dst, arg, zero));
5768 addInstr(env, PPCInstr_AvUnary(Pav_NOT, dst, dst));
5769 return dst;
5770 }
5771
5772 case Iop_CmpNEZ64x2: {
5773 HReg arg = iselVecExpr(env, e->Iex.Unop.arg, IEndianess);
5774 HReg zero = newVRegV(env);
5775 HReg dst = newVRegV(env);
5776 addInstr(env, PPCInstr_AvBinary(Pav_XOR, zero, zero, zero));
5777 addInstr(env, PPCInstr_AvBin64x2(Pav_CMPEQU, dst, arg, zero));
5778 addInstr(env, PPCInstr_AvUnary(Pav_NOT, dst, dst));
5779 return dst;
5780 }
5781
5782 case Iop_RecipEst32Fx4: fpop = Pavfp_RCPF; goto do_32Fx4_unary;
5783 case Iop_RSqrtEst32Fx4: fpop = Pavfp_RSQRTF; goto do_32Fx4_unary;
5784 case Iop_I32UtoFx4: fpop = Pavfp_CVTU2F; goto do_32Fx4_unary;
5785 case Iop_I32StoFx4: fpop = Pavfp_CVTS2F; goto do_32Fx4_unary;
5786 case Iop_QFtoI32Ux4_RZ: fpop = Pavfp_QCVTF2U; goto do_32Fx4_unary;
5787 case Iop_QFtoI32Sx4_RZ: fpop = Pavfp_QCVTF2S; goto do_32Fx4_unary;
5788 case Iop_RoundF32x4_RM: fpop = Pavfp_ROUNDM; goto do_32Fx4_unary;
5789 case Iop_RoundF32x4_RP: fpop = Pavfp_ROUNDP; goto do_32Fx4_unary;
5790 case Iop_RoundF32x4_RN: fpop = Pavfp_ROUNDN; goto do_32Fx4_unary;
5791 case Iop_RoundF32x4_RZ: fpop = Pavfp_ROUNDZ; goto do_32Fx4_unary;
5792 do_32Fx4_unary:
5793 {
5794 HReg arg = iselVecExpr(env, e->Iex.Unop.arg, IEndianess);
5795 HReg dst = newVRegV(env);
5796 addInstr(env, PPCInstr_AvUn32Fx4(fpop, dst, arg));
5797 return dst;
5798 }
5799
5800 case Iop_32UtoV128: {
5801 HReg r_aligned16, r_zeros;
5802 HReg r_src = iselWordExpr_R(env, e->Iex.Unop.arg, IEndianess);
5803 HReg dst = newVRegV(env);
5804 PPCAMode *am_off0, *am_off4, *am_off8, *am_off12;
5805 sub_from_sp( env, 32 ); // Move SP down
5806
5807 /* Get a quadword aligned address within our stack space */
5808 r_aligned16 = get_sp_aligned16( env );
5809 am_off0 = PPCAMode_IR( 0, r_aligned16 );
5810 am_off4 = PPCAMode_IR( 4, r_aligned16 );
5811 am_off8 = PPCAMode_IR( 8, r_aligned16 );
5812 am_off12 = PPCAMode_IR( 12, r_aligned16 );
5813
5814 /* Store zeros */
5815 r_zeros = newVRegI(env);
5816 addInstr(env, PPCInstr_LI(r_zeros, 0x0, mode64));
5817 if (IEndianess == Iend_LE)
5818 addInstr(env, PPCInstr_Store( 4, am_off0, r_src, mode64 ));
5819 else
5820 addInstr(env, PPCInstr_Store( 4, am_off0, r_zeros, mode64 ));
5821 addInstr(env, PPCInstr_Store( 4, am_off4, r_zeros, mode64 ));
5822 addInstr(env, PPCInstr_Store( 4, am_off8, r_zeros, mode64 ));
5823
5824 /* Store r_src in low word of quadword-aligned mem */
5825 if (IEndianess == Iend_LE)
5826 addInstr(env, PPCInstr_Store( 4, am_off12, r_zeros, mode64 ));
5827 else
5828 addInstr(env, PPCInstr_Store( 4, am_off12, r_src, mode64 ));
5829
5830 /* Load word into low word of quadword vector reg */
5831 if (IEndianess == Iend_LE)
5832 addInstr(env, PPCInstr_AvLdSt( True/*ld*/, 4, dst, am_off0 ));
5833 else
5834 addInstr(env, PPCInstr_AvLdSt( True/*ld*/, 4, dst, am_off12 ));
5835
5836 add_to_sp( env, 32 ); // Reset SP
5837 return dst;
5838 }
5839
5840 case Iop_Dup8x16:
5841 case Iop_Dup16x8:
5842 case Iop_Dup32x4:
5843 return mk_AvDuplicateRI(env, e->Iex.Unop.arg, IEndianess);
5844
5845 case Iop_CipherSV128: op = Pav_CIPHERSUBV128; goto do_AvCipherV128Un;
5846 do_AvCipherV128Un: {
5847 HReg arg = iselVecExpr(env, e->Iex.Unop.arg, IEndianess);
5848 HReg dst = newVRegV(env);
5849 addInstr(env, PPCInstr_AvCipherV128Unary(op, dst, arg));
5850 return dst;
5851 }
5852
5853 case Iop_Clz8x16: op = Pav_ZEROCNTBYTE; goto do_zerocnt;
5854 case Iop_Clz16x8: op = Pav_ZEROCNTHALF; goto do_zerocnt;
5855 case Iop_Clz32x4: op = Pav_ZEROCNTWORD; goto do_zerocnt;
5856 case Iop_Clz64x2: op = Pav_ZEROCNTDBL; goto do_zerocnt;
5857 case Iop_Ctz8x16: op = Pav_TRAILINGZEROCNTBYTE; goto do_zerocnt;
5858 case Iop_Ctz16x8: op = Pav_TRAILINGZEROCNTHALF; goto do_zerocnt;
5859 case Iop_Ctz32x4: op = Pav_TRAILINGZEROCNTWORD; goto do_zerocnt;
5860 case Iop_Ctz64x2: op = Pav_TRAILINGZEROCNTDBL; goto do_zerocnt;
5861 case Iop_PwBitMtxXpose64x2: op = Pav_BITMTXXPOSE; goto do_zerocnt;
5862 do_zerocnt:
5863 {
5864 HReg arg = iselVecExpr(env, e->Iex.Unop.arg, IEndianess);
5865 HReg dst = newVRegV(env);
5866 addInstr(env, PPCInstr_AvUnary(op, dst, arg));
5867 return dst;
5868 }
5869
5870 /* BCD Iops */
5871 case Iop_BCD128toI128S:
5872 {
5873 HReg dst = newVRegV(env);
5874 HReg arg = iselVecExpr(env, e->Iex.Unop.arg, IEndianess);
5875 addInstr(env, PPCInstr_AvUnary( Pav_BCD128toI128S, dst, arg ) );
5876 return dst;
5877 }
5878
5879 case Iop_MulI128by10: op = Pav_MulI128by10; goto do_MulI128;
5880 case Iop_MulI128by10Carry: op = Pav_MulI128by10Carry; goto do_MulI128;
5881 do_MulI128: {
5882 HReg dst = newVRegV(env);
5883 HReg arg = iselVecExpr(env, e->Iex.Unop.arg, IEndianess);
5884 addInstr(env, PPCInstr_AvUnary(op, dst, arg));
5885 return dst;
5886 }
5887
5888 default:
5889 break;
5890 } /* switch (e->Iex.Unop.op) */
5891 } /* if (e->tag == Iex_Unop) */
5892
5893 if (e->tag == Iex_Binop) {
5894 switch (e->Iex.Binop.op) {
5895
5896 case Iop_64HLtoV128: {
5897 if (!mode64) {
5898 HReg r3, r2, r1, r0, r_aligned16;
5899 PPCAMode *am_off0, *am_off4, *am_off8, *am_off12;
5900 HReg dst = newVRegV(env);
5901 /* do this via the stack (easy, convenient, etc) */
5902 sub_from_sp( env, 32 ); // Move SP down
5903
5904 // get a quadword aligned address within our stack space
5905 r_aligned16 = get_sp_aligned16( env );
5906 am_off0 = PPCAMode_IR( 0, r_aligned16 );
5907 am_off4 = PPCAMode_IR( 4, r_aligned16 );
5908 am_off8 = PPCAMode_IR( 8, r_aligned16 );
5909 am_off12 = PPCAMode_IR( 12, r_aligned16 );
5910
5911 /* Do the less significant 64 bits */
5912 iselInt64Expr(&r1, &r0, env, e->Iex.Binop.arg2, IEndianess);
5913 addInstr(env, PPCInstr_Store( 4, am_off12, r0, mode64 ));
5914 addInstr(env, PPCInstr_Store( 4, am_off8, r1, mode64 ));
5915 /* Do the more significant 64 bits */
5916 iselInt64Expr(&r3, &r2, env, e->Iex.Binop.arg1, IEndianess);
5917 addInstr(env, PPCInstr_Store( 4, am_off4, r2, mode64 ));
5918 addInstr(env, PPCInstr_Store( 4, am_off0, r3, mode64 ));
5919
5920 /* Fetch result back from stack. */
