1 //===- InterleavedLoadCombine.cpp - Combine Interleaved Loads ---*- C++ -*-===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // \file
10 //
11 // This file defines the interleaved-load-combine pass. The pass searches for
12 // ShuffleVectorInstruction that execute interleaving loads. If a matching
13 // pattern is found, it adds a combined load and further instructions in a
14 // pattern that is detectable by InterleavedAccesPass. The old instructions are
15 // left dead to be removed later. The pass is specifically designed to be
16 // executed just before InterleavedAccesPass to find any left-over instances
17 // that are not detected within former passes.
18 //
19 //===----------------------------------------------------------------------===//
20 
21 #include "llvm/ADT/Statistic.h"
22 #include "llvm/Analysis/MemorySSA.h"
23 #include "llvm/Analysis/MemorySSAUpdater.h"
24 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
25 #include "llvm/Analysis/TargetTransformInfo.h"
26 #include "llvm/CodeGen/Passes.h"
27 #include "llvm/CodeGen/TargetLowering.h"
28 #include "llvm/CodeGen/TargetPassConfig.h"
29 #include "llvm/CodeGen/TargetSubtargetInfo.h"
30 #include "llvm/IR/DataLayout.h"
31 #include "llvm/IR/Dominators.h"
32 #include "llvm/IR/Function.h"
33 #include "llvm/IR/IRBuilder.h"
34 #include "llvm/IR/Instructions.h"
35 #include "llvm/IR/Module.h"
36 #include "llvm/InitializePasses.h"
37 #include "llvm/Pass.h"
38 #include "llvm/Support/Debug.h"
39 #include "llvm/Support/ErrorHandling.h"
40 #include "llvm/Support/raw_ostream.h"
41 #include "llvm/Target/TargetMachine.h"
42 
43 #include <algorithm>
44 #include <cassert>
45 #include <list>
46 
47 using namespace llvm;
48 
49 #define DEBUG_TYPE "interleaved-load-combine"
50 
51 namespace {
52 
53 /// Statistic counter
54 STATISTIC(NumInterleavedLoadCombine, "Number of combined loads");
55 
56 /// Option to disable the pass
57 static cl::opt<bool> DisableInterleavedLoadCombine(
58     "disable-" DEBUG_TYPE, cl::init(false), cl::Hidden,
59     cl::desc("Disable combining of interleaved loads"));
60 
61 struct VectorInfo;
62 
63 struct InterleavedLoadCombineImpl {
64 public:
InterleavedLoadCombineImpl__anonb47475680111::InterleavedLoadCombineImpl65   InterleavedLoadCombineImpl(Function &F, DominatorTree &DT, MemorySSA &MSSA,
66                              TargetMachine &TM)
67       : F(F), DT(DT), MSSA(MSSA),
68         TLI(*TM.getSubtargetImpl(F)->getTargetLowering()),
69         TTI(TM.getTargetTransformInfo(F)) {}
70 
71   /// Scan the function for interleaved load candidates and execute the
72   /// replacement if applicable.
73   bool run();
74 
75 private:
76   /// Function this pass is working on
77   Function &F;
78 
79   /// Dominator Tree Analysis
80   DominatorTree &DT;
81 
82   /// Memory Alias Analyses
83   MemorySSA &MSSA;
84 
85   /// Target Lowering Information
86   const TargetLowering &TLI;
87 
88   /// Target Transform Information
89   const TargetTransformInfo TTI;
90 
91   /// Find the instruction in sets LIs that dominates all others, return nullptr
92   /// if there is none.
93   LoadInst *findFirstLoad(const std::set<LoadInst *> &LIs);
94 
95   /// Replace interleaved load candidates. It does additional
96   /// analyses if this makes sense. Returns true on success and false
97   /// of nothing has been changed.
98   bool combine(std::list<VectorInfo> &InterleavedLoad,
99                OptimizationRemarkEmitter &ORE);
100 
101   /// Given a set of VectorInfo containing candidates for a given interleave
102   /// factor, find a set that represents a 'factor' interleaved load.
103   bool findPattern(std::list<VectorInfo> &Candidates,
104                    std::list<VectorInfo> &InterleavedLoad, unsigned Factor,
105                    const DataLayout &DL);
106 }; // InterleavedLoadCombine
107 
108 /// First Order Polynomial on an n-Bit Integer Value
109 ///
110 /// Polynomial(Value) = Value * B + A + E*2^(n-e)
111 ///
112 /// A and B are the coefficients. E*2^(n-e) is an error within 'e' most
113 /// significant bits. It is introduced if an exact computation cannot be proven
114 /// (e.q. division by 2).
115 ///
116 /// As part of this optimization multiple loads will be combined. It necessary
117 /// to prove that loads are within some relative offset to each other. This
118 /// class is used to prove relative offsets of values loaded from memory.
119 ///
120 /// Representing an integer in this form is sound since addition in two's
121 /// complement is associative (trivial) and multiplication distributes over the
122 /// addition (see Proof(1) in Polynomial::mul). Further, both operations
123 /// commute.
124 //
125 // Example:
126 // declare @fn(i64 %IDX, <4 x float>* %PTR) {
127 //   %Pa1 = add i64 %IDX, 2
128 //   %Pa2 = lshr i64 %Pa1, 1
129 //   %Pa3 = getelementptr inbounds <4 x float>, <4 x float>* %PTR, i64 %Pa2
130 //   %Va = load <4 x float>, <4 x float>* %Pa3
131 //
132 //   %Pb1 = add i64 %IDX, 4
133 //   %Pb2 = lshr i64 %Pb1, 1
134 //   %Pb3 = getelementptr inbounds <4 x float>, <4 x float>* %PTR, i64 %Pb2
135 //   %Vb = load <4 x float>, <4 x float>* %Pb3
136 // ... }
137 //
138 // The goal is to prove that two loads load consecutive addresses.
139 //
140 // In this case the polynomials are constructed by the following
141 // steps.
142 //
143 // The number tag #e specifies the error bits.
144 //
145 // Pa_0 = %IDX              #0
146 // Pa_1 = %IDX + 2          #0 | add 2
147 // Pa_2 = %IDX/2 + 1        #1 | lshr 1
148 // Pa_3 = %IDX/2 + 1        #1 | GEP, step signext to i64
149 // Pa_4 = (%IDX/2)*16 + 16  #0 | GEP, multiply index by sizeof(4) for floats
150 // Pa_5 = (%IDX/2)*16 + 16  #0 | GEP, add offset of leading components
151 //
152 // Pb_0 = %IDX              #0
153 // Pb_1 = %IDX + 4          #0 | add 2
154 // Pb_2 = %IDX/2 + 2        #1 | lshr 1
155 // Pb_3 = %IDX/2 + 2        #1 | GEP, step signext to i64
156 // Pb_4 = (%IDX/2)*16 + 32  #0 | GEP, multiply index by sizeof(4) for floats
157 // Pb_5 = (%IDX/2)*16 + 16  #0 | GEP, add offset of leading components
158 //
159 // Pb_5 - Pa_5 = 16         #0 | subtract to get the offset
160 //
161 // Remark: %PTR is not maintained within this class. So in this instance the
162 // offset of 16 can only be assumed if the pointers are equal.
163 //
164 class Polynomial {
165   /// Operations on B
166   enum BOps {
167     LShr,
168     Mul,
169     SExt,
170     Trunc,
171   };
172 
173   /// Number of Error Bits e
174   unsigned ErrorMSBs = (unsigned)-1;
175 
176   /// Value
177   Value *V = nullptr;
178 
179   /// Coefficient B
180   SmallVector<std::pair<BOps, APInt>, 4> B;
181 
182   /// Coefficient A
183   APInt A;
184 
185 public:
Polynomial(Value * V)186   Polynomial(Value *V) : V(V) {
187     IntegerType *Ty = dyn_cast<IntegerType>(V->getType());
188     if (Ty) {
189       ErrorMSBs = 0;
190       this->V = V;
191       A = APInt(Ty->getBitWidth(), 0);
192     }
193   }
194 
Polynomial(const APInt & A,unsigned ErrorMSBs=0)195   Polynomial(const APInt &A, unsigned ErrorMSBs = 0)
196       : ErrorMSBs(ErrorMSBs), A(A) {}
197 
Polynomial(unsigned BitWidth,uint64_t A,unsigned ErrorMSBs=0)198   Polynomial(unsigned BitWidth, uint64_t A, unsigned ErrorMSBs = 0)
199       : ErrorMSBs(ErrorMSBs), A(BitWidth, A) {}
200 
201   Polynomial() = default;
202 
203   /// Increment and clamp the number of undefined bits.
incErrorMSBs(unsigned amt)204   void incErrorMSBs(unsigned amt) {
205     if (ErrorMSBs == (unsigned)-1)
206       return;
207 
208     ErrorMSBs += amt;
209     if (ErrorMSBs > A.getBitWidth())
210       ErrorMSBs = A.getBitWidth();
211   }
212 
213   /// Decrement and clamp the number of undefined bits.
decErrorMSBs(unsigned amt)214   void decErrorMSBs(unsigned amt) {
215     if (ErrorMSBs == (unsigned)-1)
216       return;
217 
218     if (ErrorMSBs > amt)
219       ErrorMSBs -= amt;
220     else
221       ErrorMSBs = 0;
222   }
223 
224   /// Apply an add on the polynomial
add(const APInt & C)225   Polynomial &add(const APInt &C) {
226     // Note: Addition is associative in two's complement even when in case of
227     // signed overflow.
228     //
229     // Error bits can only propagate into higher significant bits. As these are
230     // already regarded as undefined, there is no change.
231     //
232     // Theorem: Adding a constant to a polynomial does not change the error
233     // term.
234     //
235     // Proof:
236     //
237     //   Since the addition is associative and commutes:
238     //
239     //   (B + A + E*2^(n-e)) + C = B + (A + C) + E*2^(n-e)
240     // [qed]
241 
242     if (C.getBitWidth() != A.getBitWidth()) {
243       ErrorMSBs = (unsigned)-1;
244       return *this;
245     }
246 
247     A += C;
248     return *this;
249   }
250 
251   /// Apply a multiplication onto the polynomial.
mul(const APInt & C)252   Polynomial &mul(const APInt &C) {
253     // Note: Multiplication distributes over the addition
254     //
255     // Theorem: Multiplication distributes over the addition
256     //
257     // Proof(1):
258     //
259     //   (B+A)*C =-
260     //        = (B + A) + (B + A) + .. {C Times}
261     //         addition is associative and commutes, hence
262     //        = B + B + .. {C Times} .. + A + A + .. {C times}
263     //        = B*C + A*C
264     //   (see (function add) for signed values and overflows)
265     // [qed]
266     //
267     // Theorem: If C has c trailing zeros, errors bits in A or B are shifted out
268     // to the left.