5921 addInstr(env, PPCInstr_AvLdSt(True/*ld*/, 16, dst, am_off0));
5922
5923 add_to_sp( env, 32 ); // Reset SP
5924 return dst;
5925 } else {
5926 HReg rHi = iselWordExpr_R(env, e->Iex.Binop.arg1, IEndianess);
5927 HReg rLo = iselWordExpr_R(env, e->Iex.Binop.arg2, IEndianess);
5928 HReg dst = newVRegV(env);
5929 HReg r_aligned16;
5930 PPCAMode *am_off0, *am_off8;
5931 /* do this via the stack (easy, convenient, etc) */
5932 sub_from_sp( env, 32 ); // Move SP down
5933
5934 // get a quadword aligned address within our stack space
5935 r_aligned16 = get_sp_aligned16( env );
5936 am_off0 = PPCAMode_IR( 0, r_aligned16 );
5937 am_off8 = PPCAMode_IR( 8, r_aligned16 );
5938
5939 /* Store 2*I64 to stack */
5940 if (IEndianess == Iend_LE) {
5941 addInstr(env, PPCInstr_Store( 8, am_off0, rLo, mode64 ));
5942 addInstr(env, PPCInstr_Store( 8, am_off8, rHi, mode64 ));
5943 } else {
5944 addInstr(env, PPCInstr_Store( 8, am_off0, rHi, mode64 ));
5945 addInstr(env, PPCInstr_Store( 8, am_off8, rLo, mode64 ));
5946 }
5947 /* Fetch result back from stack. */
5948 addInstr(env, PPCInstr_AvLdSt(True/*ld*/, 16, dst, am_off0));
5949
5950 add_to_sp( env, 32 ); // Reset SP
5951 return dst;
5952 }
5953 }
5954
5955 case Iop_Max32Fx4: fpop = Pavfp_MAXF; goto do_32Fx4;
5956 case Iop_Min32Fx4: fpop = Pavfp_MINF; goto do_32Fx4;
5957 case Iop_CmpEQ32Fx4: fpop = Pavfp_CMPEQF; goto do_32Fx4;
5958 case Iop_CmpGT32Fx4: fpop = Pavfp_CMPGTF; goto do_32Fx4;
5959 case Iop_CmpGE32Fx4: fpop = Pavfp_CMPGEF; goto do_32Fx4;
5960 do_32Fx4:
5961 {
5962 HReg argL = iselVecExpr(env, e->Iex.Binop.arg1, IEndianess);
5963 HReg argR = iselVecExpr(env, e->Iex.Binop.arg2, IEndianess);
5964 HReg dst = newVRegV(env);
5965 addInstr(env, PPCInstr_AvBin32Fx4(fpop, dst, argL, argR));
5966 return dst;
5967 }
5968
5969 case Iop_CmpLE32Fx4: {
5970 HReg argL = iselVecExpr(env, e->Iex.Binop.arg1, IEndianess);
5971 HReg argR = iselVecExpr(env, e->Iex.Binop.arg2, IEndianess);
5972 HReg dst = newVRegV(env);
5973
5974 /* stay consistent with native ppc compares:
5975 if a left/right lane holds a nan, return zeros for that lane
5976 so: le == NOT(gt OR isNan)
5977 */
5978 HReg isNanLR = newVRegV(env);
5979 HReg isNanL = isNan(env, argL, IEndianess);
5980 HReg isNanR = isNan(env, argR, IEndianess);
5981 addInstr(env, PPCInstr_AvBinary(Pav_OR, isNanLR,
5982 isNanL, isNanR));
5983
5984 addInstr(env, PPCInstr_AvBin32Fx4(Pavfp_CMPGTF, dst,
5985 argL, argR));
5986 addInstr(env, PPCInstr_AvBinary(Pav_OR, dst, dst, isNanLR));
5987 addInstr(env, PPCInstr_AvUnary(Pav_NOT, dst, dst));
5988 return dst;
5989 }
5990
5991 case Iop_AndV128: op = Pav_AND; goto do_AvBin;
5992 case Iop_OrV128: op = Pav_OR; goto do_AvBin;
5993 case Iop_XorV128: op = Pav_XOR; goto do_AvBin;
5994 do_AvBin: {
5995 HReg arg1 = iselVecExpr(env, e->Iex.Binop.arg1, IEndianess);
5996 HReg arg2 = iselVecExpr(env, e->Iex.Binop.arg2, IEndianess);
5997 HReg dst = newVRegV(env);
5998 addInstr(env, PPCInstr_AvBinary(op, dst, arg1, arg2));
5999 return dst;
6000 }
6001
6002 case Iop_Shl8x16: op = Pav_SHL; goto do_AvBin8x16;
6003 case Iop_Shr8x16: op = Pav_SHR; goto do_AvBin8x16;
6004 case Iop_Sar8x16: op = Pav_SAR; goto do_AvBin8x16;
6005 case Iop_Rol8x16: op = Pav_ROTL; goto do_AvBin8x16;
6006 case Iop_InterleaveHI8x16: op = Pav_MRGHI; goto do_AvBin8x16;
6007 case Iop_InterleaveLO8x16: op = Pav_MRGLO; goto do_AvBin8x16;
6008 case Iop_Add8x16: op = Pav_ADDU; goto do_AvBin8x16;
6009 case Iop_QAdd8Ux16: op = Pav_QADDU; goto do_AvBin8x16;
6010 case Iop_QAdd8Sx16: op = Pav_QADDS; goto do_AvBin8x16;
6011 case Iop_Sub8x16: op = Pav_SUBU; goto do_AvBin8x16;
6012 case Iop_QSub8Ux16: op = Pav_QSUBU; goto do_AvBin8x16;
6013 case Iop_QSub8Sx16: op = Pav_QSUBS; goto do_AvBin8x16;
6014 case Iop_Avg8Ux16: op = Pav_AVGU; goto do_AvBin8x16;
6015 case Iop_Avg8Sx16: op = Pav_AVGS; goto do_AvBin8x16;
6016 case Iop_Max8Ux16: op = Pav_MAXU; goto do_AvBin8x16;
6017 case Iop_Max8Sx16: op = Pav_MAXS; goto do_AvBin8x16;
6018 case Iop_Min8Ux16: op = Pav_MINU; goto do_AvBin8x16;
6019 case Iop_Min8Sx16: op = Pav_MINS; goto do_AvBin8x16;
6020 case Iop_MullEven8Ux16: op = Pav_OMULU; goto do_AvBin8x16;
6021 case Iop_MullEven8Sx16: op = Pav_OMULS; goto do_AvBin8x16;
6022 case Iop_CmpEQ8x16: op = Pav_CMPEQU; goto do_AvBin8x16;
6023 case Iop_CmpGT8Ux16: op = Pav_CMPGTU; goto do_AvBin8x16;
6024 case Iop_CmpGT8Sx16: op = Pav_CMPGTS; goto do_AvBin8x16;
6025 case Iop_PolynomialMulAdd8x16: op = Pav_POLYMULADD; goto do_AvBin8x16;
6026 do_AvBin8x16: {
6027 HReg arg1 = iselVecExpr(env, e->Iex.Binop.arg1, IEndianess);
6028 HReg arg2 = iselVecExpr(env, e->Iex.Binop.arg2, IEndianess);
6029 HReg dst = newVRegV(env);
6030 addInstr(env, PPCInstr_AvBin8x16(op, dst, arg1, arg2));
6031 return dst;
6032 }
6033
6034 case Iop_Shl16x8: op = Pav_SHL; goto do_AvBin16x8;
6035 case Iop_Shr16x8: op = Pav_SHR; goto do_AvBin16x8;
6036 case Iop_Sar16x8: op = Pav_SAR; goto do_AvBin16x8;
6037 case Iop_Rol16x8: op = Pav_ROTL; goto do_AvBin16x8;
6038 case Iop_NarrowBin16to8x16: op = Pav_PACKUU; goto do_AvBin16x8;
6039 case Iop_QNarrowBin16Uto8Ux16: op = Pav_QPACKUU; goto do_AvBin16x8;
6040 case Iop_QNarrowBin16Sto8Sx16: op = Pav_QPACKSS; goto do_AvBin16x8;
6041 case Iop_InterleaveHI16x8: op = Pav_MRGHI; goto do_AvBin16x8;
6042 case Iop_InterleaveLO16x8: op = Pav_MRGLO; goto do_AvBin16x8;
6043 case Iop_Add16x8: op = Pav_ADDU; goto do_AvBin16x8;
6044 case Iop_QAdd16Ux8: op = Pav_QADDU; goto do_AvBin16x8;
6045 case Iop_QAdd16Sx8: op = Pav_QADDS; goto do_AvBin16x8;
6046 case Iop_Sub16x8: op = Pav_SUBU; goto do_AvBin16x8;
6047 case Iop_QSub16Ux8: op = Pav_QSUBU; goto do_AvBin16x8;
6048 case Iop_QSub16Sx8: op = Pav_QSUBS; goto do_AvBin16x8;
6049 case Iop_Avg16Ux8: op = Pav_AVGU; goto do_AvBin16x8;
6050 case Iop_Avg16Sx8: op = Pav_AVGS; goto do_AvBin16x8;
6051 case Iop_Max16Ux8: op = Pav_MAXU; goto do_AvBin16x8;
6052 case Iop_Max16Sx8: op = Pav_MAXS; goto do_AvBin16x8;
6053 case Iop_Min16Ux8: op = Pav_MINU; goto do_AvBin16x8;
6054 case Iop_Min16Sx8: op = Pav_MINS; goto do_AvBin16x8;