269     //
270     // Proof(2):
271     //
272     //   Let B' and A' be the n-Bit inputs with some unknown errors EA,
273     //   EB at e leading bits. B' and A' can be written down as:
274     //
275     //     B' = B + 2^(n-e)*EB
276     //     A' = A + 2^(n-e)*EA
277     //
278     //   Let C' be an input with c trailing zero bits. C' can be written as
279     //
280     //     C' = C*2^c
281     //
282     //   Therefore we can compute the result by using distributivity and
283     //   commutativity.
284     //
285     //     (B'*C' + A'*C') = [B + 2^(n-e)*EB] * C' + [A + 2^(n-e)*EA] * C' =
286     //                     = [B + 2^(n-e)*EB + A + 2^(n-e)*EA] * C' =
287     //                     = (B'+A') * C' =
288     //                     = [B + 2^(n-e)*EB + A + 2^(n-e)*EA] * C' =
289     //                     = [B + A + 2^(n-e)*EB + 2^(n-e)*EA] * C' =
290     //                     = (B + A) * C' + [2^(n-e)*EB + 2^(n-e)*EA)] * C' =
291     //                     = (B + A) * C' + [2^(n-e)*EB + 2^(n-e)*EA)] * C*2^c =
292     //                     = (B + A) * C' + C*(EB + EA)*2^(n-e)*2^c =
293     //
294     //   Let EC be the final error with EC = C*(EB + EA)
295     //
296     //                     = (B + A)*C' + EC*2^(n-e)*2^c =
297     //                     = (B + A)*C' + EC*2^(n-(e-c))
298     //
299     //   Since EC is multiplied by 2^(n-(e-c)) the resulting error contains c
300     //   less error bits than the input. c bits are shifted out to the left.
301     // [qed]
302 
303     if (C.getBitWidth() != A.getBitWidth()) {
304       ErrorMSBs = (unsigned)-1;
305       return *this;
306     }
307 
308     // Multiplying by one is a no-op.
309     if (C.isOne()) {
310       return *this;
311     }
312 
313     // Multiplying by zero removes the coefficient B and defines all bits.
314     if (C.isZero()) {
315       ErrorMSBs = 0;
316       deleteB();
317     }
318 
319     // See Proof(2): Trailing zero bits indicate a left shift. This removes
320     // leading bits from the result even if they are undefined.
321     decErrorMSBs(C.countTrailingZeros());
322 
323     A *= C;
324     pushBOperation(Mul, C);
325     return *this;
326   }
327 
328   /// Apply a logical shift right on the polynomial
lshr(const APInt & C)329   Polynomial &lshr(const APInt &C) {
330     // Theorem(1): (B + A + E*2^(n-e)) >> 1 => (B >> 1) + (A >> 1) + E'*2^(n-e')
331     //          where
332     //             e' = e + 1,
333     //             E is a e-bit number,
334     //             E' is a e'-bit number,
335     //   holds under the following precondition:
336     //          pre(1): A % 2 = 0
337     //          pre(2): e < n, (see Theorem(2) for the trivial case with e=n)
338     //   where >> expresses a logical shift to the right, with adding zeros.
339     //
340     //  We need to show that for every, E there is a E'
341     //
342     //  B = b_h * 2^(n-1) + b_m * 2 + b_l
343     //  A = a_h * 2^(n-1) + a_m * 2         (pre(1))
344     //
345     //  where a_h, b_h, b_l are single bits, and a_m, b_m are (n-2) bit numbers
346     //
347     //  Let X = (B + A + E*2^(n-e)) >> 1
348     //  Let Y = (B >> 1) + (A >> 1) + E*2^(n-e) >> 1
349     //
350     //    X = [B + A + E*2^(n-e)] >> 1 =
351     //      = [  b_h * 2^(n-1) + b_m * 2 + b_l +
352     //         + a_h * 2^(n-1) + a_m * 2 +
353     //         + E * 2^(n-e) ] >> 1 =
354     //
355     //    The sum is built by putting the overflow of [a_m + b+n] into the term
356     //    2^(n-1). As there are no more bits beyond 2^(n-1) the overflow within
357     //    this bit is discarded. This is expressed by % 2.
358     //
359     //    The bit in position 0 cannot overflow into the term (b_m + a_m).
360     //
361     //      = [  ([b_h + a_h + (b_m + a_m) >> (n-2)] % 2) * 2^(n-1) +
362     //         + ((b_m + a_m) % 2^(n-2)) * 2 +
363     //         + b_l + E * 2^(n-e) ] >> 1 =
364     //
365     //    The shift is computed by dividing the terms by 2 and by cutting off
366     //    b_l.