6055 case Iop_MullEven16Ux8: op = Pav_OMULU; goto do_AvBin16x8;
6056 case Iop_MullEven16Sx8: op = Pav_OMULS; goto do_AvBin16x8;
6057 case Iop_CmpEQ16x8: op = Pav_CMPEQU; goto do_AvBin16x8;
6058 case Iop_CmpGT16Ux8: op = Pav_CMPGTU; goto do_AvBin16x8;
6059 case Iop_CmpGT16Sx8: op = Pav_CMPGTS; goto do_AvBin16x8;
6060 case Iop_PolynomialMulAdd16x8: op = Pav_POLYMULADD; goto do_AvBin16x8;
6061 do_AvBin16x8: {
6062 HReg arg1 = iselVecExpr(env, e->Iex.Binop.arg1, IEndianess);
6063 HReg arg2 = iselVecExpr(env, e->Iex.Binop.arg2, IEndianess);
6064 HReg dst = newVRegV(env);
6065 addInstr(env, PPCInstr_AvBin16x8(op, dst, arg1, arg2));
6066 return dst;
6067 }
6068
6069 case Iop_Shl32x4: op = Pav_SHL; goto do_AvBin32x4;
6070 case Iop_Shr32x4: op = Pav_SHR; goto do_AvBin32x4;
6071 case Iop_Sar32x4: op = Pav_SAR; goto do_AvBin32x4;
6072 case Iop_Rol32x4: op = Pav_ROTL; goto do_AvBin32x4;
6073 case Iop_NarrowBin32to16x8: op = Pav_PACKUU; goto do_AvBin32x4;
6074 case Iop_QNarrowBin32Uto16Ux8: op = Pav_QPACKUU; goto do_AvBin32x4;
6075 case Iop_QNarrowBin32Sto16Sx8: op = Pav_QPACKSS; goto do_AvBin32x4;
6076 case Iop_InterleaveHI32x4: op = Pav_MRGHI; goto do_AvBin32x4;
6077 case Iop_InterleaveLO32x4: op = Pav_MRGLO; goto do_AvBin32x4;
6078 case Iop_Add32x4: op = Pav_ADDU; goto do_AvBin32x4;
6079 case Iop_QAdd32Ux4: op = Pav_QADDU; goto do_AvBin32x4;
6080 case Iop_QAdd32Sx4: op = Pav_QADDS; goto do_AvBin32x4;
6081 case Iop_Sub32x4: op = Pav_SUBU; goto do_AvBin32x4;
6082 case Iop_QSub32Ux4: op = Pav_QSUBU; goto do_AvBin32x4;
6083 case Iop_QSub32Sx4: op = Pav_QSUBS; goto do_AvBin32x4;
6084 case Iop_Avg32Ux4: op = Pav_AVGU; goto do_AvBin32x4;
6085 case Iop_Avg32Sx4: op = Pav_AVGS; goto do_AvBin32x4;
6086 case Iop_Max32Ux4: op = Pav_MAXU; goto do_AvBin32x4;
6087 case Iop_Max32Sx4: op = Pav_MAXS; goto do_AvBin32x4;
6088 case Iop_Min32Ux4: op = Pav_MINU; goto do_AvBin32x4;
6089 case Iop_Min32Sx4: op = Pav_MINS; goto do_AvBin32x4;
6090 case Iop_Mul32x4: op = Pav_MULU; goto do_AvBin32x4;
6091 case Iop_MullEven32Ux4: op = Pav_OMULU; goto do_AvBin32x4;
6092 case Iop_MullEven32Sx4: op = Pav_OMULS; goto do_AvBin32x4;
6093 case Iop_CmpEQ32x4: op = Pav_CMPEQU; goto do_AvBin32x4;
6094 case Iop_CmpGT32Ux4: op = Pav_CMPGTU; goto do_AvBin32x4;
6095 case Iop_CmpGT32Sx4: op = Pav_CMPGTS; goto do_AvBin32x4;
6096 case Iop_CatOddLanes32x4: op = Pav_CATODD; goto do_AvBin32x4;
6097 case Iop_CatEvenLanes32x4: op = Pav_CATEVEN; goto do_AvBin32x4;
6098 case Iop_PolynomialMulAdd32x4: op = Pav_POLYMULADD; goto do_AvBin32x4;
6099 do_AvBin32x4: {
6100 HReg arg1 = iselVecExpr(env, e->Iex.Binop.arg1, IEndianess);
6101 HReg arg2 = iselVecExpr(env, e->Iex.Binop.arg2, IEndianess);
6102 HReg dst = newVRegV(env);
6103 addInstr(env, PPCInstr_AvBin32x4(op, dst, arg1, arg2));
6104 return dst;
6105 }
6106
6107 case Iop_Shl64x2: op = Pav_SHL; goto do_AvBin64x2;
6108 case Iop_Shr64x2: op = Pav_SHR; goto do_AvBin64x2;
6109 case Iop_Sar64x2: op = Pav_SAR; goto do_AvBin64x2;
6110 case Iop_Rol64x2: op = Pav_ROTL; goto do_AvBin64x2;
6111 case Iop_NarrowBin64to32x4: op = Pav_PACKUU; goto do_AvBin64x2;
6112 case Iop_QNarrowBin64Sto32Sx4: op = Pav_QPACKSS; goto do_AvBin64x2;
6113 case Iop_QNarrowBin64Uto32Ux4: op = Pav_QPACKUU; goto do_AvBin64x2;
6114 case Iop_InterleaveHI64x2: op = Pav_MRGHI; goto do_AvBin64x2;
6115 case Iop_InterleaveLO64x2: op = Pav_MRGLO; goto do_AvBin64x2;
6116 case Iop_Add64x2: op = Pav_ADDU; goto do_AvBin64x2;
6117 case Iop_Sub64x2: op = Pav_SUBU; goto do_AvBin64x2;
6118 case Iop_Max64Ux2: op = Pav_MAXU; goto do_AvBin64x2;
6119 case Iop_Max64Sx2: op = Pav_MAXS; goto do_AvBin64x2;
6120 case Iop_Min64Ux2: op = Pav_MINU; goto do_AvBin64x2;
6121 case Iop_Min64Sx2: op = Pav_MINS; goto do_AvBin64x2;
6122 case Iop_CmpEQ64x2: op = Pav_CMPEQU; goto do_AvBin64x2;
6123 case Iop_CmpGT64Ux2: op = Pav_CMPGTU; goto do_AvBin64x2;
6124 case Iop_CmpGT64Sx2: op = Pav_CMPGTS; goto do_AvBin64x2;
6125 case Iop_PolynomialMulAdd64x2: op = Pav_POLYMULADD; goto do_AvBin64x2;
6126 do_AvBin64x2: {
6127 HReg arg1 = iselVecExpr(env, e->Iex.Binop.arg1, IEndianess);
6128 HReg arg2 = iselVecExpr(env, e->Iex.Binop.arg2, IEndianess);
6129 HReg dst = newVRegV(env);
6130 addInstr(env, PPCInstr_AvBin64x2(op, dst, arg1, arg2));
6131 return dst;
6132 }
6133
6134 case Iop_ShlN8x16: op = Pav_SHL; goto do_AvShift8x16;
6135 case Iop_SarN8x16: op = Pav_SAR; goto do_AvShift8x16;
6136 do_AvShift8x16: {
6137 HReg r_src = iselVecExpr(env, e->Iex.Binop.arg1, IEndianess);
6138 HReg dst = newVRegV(env);
6139 HReg v_shft = mk_AvDuplicateRI(env, e->Iex.Binop.arg2, IEndianess);
6140 addInstr(env, PPCInstr_AvBin8x16(op, dst, r_src, v_shft));
6141 return dst;
6142 }
6143
6144 case Iop_ShlN16x8: op = Pav_SHL; goto do_AvShift16x8;
6145 case Iop_ShrN16x8: op = Pav_SHR; goto do_AvShift16x8;
6146 case Iop_SarN16x8: op = Pav_SAR; goto do_AvShift16x8;
6147 do_AvShift16x8: {
6148 HReg r_src = iselVecExpr(env, e->Iex.Binop.arg1, IEndianess);
6149 HReg dst = newVRegV(env);
6150 HReg v_shft = mk_AvDuplicateRI(env, e->Iex.Binop.arg2, IEndianess);
6151 addInstr(env, PPCInstr_AvBin16x8(op, dst, r_src, v_shft));
6152 return dst;
6153 }
6154
6155 case Iop_ShlN32x4: op = Pav_SHL; goto do_AvShift32x4;
6156 case Iop_ShrN32x4: op = Pav_SHR; goto do_AvShift32x4;
6157 case Iop_SarN32x4: op = Pav_SAR; goto do_AvShift32x4;
6158 do_AvShift32x4: {
6159 HReg r_src = iselVecExpr(env, e->Iex.Binop.arg1, IEndianess);
6160 HReg dst = newVRegV(env);
6161 HReg v_shft = mk_AvDuplicateRI(env, e->Iex.Binop.arg2, IEndianess);
6162 addInstr(env, PPCInstr_AvBin32x4(op, dst, r_src, v_shft));
6163 return dst;
6164 }
6165
6166 case Iop_ShlN64x2: op = Pav_SHL; goto do_AvShift64x2;
6167 case Iop_ShrN64x2: op = Pav_SHR; goto do_AvShift64x2;
6168 case Iop_SarN64x2: op = Pav_SAR; goto do_AvShift64x2;
6169 do_AvShift64x2: {
6170 HReg r_src = iselVecExpr(env, e->Iex.Binop.arg1, IEndianess);
6171 HReg dst = newVRegV(env);
6172 HReg v_shft = mk_AvDuplicateRI(env, e->Iex.Binop.arg2, IEndianess);
6173 addInstr(env, PPCInstr_AvBin64x2(op, dst, r_src, v_shft));