367     //
368     //      =    ([b_h + a_h + (b_m + a_m) >> (n-2)] % 2) * 2^(n-2) +
369     //         + ((b_m + a_m) % 2^(n-2)) +
370     //         + E * 2^(n-(e+1)) =
371     //
372     //    by the definition in the Theorem e+1 = e'
373     //
374     //      =    ([b_h + a_h + (b_m + a_m) >> (n-2)] % 2) * 2^(n-2) +
375     //         + ((b_m + a_m) % 2^(n-2)) +
376     //         + E * 2^(n-e') =
377     //
378     //    Compute Y by applying distributivity first
379     //
380     //    Y =  (B >> 1) + (A >> 1) + E*2^(n-e') =
381     //      =    (b_h * 2^(n-1) + b_m * 2 + b_l) >> 1 +
382     //         + (a_h * 2^(n-1) + a_m * 2) >> 1 +
383     //         + E * 2^(n-e) >> 1 =
384     //
385     //    Again, the shift is computed by dividing the terms by 2 and by cutting
386     //    off b_l.
387     //
388     //      =     b_h * 2^(n-2) + b_m +
389     //         +  a_h * 2^(n-2) + a_m +
390     //         +  E * 2^(n-(e+1)) =
391     //
392     //    Again, the sum is built by putting the overflow of [a_m + b+n] into
393     //    the term 2^(n-1). But this time there is room for a second bit in the
394     //    term 2^(n-2) we add this bit to a new term and denote it o_h in a
395     //    second step.
396     //
397     //      =    ([b_h + a_h + (b_m + a_m) >> (n-2)] >> 1) * 2^(n-1) +
398     //         + ([b_h + a_h + (b_m + a_m) >> (n-2)] % 2) * 2^(n-2) +
399     //         + ((b_m + a_m) % 2^(n-2)) +
400     //         + E * 2^(n-(e+1)) =
401     //
402     //    Let o_h = [b_h + a_h + (b_m + a_m) >> (n-2)] >> 1
403     //    Further replace e+1 by e'.
404     //
405     //      =    o_h * 2^(n-1) +
406     //         + ([b_h + a_h + (b_m + a_m) >> (n-2)] % 2) * 2^(n-2) +
407     //         + ((b_m + a_m) % 2^(n-2)) +
408     //         + E * 2^(n-e') =
409     //
410     //    Move o_h into the error term and construct E'. To ensure that there is
411     //    no 2^x with negative x, this step requires pre(2) (e < n).
412     //
413     //      =    ([b_h + a_h + (b_m + a_m) >> (n-2)] % 2) * 2^(n-2) +
414     //         + ((b_m + a_m) % 2^(n-2)) +
415     //         + o_h * 2^(e'-1) * 2^(n-e') +               | pre(2), move 2^(e'-1)
416     //                                                     | out of the old exponent
417     //         + E * 2^(n-e') =
418     //      =    ([b_h + a_h + (b_m + a_m) >> (n-2)] % 2) * 2^(n-2) +
419     //         + ((b_m + a_m) % 2^(n-2)) +
420     //         + [o_h * 2^(e'-1) + E] * 2^(n-e') +         | move 2^(e'-1) out of
421     //                                                     | the old exponent
422     //
423     //    Let E' = o_h * 2^(e'-1) + E
424     //
425     //      =    ([b_h + a_h + (b_m + a_m) >> (n-2)] % 2) * 2^(n-2) +
426     //         + ((b_m + a_m) % 2^(n-2)) +
427     //         + E' * 2^(n-e')
428     //
429     //    Because X and Y are distinct only in there error terms and E' can be
430     //    constructed as shown the theorem holds.
431     // [qed]
432     //
433     // For completeness in case of the case e=n it is also required to show that
434     // distributivity can be applied.
435     //
436     // In this case Theorem(1) transforms to (the pre-condition on A can also be
437     // dropped)
438     //
439     // Theorem(2): (B + A + E) >> 1 => (B >> 1) + (A >> 1) + E'
440     //          where
441     //             A, B, E, E' are two's complement numbers with the same bit
442     //             width
443     //
444     //   Let A + B + E = X
445     //   Let (B >> 1) + (A >> 1) = Y
446     //
447     //   Therefore we need to show that for every X and Y there is an E' which
448     //   makes the equation
449     //
450     //     X = Y + E'
451     //
452     //   hold. This is trivially the case for E' = X - Y.
453     //
454     // [qed]
455     //
456     // Remark: Distributing lshr with and arbitrary number n can be expressed as
457     //   ((((B + A) lshr 1) lshr 1) ... ) {n times}.
458     // This construction induces n additional error bits at the left.
459 
460     if (C.getBitWidth() != A.getBitWidth()) {
461       ErrorMSBs = (unsigned)-1;
462       return *this;
463     }
464 
465     if (C.isZero())
466       return *this;
467 
468     // Test if the result will be zero
469     unsigned shiftAmt = C.getZExtValue();
470     if (shiftAmt >= C.getBitWidth())
471       return mul(APInt(C.getBitWidth(), 0));
472 
473     // The proof that shiftAmt LSBs are zero for at least one summand is only
474     // possible for the constant number.
475     //
476     // If this can be proven add shiftAmt to the error counter
477     // `ErrorMSBs`. Otherwise set all bits as undefined.
478     if (A.countTrailingZeros() < shiftAmt)
479       ErrorMSBs = A.getBitWidth();
480     else
481       incErrorMSBs(shiftAmt);
482 
483     // Apply the operation.
484     pushBOperation(LShr, C);
485     A = A.lshr(shiftAmt);
486 
487     return *this;
488   }
489 
490   /// Apply a sign-extend or truncate operation on the polynomial.
sextOrTrunc(unsigned n)491   Polynomial &sextOrTrunc(unsigned n) {
492     if (n < A.getBitWidth()) {
493       // Truncate: Clearly undefined Bits on the MSB side are removed
494       // if there are any.
495       decErrorMSBs(A.getBitWidth() - n);
496       A = A.trunc(n);
497       pushBOperation(Trunc, APInt(sizeof(n) * 8, n));
498     }
499     if (n > A.getBitWidth()) {
500       // Extend: Clearly extending first and adding later is different
501       // to adding first and extending later in all extended bits.
502       incErrorMSBs(n - A.getBitWidth());
503       A = A.sext(n);
504       pushBOperation(SExt, APInt(sizeof(n) * 8, n));
505     }
506 
507     return *this;
508   }
509 
510   /// Test if there is a coefficient B.
isFirstOrder() const511   bool isFirstOrder() const { return V != nullptr; }
512 
513   /// Test coefficient B of two Polynomials are equal.
isCompatibleTo(const Polynomial & o) const514   bool isCompatibleTo(const Polynomial &o) const {
515     // The polynomial use different bit width.
516     if (A.getBitWidth() != o.A.getBitWidth())
517       return false;
518 
519     // If neither Polynomial has the Coefficient B.
520     if (!isFirstOrder() && !o.isFirstOrder())
521       return true;
522 
523     // The index variable is different.
524     if (V != o.V)
525       return false;
526 
527     // Check the operations.
528     if (B.size() != o.B.size())
529       return false;
530 
531     auto *ob = o.B.begin();
532     for (const auto &b : B) {
533       if (b != *ob)
534         return false;
535       ob++;
536     }
537 
538     return true;
539   }
540 
541   /// Subtract two polynomials, return an undefined polynomial if
542   /// subtraction is not possible.
operator -(const Polynomial & o) const543   Polynomial operator-(const Polynomial &o) const {
544     // Return an undefined polynomial if incompatible.
545     if (!isCompatibleTo(o))
546       return Polynomial();
547 
548     // If the polynomials are compatible (meaning they have the same
549     // coefficient on B), B is eliminated. Thus a polynomial solely
550     // containing A is returned
551     return Polynomial(A - o.A, std::max(ErrorMSBs, o.ErrorMSBs));
552   }
553 
554   /// Subtract a constant from a polynomial,
operator -(uint64_t C) const555   Polynomial operator-(uint64_t C) const {
556     Polynomial Result(*this);
557     Result.A -= C;
558     return Result;
559   }
560 
561   /// Add a constant to a polynomial,
operator +(uint64_t C) const562   Polynomial operator+(uint64_t C) const {
563     Polynomial Result(*this);
564     Result.A += C;
565     return Result;
566   }
567 
568   /// Returns true if it can be proven that two Polynomials are equal.
isProvenEqualTo(const Polynomial & o)569   bool isProvenEqualTo(const Polynomial &o) {
570     // Subtract both polynomials and test if it is fully defined and zero.
571     Polynomial r = *this - o;
572     return (r.ErrorMSBs == 0) && (!r.isFirstOrder()) && (r.A.isZero());
573   }
574 
575   /// Print the polynomial into a stream.
print(raw_ostream & OS) const576   void print(raw_ostream &OS) const {
577     OS << "[{#ErrBits:" << ErrorMSBs << "} ";
578 
579     if (V) {
580       for (auto b : B)
581         OS << "(";
582       OS << "(" << *V << ") ";
583 
584       for (auto b : B) {
585         switch (b.first) {
586         case LShr:
587           OS << "LShr ";
588           break;
589         case Mul:
590           OS << "Mul ";
591           break;
592         case SExt:
593           OS << "SExt ";
594           break;
595         case Trunc:
596           OS << "Trunc ";
597           break;
598         }
599 
600         OS << b.second << ") ";
601       }
602     }
603 
604     OS << "+ " << A << "]";
605   }
606 
607 private:
deleteB()608   void deleteB() {
609     V = nullptr;
610     B.clear();
611   }
612 
pushBOperation(const BOps Op,const APInt & C)613   void pushBOperation(const BOps Op, const APInt &C) {
614     if (isFirstOrder()) {
615       B.push_back(std::make_pair(Op, C));
616       return;
617     }
618   }
619 };
620 
621 #ifndef NDEBUG
operator <<(raw_ostream & OS,const Polynomial & S)622 static raw_ostream &operator<<(raw_ostream &OS, const Polynomial &S) {
623   S.print(OS);
624   return OS;
625 }
626 #endif
627 
628 /// VectorInfo stores abstract the following information for each vector
629 /// element:
630 ///
631 /// 1) The the memory address loaded into the element as Polynomial
632 /// 2) a set of load instruction necessary to construct the vector,
633 /// 3) a set of all other instructions that are necessary to create the vector and
634 /// 4) a pointer value that can be used as relative base for all elements.
635 struct VectorInfo {
636 private:
VectorInfo__anonb47475680111::VectorInfo637   VectorInfo(const VectorInfo &c) : VTy(c.VTy) {
638     llvm_unreachable(
639         "Copying VectorInfo is neither implemented nor necessary,");
640   }
641 
642 public:
643   /// Information of a Vector Element
644   struct ElementInfo {
645     /// Offset Polynomial.