6174 return dst;
6175 }
6176
6177 case Iop_ShrV128: op = Pav_SHR; goto do_AvShiftV128;
6178 case Iop_ShlV128: op = Pav_SHL; goto do_AvShiftV128;
6179 do_AvShiftV128: {
6180 HReg dst = newVRegV(env);
6181 HReg r_src = iselVecExpr(env, e->Iex.Binop.arg1, IEndianess);
6182 HReg v_shft = mk_AvDuplicateRI(env, e->Iex.Binop.arg2, IEndianess);
6183 /* Note: shift value gets masked by 127 */
6184 addInstr(env, PPCInstr_AvBinary(op, dst, r_src, v_shft));
6185 return dst;
6186 }
6187
6188 case Iop_Perm8x16: {
6189 HReg dst = newVRegV(env);
6190 HReg v_src = iselVecExpr(env, e->Iex.Binop.arg1, IEndianess);
6191 HReg v_ctl = iselVecExpr(env, e->Iex.Binop.arg2, IEndianess);
6192 addInstr(env, PPCInstr_AvPerm(dst, v_src, v_src, v_ctl));
6193 return dst;
6194 }
6195
6196 case Iop_CipherV128: op = Pav_CIPHERV128; goto do_AvCipherV128;
6197 case Iop_CipherLV128: op = Pav_CIPHERLV128; goto do_AvCipherV128;
6198 case Iop_NCipherV128: op = Pav_NCIPHERV128; goto do_AvCipherV128;
6199 case Iop_NCipherLV128:op = Pav_NCIPHERLV128; goto do_AvCipherV128;
6200 do_AvCipherV128: {
6201 HReg arg1 = iselVecExpr(env, e->Iex.Binop.arg1, IEndianess);
6202 HReg arg2 = iselVecExpr(env, e->Iex.Binop.arg2, IEndianess);
6203 HReg dst = newVRegV(env);
6204 addInstr(env, PPCInstr_AvCipherV128Binary(op, dst, arg1, arg2));
6205 return dst;
6206 }
6207
6208 case Iop_SHA256:op = Pav_SHA256; goto do_AvHashV128;
6209 case Iop_SHA512:op = Pav_SHA512; goto do_AvHashV128;
6210 do_AvHashV128: {
6211 HReg arg1 = iselVecExpr(env, e->Iex.Binop.arg1, IEndianess);
6212 HReg dst = newVRegV(env);
6213 PPCRI* s_field = iselWordExpr_RI(env, e->Iex.Binop.arg2, IEndianess);
6214 addInstr(env, PPCInstr_AvHashV128Binary(op, dst, arg1, s_field));
6215 return dst;
6216 }
6217
6218 /* BCD Iops */
6219 case Iop_I128StoBCD128:
6220 {
6221 HReg dst = newVRegV(env);
6222 HReg arg = iselVecExpr(env, e->Iex.Binop.arg1, IEndianess);
6223 PPCRI* ps = iselWordExpr_RI(env, e->Iex.Binop.arg2, IEndianess);
6224
6225 addInstr(env, PPCInstr_AvBinaryInt( Pav_I128StoBCD128, dst, arg,
6226 ps ) );
6227 return dst;
6228 }
6229
6230 case Iop_MulI128by10E: op = Pav_MulI128by10E; goto do_MulI128E;
6231 case Iop_MulI128by10ECarry: op = Pav_MulI128by10ECarry; goto do_MulI128E;
6232 do_MulI128E: {
6233 HReg dst = newVRegV(env);
6234 HReg argL = iselVecExpr(env, e->Iex.Binop.arg1, IEndianess);
6235 HReg argR = iselVecExpr(env, e->Iex.Binop.arg2, IEndianess);
6236 addInstr(env, PPCInstr_AvBinary(op, dst, argL, argR));
6237 return dst;
6238 }
6239
6240 case Iop_BCDAdd:op = Pav_BCDAdd; goto do_AvBCDV128;
6241 case Iop_BCDSub:op = Pav_BCDSub; goto do_AvBCDV128;
6242 do_AvBCDV128: {
6243 HReg arg1 = iselVecExpr(env, e->Iex.Binop.arg1, IEndianess);
6244 HReg arg2 = iselVecExpr(env, e->Iex.Binop.arg2, IEndianess);
6245 HReg dst = newVRegV(env);
6246 addInstr(env, PPCInstr_AvBCDV128Binary(op, dst, arg1, arg2));
6247 return dst;
6248 }
6249
6250 default:
6251 break;
6252 } /* switch (e->Iex.Binop.op) */
6253 } /* if (e->tag == Iex_Binop) */
6254
6255 if (e->tag == Iex_Triop) {
6256 IRTriop *triop = e->Iex.Triop.details;
6257 switch (triop->op) {
6258 case Iop_Add32Fx4: fpop = Pavfp_ADDF; goto do_32Fx4_with_rm;
6259 case Iop_Sub32Fx4: fpop = Pavfp_SUBF; goto do_32Fx4_with_rm;
6260 case Iop_Mul32Fx4: fpop = Pavfp_MULF; goto do_32Fx4_with_rm;
6261 do_32Fx4_with_rm:
6262 {
6263 HReg argL = iselVecExpr(env, triop->arg2, IEndianess);
6264 HReg argR = iselVecExpr(env, triop->arg3, IEndianess);
6265 HReg dst = newVRegV(env);
6266 /* FIXME: this is bogus, in the sense that Altivec ignores
6267 FPSCR.RM, at least for some FP operations. So setting the
6268 RM is pointless. This is only really correct in the case
6269 where the RM is known, at JIT time, to be Irrm_NEAREST,
6270 since -- at least for Altivec FP add/sub/mul -- the
6271 emitted insn is hardwired to round to nearest. */
6272 set_FPU_rounding_mode(env, triop->arg1, IEndianess);
6273 addInstr(env, PPCInstr_AvBin32Fx4(fpop, dst, argL, argR));
6274 return dst;
6275 }
6276
6277 default:
6278 break;
6279 } /* switch (e->Iex.Triop.op) */
6280 } /* if (e->tag == Iex_Trinop) */
6281
6282
6283 if (e->tag == Iex_Const ) {
6284 vassert(e->Iex.Const.con->tag == Ico_V128);
6285 if (e->Iex.Const.con->Ico.V128 == 0x0000) {
6286 return generate_zeroes_V128(env);
6287 }
6288 else if (e->Iex.Const.con->Ico.V128 == 0xffff) {
6289 return generate_ones_V128(env);
6290 }
6291 }
6292
6293 vex_printf("iselVecExpr(ppc) (subarch = %s): can't reduce\n",
6294 LibVEX_ppVexHwCaps(mode64 ? VexArchPPC64 : VexArchPPC32,
6295 env->hwcaps));
6296 ppIRExpr(e);
6297 vpanic("iselVecExpr_wrk(ppc)");
6298 }
6299
6300
6301 /*---------------------------------------------------------*/
6302 /*--- ISEL: Statements ---*/
6303 /*---------------------------------------------------------*/
6304
iselStmt(ISelEnv * env,IRStmt * stmt,IREndness IEndianess)6305 static void iselStmt ( ISelEnv* env, IRStmt* stmt, IREndness IEndianess )
6306 {
6307 Bool mode64 = env->mode64;
6308 if (vex_traceflags & VEX_TRACE_VCODE) {
6309 vex_printf("\n -- ");
6310 ppIRStmt(stmt);
6311 vex_printf("\n");
6312 }
6313
6314 switch (stmt->tag) {
6315
6316 /* --------- STORE --------- */
6317 case Ist_Store: {
6318 IRType tya = typeOfIRExpr(env->type_env, stmt->Ist.Store.addr);
6319 IRType tyd = typeOfIRExpr(env->type_env, stmt->Ist.Store.data);
6320 IREndness end = stmt->Ist.Store.end;
6321
6322 if (end != IEndianess)
6323 goto stmt_fail;
6324 if (!mode64 && (tya != Ity_I32))
6325 goto stmt_fail;
6326 if (mode64 && (tya != Ity_I64))
6327 goto stmt_fail;
6328
6329 if (tyd == Ity_I8 || tyd == Ity_I16 || tyd == Ity_I32 ||
6330 (mode64 && (tyd == Ity_I64))) {
6331 PPCAMode* am_addr
6332 = iselWordExpr_AMode(env, stmt->Ist.Store.addr, tyd/*of xfer*/,
6333 IEndianess);
6334 HReg r_src = iselWordExpr_R(env, stmt->Ist.Store.data, IEndianess);
6335 addInstr(env, PPCInstr_Store( toUChar(sizeofIRType(tyd)),
6336 am_addr, r_src, mode64 ));
6337 return;
6338 }
6339 if (tyd == Ity_F64) {