646     Polynomial Ofs;
647 
648     /// The Load Instruction used to Load the entry. LI is null if the pointer
649     /// of the load instruction does not point on to the entry
650     LoadInst *LI;
651 
ElementInfo__anonb47475680111::VectorInfo::ElementInfo652     ElementInfo(Polynomial Offset = Polynomial(), LoadInst *LI = nullptr)
653         : Ofs(Offset), LI(LI) {}
654   };
655 
656   /// Basic-block the load instructions are within
657   BasicBlock *BB = nullptr;
658 
659   /// Pointer value of all participation load instructions
660   Value *PV = nullptr;
661 
662   /// Participating load instructions
663   std::set<LoadInst *> LIs;
664 
665   /// Participating instructions
666   std::set<Instruction *> Is;
667 
668   /// Final shuffle-vector instruction
669   ShuffleVectorInst *SVI = nullptr;
670 
671   /// Information of the offset for each vector element
672   ElementInfo *EI;
673 
674   /// Vector Type
675   FixedVectorType *const VTy;
676 
VectorInfo__anonb47475680111::VectorInfo677   VectorInfo(FixedVectorType *VTy) : VTy(VTy) {
678     EI = new ElementInfo[VTy->getNumElements()];
679   }
680 
~VectorInfo__anonb47475680111::VectorInfo681   virtual ~VectorInfo() { delete[] EI; }
682 
getDimension__anonb47475680111::VectorInfo683   unsigned getDimension() const { return VTy->getNumElements(); }
684 
685   /// Test if the VectorInfo can be part of an interleaved load with the
686   /// specified factor.
687   ///
688   /// \param Factor of the interleave
689   /// \param DL Targets Datalayout
690   ///
691   /// \returns true if this is possible and false if not
isInterleaved__anonb47475680111::VectorInfo692   bool isInterleaved(unsigned Factor, const DataLayout &DL) const {
693     unsigned Size = DL.getTypeAllocSize(VTy->getElementType());
694     for (unsigned i = 1; i < getDimension(); i++) {
695       if (!EI[i].Ofs.isProvenEqualTo(EI[0].Ofs + i * Factor * Size)) {
696         return false;
697       }
698     }
699     return true;
700   }
701 
702   /// Recursively computes the vector information stored in V.
703   ///
704   /// This function delegates the work to specialized implementations
705   ///
706   /// \param V Value to operate on
707   /// \param Result Result of the computation
708   ///
709   /// \returns false if no sensible information can be gathered.
compute__anonb47475680111::VectorInfo710   static bool compute(Value *V, VectorInfo &Result, const DataLayout &DL) {
711     ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V);
712     if (SVI)
713       return computeFromSVI(SVI, Result, DL);
714     LoadInst *LI = dyn_cast<LoadInst>(V);
715     if (LI)
716       return computeFromLI(LI, Result, DL);
717     BitCastInst *BCI = dyn_cast<BitCastInst>(V);
718     if (BCI)
719       return computeFromBCI(BCI, Result, DL);
720     return false;
721   }
722 
723   /// BitCastInst specialization to compute the vector information.
724   ///
725   /// \param BCI BitCastInst to operate on
726   /// \param Result Result of the computation
727   ///
728   /// \returns false if no sensible information can be gathered.
computeFromBCI__anonb47475680111::VectorInfo729   static bool computeFromBCI(BitCastInst *BCI, VectorInfo &Result,
730                              const DataLayout &DL) {
731     Instruction *Op = dyn_cast<Instruction>(BCI->getOperand(0));
732 
733     if (!Op)
734       return false;
735 
736     FixedVectorType *VTy = dyn_cast<FixedVectorType>(Op->getType());
737     if (!VTy)
738       return false;
739 
740     // We can only cast from large to smaller vectors
741     if (Result.VTy->getNumElements() % VTy->getNumElements())
742       return false;
743 
744     unsigned Factor = Result.VTy->getNumElements() / VTy->getNumElements();
745     unsigned NewSize = DL.getTypeAllocSize(Result.VTy->getElementType());
746     unsigned OldSize = DL.getTypeAllocSize(VTy->getElementType());
747 
748     if (NewSize * Factor != OldSize)
749       return false;
750 
751     VectorInfo Old(VTy);
752     if (!compute(Op, Old, DL))
753       return false;
754 
755     for (unsigned i = 0; i < Result.VTy->getNumElements(); i += Factor) {
756       for (unsigned j = 0; j < Factor; j++) {
757         Result.EI[i + j] =
758             ElementInfo(Old.EI[i / Factor].Ofs + j * NewSize,
759                         j == 0 ? Old.EI[i / Factor].LI : nullptr);
760       }
761     }
762 
763     Result.BB = Old.BB;
764     Result.PV = Old.PV;
765     Result.LIs.insert(Old.LIs.begin(), Old.LIs.end());
766     Result.Is.insert(Old.Is.begin(), Old.Is.end());
767     Result.Is.insert(BCI);
768     Result.SVI = nullptr;
769 
770     return true;
771   }
772 
773   /// ShuffleVectorInst specialization to compute vector information.
774   ///
775   /// \param SVI ShuffleVectorInst to operate on
776   /// \param Result Result of the computation
777   ///
778   /// Compute the left and the right side vector information and merge them by
779   /// applying the shuffle operation. This function also ensures that the left
780   /// and right side have compatible loads. This means that all loads are with
781   /// in the same basic block and are based on the same pointer.