6340 PPCAMode* am_addr
6341 = iselWordExpr_AMode(env, stmt->Ist.Store.addr, tyd/*of xfer*/,
6342 IEndianess);
6343 HReg fr_src = iselDblExpr(env, stmt->Ist.Store.data, IEndianess);
6344 addInstr(env,
6345 PPCInstr_FpLdSt(False/*store*/, 8, fr_src, am_addr));
6346 return;
6347 }
6348 if (tyd == Ity_F32) {
6349 PPCAMode* am_addr
6350 = iselWordExpr_AMode(env, stmt->Ist.Store.addr, tyd/*of xfer*/,
6351 IEndianess);
6352 HReg fr_src = iselFltExpr(env, stmt->Ist.Store.data, IEndianess);
6353 addInstr(env,
6354 PPCInstr_FpLdSt(False/*store*/, 4, fr_src, am_addr));
6355 return;
6356 }
6357 if (tyd == Ity_D64) {
6358 PPCAMode* am_addr
6359 = iselWordExpr_AMode(env, stmt->Ist.Store.addr, tyd/*of xfer*/,
6360 IEndianess);
6361 HReg fr_src = iselDfp64Expr(env, stmt->Ist.Store.data, IEndianess);
6362 addInstr(env,
6363 PPCInstr_FpLdSt(False/*store*/, 8, fr_src, am_addr));
6364 return;
6365 }
6366 if (tyd == Ity_D32) {
6367 PPCAMode* am_addr
6368 = iselWordExpr_AMode(env, stmt->Ist.Store.addr, tyd/*of xfer*/,
6369 IEndianess);
6370 HReg fr_src = iselDfp32Expr(env, stmt->Ist.Store.data, IEndianess);
6371 addInstr(env,
6372 PPCInstr_FpLdSt(False/*store*/, 4, fr_src, am_addr));
6373 return;
6374 }
6375 if (tyd == Ity_V128) {
6376 PPCAMode* am_addr
6377 = iselWordExpr_AMode(env, stmt->Ist.Store.addr, tyd/*of xfer*/,
6378 IEndianess);
6379 HReg v_src = iselVecExpr(env, stmt->Ist.Store.data, IEndianess);
6380 addInstr(env,
6381 PPCInstr_AvLdSt(False/*store*/, 16, v_src, am_addr));
6382 return;
6383 }
6384 if (tyd == Ity_I64 && !mode64) {
6385 /* Just calculate the address in the register. Life is too
6386 short to arse around trying and possibly failing to adjust
6387 the offset in a 'reg+offset' style amode. */
6388 HReg rHi32, rLo32;
6389 HReg r_addr = iselWordExpr_R(env, stmt->Ist.Store.addr, IEndianess);
6390 iselInt64Expr( &rHi32, &rLo32, env, stmt->Ist.Store.data,
6391 IEndianess );
6392 addInstr(env, PPCInstr_Store( 4/*byte-store*/,
6393 PPCAMode_IR( 0, r_addr ),
6394 rHi32,
6395 False/*32-bit insn please*/) );
6396 addInstr(env, PPCInstr_Store( 4/*byte-store*/,
6397 PPCAMode_IR( 4, r_addr ),
6398 rLo32,
6399 False/*32-bit insn please*/) );
6400 return;
6401 }
6402 break;
6403 }
6404
6405 /* --------- PUT --------- */
6406 case Ist_Put: {
6407 IRType ty = typeOfIRExpr(env->type_env, stmt->Ist.Put.data);
6408 if (ty == Ity_I8 || ty == Ity_I16 ||
6409 ty == Ity_I32 || ((ty == Ity_I64) && mode64)) {
6410 HReg r_src = iselWordExpr_R(env, stmt->Ist.Put.data, IEndianess);
6411 PPCAMode* am_addr = PPCAMode_IR( stmt->Ist.Put.offset,
6412 GuestStatePtr(mode64) );
6413 addInstr(env, PPCInstr_Store( toUChar(sizeofIRType(ty)),
6414 am_addr, r_src, mode64 ));
6415 return;
6416 }
6417 if (!mode64 && ty == Ity_I64) {
6418 HReg rHi, rLo;
6419 PPCAMode* am_addr = PPCAMode_IR( stmt->Ist.Put.offset,
6420 GuestStatePtr(mode64) );
6421 PPCAMode* am_addr4 = advance4(env, am_addr);
6422 iselInt64Expr(&rHi,&rLo, env, stmt->Ist.Put.data, IEndianess);
6423 addInstr(env, PPCInstr_Store( 4, am_addr, rHi, mode64 ));
6424 addInstr(env, PPCInstr_Store( 4, am_addr4, rLo, mode64 ));
6425 return;
6426 }
6427 if (ty == Ity_I128) {
6428 HReg rHi, rLo;
6429 PPCAMode* am_addr = PPCAMode_IR( stmt->Ist.Put.offset,
6430 GuestStatePtr(mode64) );
6431 PPCAMode* am_addr4 = advance4(env, am_addr);
6432
6433 iselInt128Expr(&rHi,&rLo, env, stmt->Ist.Put.data, IEndianess);
6434 addInstr(env, PPCInstr_Store( 4, am_addr, rHi, mode64 ));
6435 addInstr(env, PPCInstr_Store( 4, am_addr4, rLo, mode64 ));
6436 return;
6437 }
6438 if (ty == Ity_F128) {
6439 /* Guest state vectors are 16byte aligned,
6440 so don't need to worry here */
6441 HReg v_src = iselFp128Expr(env, stmt->Ist.Put.data, IEndianess);
6442
6443 PPCAMode* am_addr = PPCAMode_IR( stmt->Ist.Put.offset,
6444 GuestStatePtr(mode64) );
6445 addInstr(env,
6446 PPCInstr_AvLdSt(False/*store*/, 16, v_src, am_addr));
6447 return;
6448 }
6449 if (ty == Ity_V128) {
6450 /* Guest state vectors are 16byte aligned,
6451 so don't need to worry here */
6452 HReg v_src = iselVecExpr(env, stmt->Ist.Put.data, IEndianess);
6453 PPCAMode* am_addr = PPCAMode_IR( stmt->Ist.Put.offset,
6454 GuestStatePtr(mode64) );
6455 addInstr(env,
6456 PPCInstr_AvLdSt(False/*store*/, 16, v_src, am_addr));
6457 return;
6458 }
6459 if (ty == Ity_F64) {
6460 HReg fr_src = iselDblExpr(env, stmt->Ist.Put.data, IEndianess);
6461 PPCAMode* am_addr = PPCAMode_IR( stmt->Ist.Put.offset,
6462 GuestStatePtr(mode64) );
6463 addInstr(env, PPCInstr_FpLdSt( False/*store*/, 8,
6464 fr_src, am_addr ));
6465 return;
6466 }
6467 if (ty == Ity_D32) {
6468 /* The 32-bit value is stored in a 64-bit register */
6469 HReg fr_src = iselDfp32Expr( env, stmt->Ist.Put.data, IEndianess );
6470 PPCAMode* am_addr = PPCAMode_IR( stmt->Ist.Put.offset,
6471 GuestStatePtr(mode64) );
6472 addInstr( env, PPCInstr_FpLdSt( False/*store*/, 8,
6473 fr_src, am_addr ) );
6474 return;
6475 }
6476 if (ty == Ity_D64) {
6477 HReg fr_src = iselDfp64Expr( env, stmt->Ist.Put.data, IEndianess );
6478 PPCAMode* am_addr = PPCAMode_IR( stmt->Ist.Put.offset,
6479 GuestStatePtr(mode64) );
6480 addInstr( env, PPCInstr_FpLdSt( False/*store*/, 8, fr_src, am_addr ) );
6481 return;
6482 }
6483 break;
6484 }
6485
6486 /* --------- Indexed PUT --------- */
6487 case Ist_PutI: {
6488 IRPutI *puti = stmt->Ist.PutI.details;
6489
6490 PPCAMode* dst_am
6491 = genGuestArrayOffset(
6492 env, puti->descr,
6493 puti->ix, puti->bias,
6494 IEndianess );
6495 IRType ty = typeOfIRExpr(env->type_env, puti->data);
6496 if (mode64 && ty == Ity_I64) {
6497 HReg r_src = iselWordExpr_R(env, puti->data, IEndianess);
6498 addInstr(env, PPCInstr_Store( toUChar(8),
6499 dst_am, r_src, mode64 ));
6500 return;
6501 }
6502 if ((!mode64) && ty == Ity_I32) {
6503 HReg r_src = iselWordExpr_R(env, puti->data, IEndianess);
6504 addInstr(env, PPCInstr_Store( toUChar(4),
6505 dst_am, r_src, mode64 ));
6506 return;
6507 }
6508 break;
6509 }
6510
6511 /* --------- TMP --------- */
6512 case Ist_WrTmp: {
6513 IRTemp tmp = stmt->Ist.WrTmp.tmp;
6514 IRType ty = typeOfIRTemp(env->type_env, tmp);
6515 if (ty == Ity_I8 || ty == Ity_I16 ||
6516 ty == Ity_I32 || ((ty == Ity_I64) && mode64)) {
6517 HReg r_dst = lookupIRTemp(env, tmp);
6518 HReg r_src = iselWordExpr_R(env, stmt->Ist.WrTmp.data, IEndianess);
6519 addInstr(env, mk_iMOVds_RR( r_dst, r_src ));
6520 return;
6521 }
6522 if (!mode64 && ty == Ity_I64) {
6523 HReg r_srcHi, r_srcLo, r_dstHi, r_dstLo;
6524
6525 iselInt64Expr(&r_srcHi,&r_srcLo, env, stmt->Ist.WrTmp.data,
6526 IEndianess);
6527 lookupIRTempPair( &r_dstHi, &r_dstLo, env, tmp);
6528 addInstr(env, mk_iMOVds_RR(r_dstHi, r_srcHi) );
6529 addInstr(env, mk_iMOVds_RR(r_dstLo, r_srcLo) );
6530 return;
6531 }
6532 if (mode64 && ty == Ity_I128) {
6533 HReg r_srcHi, r_srcLo, r_dstHi, r_dstLo;
6534 iselInt128Expr(&r_srcHi,&r_srcLo, env, stmt->Ist.WrTmp.data,
6535 IEndianess);
6536 lookupIRTempPair( &r_dstHi, &r_dstLo, env, tmp);
6537 addInstr(env, mk_iMOVds_RR(r_dstHi, r_srcHi) );
6538 addInstr(env, mk_iMOVds_RR(r_dstLo, r_srcLo) );
6539 return;
6540 }
6541 if (!mode64 && ty == Ity_I128) {
6542 HReg r_srcHi, r_srcMedHi, r_srcMedLo, r_srcLo;
6543 HReg r_dstHi, r_dstMedHi, r_dstMedLo, r_dstLo;
6544
6545 iselInt128Expr_to_32x4(&r_srcHi, &r_srcMedHi,
6546 &r_srcMedLo, &r_srcLo,
6547 env, stmt->Ist.WrTmp.data, IEndianess);
6548
6549 lookupIRTempQuad( &r_dstHi, &r_dstMedHi, &r_dstMedLo,
6550 &r_dstLo, env, tmp);
6551
6552 addInstr(env, mk_iMOVds_RR(r_dstHi, r_srcHi) );
6553 addInstr(env, mk_iMOVds_RR(r_dstMedHi, r_srcMedHi) );
6554 addInstr(env, mk_iMOVds_RR(r_dstMedLo, r_srcMedLo) );
6555 addInstr(env, mk_iMOVds_RR(r_dstLo, r_srcLo) );
6556 return;
6557 }
6558 if (ty == Ity_I1) {
6559 PPCCondCode cond = iselCondCode(env, stmt->Ist.WrTmp.data,
6560 IEndianess);
6561 HReg r_dst = lookupIRTemp(env, tmp);
6562 addInstr(env, PPCInstr_Set(cond, r_dst));
6563 return;
6564 }
6565 if (ty == Ity_F64) {
6566 HReg fr_dst = lookupIRTemp(env, tmp);
6567 HReg fr_src = iselDblExpr(env, stmt->Ist.WrTmp.data, IEndianess);
6568 addInstr(env, PPCInstr_FpUnary(Pfp_MOV, fr_dst, fr_src));
6569 return;
6570 }
6571 if (ty == Ity_F32) {
6572 HReg fr_dst = lookupIRTemp(env, tmp);
6573 HReg fr_src = iselFltExpr(env, stmt->Ist.WrTmp.data, IEndianess);
6574 addInstr(env, PPCInstr_FpUnary(Pfp_MOV, fr_dst, fr_src));
6575 return;
6576 }
6577 if (ty == Ity_D32) {
6578 HReg fr_dst = lookupIRTemp(env, tmp);
6579 HReg fr_src = iselDfp32Expr(env, stmt->Ist.WrTmp.data, IEndianess);
6580 addInstr(env, PPCInstr_Dfp64Unary(Pfp_MOV, fr_dst, fr_src));
6581 return;
6582 }
6583 if (ty == Ity_F128) {
6584 HReg v_dst = lookupIRTemp(env, tmp);
6585 HReg v_src = iselFp128Expr(env, stmt->Ist.WrTmp.data, IEndianess);
6586 addInstr(env, PPCInstr_AvUnary(Pav_MOV, v_dst, v_src));
6587 return;
6588 }
6589 if (ty == Ity_V128) {
6590 HReg v_dst = lookupIRTemp(env, tmp);
6591 HReg v_src = iselVecExpr(env, stmt->Ist.WrTmp.data, IEndianess);
6592 addInstr(env, PPCInstr_AvUnary(Pav_MOV, v_dst, v_src));
6593 return;
6594 }
6595 if (ty == Ity_D64) {
6596 HReg fr_dst = lookupIRTemp( env, tmp );
6597 HReg fr_src = iselDfp64Expr( env, stmt->Ist.WrTmp.data, IEndianess );
6598 addInstr( env, PPCInstr_Dfp64Unary( Pfp_MOV, fr_dst, fr_src ) );
6599 return;
6600 }
6601 if (ty == Ity_D128) {
6602 HReg fr_srcHi, fr_srcLo, fr_dstHi, fr_dstLo;
6603 // lookupDfp128IRTempPair( &fr_dstHi, &fr_dstLo, env, tmp );
6604 lookupIRTempPair( &fr_dstHi, &fr_dstLo, env, tmp );
6605 iselDfp128Expr( &fr_srcHi, &fr_srcLo, env, stmt->Ist.WrTmp.data,
6606 IEndianess );
6607 addInstr( env, PPCInstr_Dfp64Unary( Pfp_MOV, fr_dstHi, fr_srcHi ) );
6608 addInstr( env, PPCInstr_Dfp64Unary( Pfp_MOV, fr_dstLo, fr_srcLo ) );
6609 return;
6610 }
6611 break;
6612 }
6613
6614 /* --------- Load Linked or Store Conditional --------- */
6615 case Ist_LLSC: {
6616 IRTemp res = stmt->Ist.LLSC.result;
6617 IRType tyRes = typeOfIRTemp(env->type_env, res);
6618 IRType tyAddr = typeOfIRExpr(env->type_env, stmt->Ist.LLSC.addr);
6619
6620 if (stmt->Ist.LLSC.end != IEndianess)
6621 goto stmt_fail;
6622 if (!mode64 && (tyAddr != Ity_I32))
6623 goto stmt_fail;
6624 if (mode64 && (tyAddr != Ity_I64))
6625 goto stmt_fail;
6626
6627 if (stmt->Ist.LLSC.storedata == NULL) {
6628 /* LL */
6629 HReg r_addr = iselWordExpr_R( env, stmt->Ist.LLSC.addr, IEndianess );
6630 HReg r_dst = lookupIRTemp(env, res);
6631 if (tyRes == Ity_I8) {
6632 addInstr(env, PPCInstr_LoadL( 1, r_dst, r_addr, mode64 ));
6633 return;
6634 }
6635 if (tyRes == Ity_I16) {
6636 addInstr(env, PPCInstr_LoadL( 2, r_dst, r_addr, mode64 ));
6637 return;
6638 }
6639 if (tyRes == Ity_I32) {
6640 addInstr(env, PPCInstr_LoadL( 4, r_dst, r_addr, mode64 ));
6641 return;
6642 }
6643 if (tyRes == Ity_I64 && mode64) {
6644 addInstr(env, PPCInstr_LoadL( 8, r_dst, r_addr, mode64 ));
6645 return;
6646 }
6647 /* fallthru */;
6648 } else {
6649 /* SC */
6650 HReg r_res = lookupIRTemp(env, res); /* :: Ity_I1 */
6651 HReg r_a = iselWordExpr_R(env, stmt->Ist.LLSC.addr, IEndianess);
6652 HReg r_src = iselWordExpr_R(env, stmt->Ist.LLSC.storedata,
6653 IEndianess);
6654 HReg r_tmp = newVRegI(env);
6655 IRType tyData = typeOfIRExpr(env->type_env,
6656 stmt->Ist.LLSC.storedata);
6657 vassert(tyRes == Ity_I1);
6658 if (tyData == Ity_I8 || tyData == Ity_I16 || tyData == Ity_I32 ||
6659 (tyData == Ity_I64 && mode64)) {
6660 int size = 0;
6661
6662 if (tyData == Ity_I64)
6663 size = 8;
6664 else if (tyData == Ity_I32)
6665 size = 4;
6666 else if (tyData == Ity_I16)
6667 size = 2;
6668 else if (tyData == Ity_I8)
6669 size = 1;
6670
6671 addInstr(env, PPCInstr_StoreC( size,
6672 r_a, r_src, mode64 ));
6673 addInstr(env, PPCInstr_MfCR( r_tmp ));
6674 addInstr(env, PPCInstr_Shft(
6675 Pshft_SHR,
6676 env->mode64 ? False : True
6677 /*F:64-bit, T:32-bit shift*/,
6678 r_tmp, r_tmp,
6679 PPCRH_Imm(False/*unsigned*/, 29)));
6680 /* Probably unnecessary, since the IR dest type is Ity_I1,
6681 and so we are entitled to leave whatever junk we like
6682 drifting round in the upper 31 or 63 bits of r_res.