782   ///
783   /// \returns false if no sensible information can be gathered.
computeFromSVI__anonb47475680111::VectorInfo784   static bool computeFromSVI(ShuffleVectorInst *SVI, VectorInfo &Result,
785                              const DataLayout &DL) {
786     FixedVectorType *ArgTy =
787         cast<FixedVectorType>(SVI->getOperand(0)->getType());
788 
789     // Compute the left hand vector information.
790     VectorInfo LHS(ArgTy);
791     if (!compute(SVI->getOperand(0), LHS, DL))
792       LHS.BB = nullptr;
793 
794     // Compute the right hand vector information.
795     VectorInfo RHS(ArgTy);
796     if (!compute(SVI->getOperand(1), RHS, DL))
797       RHS.BB = nullptr;
798 
799     // Neither operand produced sensible results?
800     if (!LHS.BB && !RHS.BB)
801       return false;
802     // Only RHS produced sensible results?
803     else if (!LHS.BB) {
804       Result.BB = RHS.BB;
805       Result.PV = RHS.PV;
806     }
807     // Only LHS produced sensible results?
808     else if (!RHS.BB) {
809       Result.BB = LHS.BB;
810       Result.PV = LHS.PV;
811     }
812     // Both operands produced sensible results?
813     else if ((LHS.BB == RHS.BB) && (LHS.PV == RHS.PV)) {
814       Result.BB = LHS.BB;
815       Result.PV = LHS.PV;
816     }
817     // Both operands produced sensible results but they are incompatible.
818     else {
819       return false;
820     }
821 
822     // Merge and apply the operation on the offset information.
823     if (LHS.BB) {
824       Result.LIs.insert(LHS.LIs.begin(), LHS.LIs.end());
825       Result.Is.insert(LHS.Is.begin(), LHS.Is.end());
826     }
827     if (RHS.BB) {
828       Result.LIs.insert(RHS.LIs.begin(), RHS.LIs.end());
829       Result.Is.insert(RHS.Is.begin(), RHS.Is.end());
830     }
831     Result.Is.insert(SVI);
832     Result.SVI = SVI;
833 
834     int j = 0;
835     for (int i : SVI->getShuffleMask()) {
836       assert((i < 2 * (signed)ArgTy->getNumElements()) &&
837              "Invalid ShuffleVectorInst (index out of bounds)");
838 
839       if (i < 0)
840         Result.EI[j] = ElementInfo();
841       else if (i < (signed)ArgTy->getNumElements()) {
842         if (LHS.BB)
843           Result.EI[j] = LHS.EI[i];
844         else
845           Result.EI[j] = ElementInfo();
846       } else {
847         if (RHS.BB)
848           Result.EI[j] = RHS.EI[i - ArgTy->getNumElements()];
849         else
850           Result.EI[j] = ElementInfo();
851       }
852       j++;
853     }
854 
855     return true;
856   }
857 
858   /// LoadInst specialization to compute vector information.
859   ///
860   /// This function also acts as abort condition to the recursion.
861   ///
862   /// \param LI LoadInst to operate on
863   /// \param Result Result of the computation
864   ///
865   /// \returns false if no sensible information can be gathered.
computeFromLI__anonb47475680111::VectorInfo866   static bool computeFromLI(LoadInst *LI, VectorInfo &Result,
867                             const DataLayout &DL) {
868     Value *BasePtr;
869     Polynomial Offset;
870 
871     if (LI->isVolatile())
872       return false;
873 
874     if (LI->isAtomic())
875       return false;
876 
877     // Get the base polynomial
878     computePolynomialFromPointer(*LI->getPointerOperand(), Offset, BasePtr, DL);
879 
880     Result.BB = LI->getParent();
881     Result.PV = BasePtr;
882     Result.LIs.insert(LI);
883     Result.Is.insert(LI);
884 
885     for (unsigned i = 0; i < Result.getDimension(); i++) {
886       Value *Idx[2] = {
887           ConstantInt::get(Type::getInt32Ty(LI->getContext()), 0),
888           ConstantInt::get(Type::getInt32Ty(LI->getContext()), i),
889       };
890       int64_t Ofs = DL.getIndexedOffsetInType(Result.VTy, ArrayRef(Idx, 2));
891       Result.EI[i] = ElementInfo(Offset + Ofs, i == 0 ? LI : nullptr);
892     }
893 
894     return true;
895   }
896 
897   /// Recursively compute polynomial of a value.
898   ///
899   /// \param BO Input binary operation
900   /// \param Result Result polynomial
computePolynomialBinOp__anonb47475680111::VectorInfo901   static void computePolynomialBinOp(BinaryOperator &BO, Polynomial &Result) {
902     Value *LHS = BO.getOperand(0);
903     Value *RHS = BO.getOperand(1);
904 
905     // Find the RHS Constant if any
906     ConstantInt *C = dyn_cast<ConstantInt>(RHS);
907     if ((!C) && BO.isCommutative()) {
908       C = dyn_cast<ConstantInt>(LHS);
909       if (C)
910         std::swap(LHS, RHS);
911     }
912 
913     switch (BO.getOpcode()) {
914     case Instruction::Add:
915       if (!C)
916         break;
917 
918       computePolynomial(*LHS, Result);
919       Result.add(C->getValue());
920       return;
921 
922     case Instruction::LShr:
923       if (!C)
924         break;
925 
926       computePolynomial(*LHS, Result);
927       Result.lshr(C->getValue());
928       return;
929 
930     default:
931       break;
932     }
933 
934     Result = Polynomial(&BO);
935   }
936 
937   /// Recursively compute polynomial of a value
938   ///
939   /// \param V input value
940   /// \param Result result polynomial
computePolynomial__anonb47475680111::VectorInfo941   static void computePolynomial(Value &V, Polynomial &Result) {
942     if (auto *BO = dyn_cast<BinaryOperator>(&V))
943       computePolynomialBinOp(*BO, Result);
944     else
945       Result = Polynomial(&V);
946   }
947 
948   /// Compute the Polynomial representation of a Pointer type.