6683 However, for the sake of conservativeness .. */
6684 addInstr(env, PPCInstr_Alu(
6685 Palu_AND,
6686 r_res, r_tmp,
6687 PPCRH_Imm(False/*signed*/, 1)));
6688 return;
6689 }
6690 /* fallthru */
6691 }
6692 goto stmt_fail;
6693 /*NOTREACHED*/
6694 }
6695
6696 /* --------- Call to DIRTY helper --------- */
6697 case Ist_Dirty: {
6698 IRDirty* d = stmt->Ist.Dirty.details;
6699
6700 /* Figure out the return type, if any. */
6701 IRType retty = Ity_INVALID;
6702 if (d->tmp != IRTemp_INVALID)
6703 retty = typeOfIRTemp(env->type_env, d->tmp);
6704
6705 /* Throw out any return types we don't know about. The set of
6706 acceptable return types is the same in both 32- and 64-bit
6707 mode, so we don't need to inspect mode64 to make a
6708 decision. */
6709 Bool retty_ok = False;
6710 switch (retty) {
6711 case Ity_INVALID: /* function doesn't return anything */
6712 case Ity_V128:
6713 case Ity_I64: case Ity_I32: case Ity_I16: case Ity_I8:
6714 retty_ok = True; break;
6715 default:
6716 break;
6717 }
6718 if (!retty_ok)
6719 break; /* will go to stmt_fail: */
6720
6721 /* Marshal args, do the call, clear stack, set the return value
6722 to 0x555..555 if this is a conditional call that returns a
6723 value and the call is skipped. */
6724 UInt addToSp = 0;
6725 RetLoc rloc = mk_RetLoc_INVALID();
6726 doHelperCall( &addToSp, &rloc, env, d->guard, d->cee, retty, d->args,
6727 IEndianess );
6728 vassert(is_sane_RetLoc(rloc));
6729
6730 /* Now figure out what to do with the returned value, if any. */
6731 switch (retty) {
6732 case Ity_INVALID: {
6733 /* No return value. Nothing to do. */
6734 vassert(d->tmp == IRTemp_INVALID);
6735 vassert(rloc.pri == RLPri_None);
6736 vassert(addToSp == 0);
6737 return;
6738 }
6739 case Ity_I32: case Ity_I16: case Ity_I8: {
6740 /* The returned value is in %r3. Park it in the register
6741 associated with tmp. */
6742 HReg r_dst = lookupIRTemp(env, d->tmp);
6743 addInstr(env, mk_iMOVds_RR(r_dst, hregPPC_GPR3(mode64)));
6744 vassert(rloc.pri == RLPri_Int);
6745 vassert(addToSp == 0);
6746 return;
6747 }
6748 case Ity_I64:
6749 if (mode64) {
6750 /* The returned value is in %r3. Park it in the register
6751 associated with tmp. */
6752 HReg r_dst = lookupIRTemp(env, d->tmp);
6753 addInstr(env, mk_iMOVds_RR(r_dst, hregPPC_GPR3(mode64)));
6754 vassert(rloc.pri == RLPri_Int);
6755 vassert(addToSp == 0);
6756 } else {
6757 /* The returned value is in %r3:%r4. Park it in the
6758 register-pair associated with tmp. */
6759 HReg r_dstHi = INVALID_HREG;
6760 HReg r_dstLo = INVALID_HREG;
6761 lookupIRTempPair( &r_dstHi, &r_dstLo, env, d->tmp);
6762 addInstr(env, mk_iMOVds_RR(r_dstHi, hregPPC_GPR3(mode64)));
6763 addInstr(env, mk_iMOVds_RR(r_dstLo, hregPPC_GPR4(mode64)));
6764 vassert(rloc.pri == RLPri_2Int);
6765 vassert(addToSp == 0);
6766 }
6767 return;
6768 case Ity_V128: {
6769 /* The returned value is on the stack, and *retloc tells
6770 us where. Fish it off the stack and then move the
6771 stack pointer upwards to clear it, as directed by
6772 doHelperCall. */
6773 vassert(rloc.pri == RLPri_V128SpRel);
6774 vassert(addToSp >= 16);
6775 HReg dst = lookupIRTemp(env, d->tmp);
6776 PPCAMode* am = PPCAMode_IR(rloc.spOff, StackFramePtr(mode64));
6777 addInstr(env, PPCInstr_AvLdSt( True/*load*/, 16, dst, am ));
6778 add_to_sp(env, addToSp);
6779 return;
6780 }
6781 default:
6782 /*NOTREACHED*/
6783 vassert(0);
6784 }
6785 }
6786
6787 /* --------- MEM FENCE --------- */
6788 case Ist_MBE:
6789 switch (stmt->Ist.MBE.event) {
6790 case Imbe_Fence:
6791 addInstr(env, PPCInstr_MFence());
6792 return;
6793 default:
6794 break;
6795 }
6796 break;
6797
6798 /* --------- INSTR MARK --------- */
6799 /* Doesn't generate any executable code ... */
6800 case Ist_IMark:
6801 return;
6802
6803 /* --------- ABI HINT --------- */
6804 /* These have no meaning (denotation in the IR) and so we ignore
6805 them ... if any actually made it this far. */
6806 case Ist_AbiHint:
6807 return;
6808
6809 /* --------- NO-OP --------- */
6810 /* Fairly self-explanatory, wouldn't you say? */
6811 case Ist_NoOp:
6812 return;
6813
6814 /* --------- EXIT --------- */
6815 case Ist_Exit: {
6816 IRConst* dst = stmt->Ist.Exit.dst;
6817 if (!mode64 && dst->tag != Ico_U32)
6818 vpanic("iselStmt(ppc): Ist_Exit: dst is not a 32-bit value");
6819 if (mode64 && dst->tag != Ico_U64)
6820 vpanic("iselStmt(ppc64): Ist_Exit: dst is not a 64-bit value");
6821
6822 PPCCondCode cc = iselCondCode(env, stmt->Ist.Exit.guard, IEndianess);
6823 PPCAMode* amCIA = PPCAMode_IR(stmt->Ist.Exit.offsIP,
6824 hregPPC_GPR31(mode64));
6825
6826 /* Case: boring transfer to known address */
6827 if (stmt->Ist.Exit.jk == Ijk_Boring
6828 || stmt->Ist.Exit.jk == Ijk_Call
6829 /* || stmt->Ist.Exit.jk == Ijk_Ret */) {
6830 if (env->chainingAllowed) {
6831 /* .. almost always true .. */
6832 /* Skip the event check at the dst if this is a forwards
6833 edge. */
6834 Bool toFastEP
6835 = mode64
6836 ? (((Addr64)stmt->Ist.Exit.dst->Ico.U64) > (Addr64)env->max_ga)
6837 : (((Addr32)stmt->Ist.Exit.dst->Ico.U32) > (Addr32)env->max_ga);
6838 if (0) vex_printf("%s", toFastEP ? "Y" : ",");
6839 addInstr(env, PPCInstr_XDirect(
6840 mode64 ? (Addr64)stmt->Ist.Exit.dst->Ico.U64
6841 : (Addr64)stmt->Ist.Exit.dst->Ico.U32,
6842 amCIA, cc, toFastEP));
6843 } else {
6844 /* .. very occasionally .. */
6845 /* We can't use chaining, so ask for an assisted transfer,
6846 as that's the only alternative that is allowable. */
6847 HReg r = iselWordExpr_R(env, IRExpr_Const(stmt->Ist.Exit.dst),
6848 IEndianess);
6849 addInstr(env, PPCInstr_XAssisted(r, amCIA, cc, Ijk_Boring));
6850 }
6851 return;
6852 }
6853
6854 /* Case: assisted transfer to arbitrary address */
6855 switch (stmt->Ist.Exit.jk) {
6856 /* Keep this list in sync with that in iselNext below */
6857 case Ijk_ClientReq:
6858 case Ijk_EmFail:
6859 case Ijk_EmWarn:
6860 case Ijk_NoDecode:
6861 case Ijk_NoRedir:
6862 case Ijk_SigBUS:
6863 case Ijk_SigTRAP:
6864 case Ijk_Sys_syscall:
6865 case Ijk_InvalICache:
6866 {
6867 HReg r = iselWordExpr_R(env, IRExpr_Const(stmt->Ist.Exit.dst),
6868 IEndianess);
6869 addInstr(env, PPCInstr_XAssisted(r, amCIA, cc,
6870 stmt->Ist.Exit.jk));
6871 return;
6872 }
6873 default:
6874 break;
6875 }
6876
6877 /* Do we ever expect to see any other kind? */
6878 goto stmt_fail;
6879 }
6880
6881 default: break;
6882 }
6883 stmt_fail:
6884 ppIRStmt(stmt);
6885 vpanic("iselStmt(ppc)");
6886 }
6887
6888
6889 /*---------------------------------------------------------*/
6890 /*--- ISEL: Basic block terminators (Nexts) ---*/
6891 /*---------------------------------------------------------*/
6892
iselNext(ISelEnv * env,IRExpr * next,IRJumpKind jk,Int offsIP,IREndness IEndianess)6893 static void iselNext ( ISelEnv* env,
6894 IRExpr* next, IRJumpKind jk, Int offsIP,
6895 IREndness IEndianess)
6896 {
6897 if (vex_traceflags & VEX_TRACE_VCODE) {
6898 vex_printf( "\n-- PUT(%d) = ", offsIP);
6899 ppIRExpr( next );
6900 vex_printf( "; exit-");
6901 ppIRJumpKind(jk);
6902 vex_printf( "\n");
6903 }
6904
6905 PPCCondCode always = mk_PPCCondCode( Pct_ALWAYS, Pcf_NONE );
6906
6907 /* Case: boring transfer to known address */
6908 if (next->tag == Iex_Const) {
6909 IRConst* cdst = next->Iex.Const.con;
6910 vassert(cdst->tag == (env->mode64 ? Ico_U64 :Ico_U32));
6911 if (jk == Ijk_Boring || jk == Ijk_Call) {
6912 /* Boring transfer to known address */
6913 PPCAMode* amCIA = PPCAMode_IR(offsIP, hregPPC_GPR31(env->mode64));
6914 if (env->chainingAllowed) {
6915 /* .. almost always true .. */
6916 /* Skip the event check at the dst if this is a forwards
6917 edge. */
6918 Bool toFastEP
6919 = env->mode64
6920 ? (((Addr64)cdst->Ico.U64) > (Addr64)env->max_ga)
6921 : (((Addr32)cdst->Ico.U32) > (Addr32)env->max_ga);
6922 if (0) vex_printf("%s", toFastEP ? "X" : ".");
6923 addInstr(env, PPCInstr_XDirect(
6924 env->mode64 ? (Addr64)cdst->Ico.U64
6925 : (Addr64)cdst->Ico.U32,
6926 amCIA, always, toFastEP));
6927 } else {
6928 /* .. very occasionally .. */
6929 /* We can't use chaining, so ask for an assisted transfer,
6930 as that's the only alternative that is allowable. */
6931 HReg r = iselWordExpr_R(env, next, IEndianess);
6932 addInstr(env, PPCInstr_XAssisted(r, amCIA, always,
6933 Ijk_Boring));
6934 }
6935 return;
6936 }
6937 }
6938
6939 /* Case: call/return (==boring) transfer to any address */
6940 switch (jk) {
6941 case Ijk_Boring: case Ijk_Ret: case Ijk_Call: {
6942 HReg r = iselWordExpr_R(env, next, IEndianess);
6943 PPCAMode* amCIA = PPCAMode_IR(offsIP, hregPPC_GPR31(env->mode64));
6944 if (env->chainingAllowed) {
6945 addInstr(env, PPCInstr_XIndir(r, amCIA, always));
6946 } else {
6947 addInstr(env, PPCInstr_XAssisted(r, amCIA, always,
6948 Ijk_Boring));
6949 }
6950 return;
6951 }
6952 default:
6953 break;
6954 }
6955
6956 /* Case: assisted transfer to arbitrary address */
6957 switch (jk) {
6958 /* Keep this list in sync with that for Ist_Exit above */
6959 case Ijk_ClientReq:
6960 case Ijk_EmFail:
6961 case Ijk_EmWarn:
6962 case Ijk_NoDecode:
6963 case Ijk_NoRedir:
6964 case Ijk_SigBUS:
6965 case Ijk_SigTRAP:
6966 case Ijk_Sys_syscall:
6967 case Ijk_InvalICache:
6968 {
6969 HReg r = iselWordExpr_R(env, next, IEndianess);
6970 PPCAMode* amCIA = PPCAMode_IR(offsIP, hregPPC_GPR31(env->mode64));
6971 addInstr(env, PPCInstr_XAssisted(r, amCIA, always, jk));
6972 return;
6973 }
6974 default:
6975 break;
6976 }
6977
6978 vex_printf( "\n-- PUT(%d) = ", offsIP);
6979 ppIRExpr( next );
6980 vex_printf( "; exit-");
6981 ppIRJumpKind(jk);
6982 vex_printf( "\n");
6983 vassert(0); // are we expecting any other kind?
6984 }
6985
6986
6987 /*---------------------------------------------------------*/
6988 /*--- Insn selector top-level ---*/
6989 /*---------------------------------------------------------*/
6990
6991 /* Translate an entire SB to ppc code. */
iselSB_PPC(const IRSB * bb,VexArch arch_host,const VexArchInfo * archinfo_host,const VexAbiInfo * vbi,Int offs_Host_EvC_Counter,Int offs_Host_EvC_FailAddr,Bool chainingAllowed,Bool addProfInc,Addr max_ga)6992 HInstrArray* iselSB_PPC ( const IRSB* bb,
6993 VexArch arch_host,
6994 const VexArchInfo* archinfo_host,
6995 const VexAbiInfo* vbi,
6996 Int offs_Host_EvC_Counter,
6997 Int offs_Host_EvC_FailAddr,
6998 Bool chainingAllowed,
6999 Bool addProfInc,
7000 Addr max_ga)
7001
7002 {
7003 Int i, j;
7004 HReg hregLo, hregMedLo, hregMedHi, hregHi;
7005 ISelEnv* env;
7006 UInt hwcaps_host = archinfo_host->hwcaps;
7007 Bool mode64 = False;
7008 UInt mask32, mask64;
7009 PPCAMode *amCounter, *amFailAddr;
7010 IREndness IEndianess;
7011
7012 vassert(arch_host == VexArchPPC32 || arch_host == VexArchPPC64);
7013 mode64 = arch_host == VexArchPPC64;
7014
7015 /* do some sanity checks,
7016 * Note: no 32-bit support for ISA 3.0
7017 */
7018 mask32 = VEX_HWCAPS_PPC32_F | VEX_HWCAPS_PPC32_V
7019 | VEX_HWCAPS_PPC32_FX | VEX_HWCAPS_PPC32_GX | VEX_HWCAPS_PPC32_VX
7020 | VEX_HWCAPS_PPC32_DFP | VEX_HWCAPS_PPC32_ISA2_07;
7021
7022 mask64 = VEX_HWCAPS_PPC64_V | VEX_HWCAPS_PPC64_FX
7023 | VEX_HWCAPS_PPC64_GX | VEX_HWCAPS_PPC64_VX | VEX_HWCAPS_PPC64_DFP
7024 | VEX_HWCAPS_PPC64_ISA2_07 | VEX_HWCAPS_PPC64_ISA3_0;
7025
7026 if (mode64) {
7027 vassert((hwcaps_host & mask32) == 0);
7028 } else {
7029 vassert((hwcaps_host & mask64) == 0);
7030 }
7031
7032 /* Check that the host's endianness is as expected. */
7033 vassert((archinfo_host->endness == VexEndnessBE) ||
7034 (archinfo_host->endness == VexEndnessLE));
7035
7036 if (archinfo_host->endness == VexEndnessBE)
7037 IEndianess = Iend_BE;
7038 else
7039 IEndianess = Iend_LE;
7040
7041 /* Make up an initial environment to use. */
7042 env = LibVEX_Alloc_inline(sizeof(ISelEnv));
7043 env->vreg_ctr = 0;
7044
7045 /* Are we being ppc32 or ppc64? */
7046 env->mode64 = mode64;
7047
7048 /* Set up output code array. */
7049 env->code = newHInstrArray();
7050
7051 /* Copy BB's type env. */
7052 env->type_env = bb->tyenv;
7053
7054 /* Make up an IRTemp -> virtual HReg mapping. This doesn't
7055 * change as we go along.
7056 *
7057 * vregmap2 and vregmap3 are only used in 32 bit mode
7058 * for supporting I128 in 32-bit mode
7059 */
7060 env->n_vregmap = bb->tyenv->types_used;
7061 env->vregmapLo = LibVEX_Alloc_inline(env->n_vregmap * sizeof(HReg));
7062 env->vregmapMedLo = LibVEX_Alloc_inline(env->n_vregmap * sizeof(HReg));
7063 if (mode64) {
7064 env->vregmapMedHi = NULL;
7065 env->vregmapHi = NULL;
7066 } else {
7067 env->vregmapMedHi = LibVEX_Alloc_inline(env->n_vregmap * sizeof(HReg));
7068 env->vregmapHi = LibVEX_Alloc_inline(env->n_vregmap * sizeof(HReg));
7069 }
7070
7071 /* and finally ... */
7072 env->chainingAllowed = chainingAllowed;
7073 env->max_ga = max_ga;
7074 env->hwcaps = hwcaps_host;
7075 env->previous_rm = NULL;
7076 env->vbi = vbi;
7077
7078 /* For each IR temporary, allocate a suitably-kinded virtual
7079 register. */
7080 j = 0;
7081 for (i = 0; i < env->n_vregmap; i++) {
7082 hregLo = hregMedLo = hregMedHi = hregHi = INVALID_HREG;
7083 switch (bb->tyenv->types[i]) {
7084 case Ity_I1:
7085 case Ity_I8:
7086 case Ity_I16:
7087 case Ity_I32:
7088 if (mode64) {
7089 hregLo = mkHReg(True, HRcInt64, 0, j++);
7090 } else {
7091 hregLo = mkHReg(True, HRcInt32, 0, j++);
7092 }
7093 break;
7094 case Ity_I64:
7095 if (mode64) {
7096 hregLo = mkHReg(True, HRcInt64, 0, j++);
7097 } else {
7098 hregLo = mkHReg(True, HRcInt32, 0, j++);
7099 hregMedLo = mkHReg(True, HRcInt32, 0, j++);
7100 }
7101 break;
7102 case Ity_I128:
7103 if (mode64) {
7104 hregLo = mkHReg(True, HRcInt64, 0, j++);
7105 hregMedLo = mkHReg(True, HRcInt64, 0, j++);
7106 } else {
7107 hregLo = mkHReg(True, HRcInt32, 0, j++);
7108 hregMedLo = mkHReg(True, HRcInt32, 0, j++);
7109 hregMedHi = mkHReg(True, HRcInt32, 0, j++);
7110 hregHi = mkHReg(True, HRcInt32, 0, j++);
7111 }
7112 break;
7113 case Ity_F32:
7114 case Ity_F64:
7115 hregLo = mkHReg(True, HRcFlt64, 0, j++);
7116 break;
7117 case Ity_F128:
7118 case Ity_V128:
7119 hregLo = mkHReg(True, HRcVec128, 0, j++);
7120 break;
7121 case Ity_D32:
7122 case Ity_D64:
7123 hregLo = mkHReg(True, HRcFlt64, 0, j++);
7124 break;
7125 case Ity_D128:
7126 hregLo = mkHReg(True, HRcFlt64, 0, j++);
7127 hregMedLo = mkHReg(True, HRcFlt64, 0, j++);
7128 break;
7129 default:
7130 ppIRType(bb->tyenv->types[i]);
7131 vpanic("iselBB(ppc): IRTemp type");
7132 }
7133 env->vregmapLo[i] = hregLo;
7134 env->vregmapMedLo[i] = hregMedLo;
7135 if (!mode64) {
7136 env->vregmapMedHi[i] = hregMedHi;
7137 env->vregmapHi[i] = hregHi;
7138 }
7139 }
7140 env->vreg_ctr = j;
7141
7142 /* The very first instruction must be an event check. */
7143 amCounter = PPCAMode_IR(offs_Host_EvC_Counter, hregPPC_GPR31(mode64));
7144 amFailAddr = PPCAMode_IR(offs_Host_EvC_FailAddr, hregPPC_GPR31(mode64));
7145 addInstr(env, PPCInstr_EvCheck(amCounter, amFailAddr));
7146
7147 /* Possibly a block counter increment (for profiling). At this
7148 point we don't know the address of the counter, so just pretend
7149 it is zero. It will have to be patched later, but before this
7150 translation is used, by a call to LibVEX_patchProfCtr. */
7151 if (addProfInc) {
7152 addInstr(env, PPCInstr_ProfInc());
7153 }
7154
7155 /* Ok, finally we can iterate over the statements. */
7156 for (i = 0; i < bb->stmts_used; i++)
7157 iselStmt(env, bb->stmts[i], IEndianess);
7158
7159 iselNext(env, bb->next, bb->jumpkind, bb->offsIP, IEndianess);
7160
7161 /* record the number of vregs we used. */
7162 env->code->n_vregs = env->vreg_ctr;
7163 return env->code;
7164 }
7165
7166
7167 /*---------------------------------------------------------------*/
7168 /*--- end host_ppc_isel.c ---*/
7169 /*---------------------------------------------------------------*/
7170