949   ///
950   /// \param Ptr input pointer value
951   /// \param Result result polynomial
952   /// \param BasePtr pointer the polynomial is based on
953   /// \param DL Datalayout of the target machine
computePolynomialFromPointer__anonb47475680111::VectorInfo954   static void computePolynomialFromPointer(Value &Ptr, Polynomial &Result,
955                                            Value *&BasePtr,
956                                            const DataLayout &DL) {
957     // Not a pointer type? Return an undefined polynomial
958     PointerType *PtrTy = dyn_cast<PointerType>(Ptr.getType());
959     if (!PtrTy) {
960       Result = Polynomial();
961       BasePtr = nullptr;
962       return;
963     }
964     unsigned PointerBits =
965         DL.getIndexSizeInBits(PtrTy->getPointerAddressSpace());
966 
967     /// Skip pointer casts. Return Zero polynomial otherwise
968     if (isa<CastInst>(&Ptr)) {
969       CastInst &CI = *cast<CastInst>(&Ptr);
970       switch (CI.getOpcode()) {
971       case Instruction::BitCast:
972         computePolynomialFromPointer(*CI.getOperand(0), Result, BasePtr, DL);
973         break;
974       default:
975         BasePtr = &Ptr;
976         Polynomial(PointerBits, 0);
977         break;
978       }
979     }
980     /// Resolve GetElementPtrInst.
981     else if (isa<GetElementPtrInst>(&Ptr)) {
982       GetElementPtrInst &GEP = *cast<GetElementPtrInst>(&Ptr);
983 
984       APInt BaseOffset(PointerBits, 0);
985 
986       // Check if we can compute the Offset with accumulateConstantOffset
987       if (GEP.accumulateConstantOffset(DL, BaseOffset)) {
988         Result = Polynomial(BaseOffset);
989         BasePtr = GEP.getPointerOperand();
990         return;
991       } else {
992         // Otherwise we allow that the last index operand of the GEP is
993         // non-constant.
994         unsigned idxOperand, e;
995         SmallVector<Value *, 4> Indices;
996         for (idxOperand = 1, e = GEP.getNumOperands(); idxOperand < e;
997              idxOperand++) {
998           ConstantInt *IDX = dyn_cast<ConstantInt>(GEP.getOperand(idxOperand));
999           if (!IDX)
1000             break;
1001           Indices.push_back(IDX);
1002         }
1003 
1004         // It must also be the last operand.
1005         if (idxOperand + 1 != e) {
1006           Result = Polynomial();
1007           BasePtr = nullptr;
1008           return;
1009         }
1010 
1011         // Compute the polynomial of the index operand.
1012         computePolynomial(*GEP.getOperand(idxOperand), Result);
1013 
1014         // Compute base offset from zero based index, excluding the last
1015         // variable operand.
1016         BaseOffset =
1017             DL.getIndexedOffsetInType(GEP.getSourceElementType(), Indices);
1018 
1019         // Apply the operations of GEP to the polynomial.
1020         unsigned ResultSize = DL.getTypeAllocSize(GEP.getResultElementType());
1021         Result.sextOrTrunc(PointerBits);
1022         Result.mul(APInt(PointerBits, ResultSize));
1023         Result.add(BaseOffset);
1024         BasePtr = GEP.getPointerOperand();
1025       }
1026     }
1027     // All other instructions are handled by using the value as base pointer and
1028     // a zero polynomial.
1029     else {
1030       BasePtr = &Ptr;
1031       Polynomial(DL.getIndexSizeInBits(PtrTy->getPointerAddressSpace()), 0);
1032     }
1033   }
1034 
1035 #ifndef NDEBUG
print__anonb47475680111::VectorInfo1036   void print(raw_ostream &OS) const {
1037     if (PV)
1038       OS << *PV;
1039     else
1040       OS << "(none)";
1041     OS << " + ";
1042     for (unsigned i = 0; i < getDimension(); i++)
1043       OS << ((i == 0) ? "[" : ", ") << EI[i].Ofs;
1044     OS << "]";
1045   }
1046 #endif
1047 };
1048 
1049 } // anonymous namespace
1050 
findPattern(std::list<VectorInfo> & Candidates,std::list<VectorInfo> & InterleavedLoad,unsigned Factor,const DataLayout & DL)1051 bool InterleavedLoadCombineImpl::findPattern(
1052     std::list<VectorInfo> &Candidates, std::list<VectorInfo> &InterleavedLoad,
1053     unsigned Factor, const DataLayout &DL) {
1054   for (auto C0 = Candidates.begin(), E0 = Candidates.end(); C0 != E0; ++C0) {
1055     unsigned i;
1056     // Try to find an interleaved load using the front of Worklist as first line
1057     unsigned Size = DL.getTypeAllocSize(C0->VTy->getElementType());
1058 
1059     // List containing iterators pointing to the VectorInfos of the candidates
1060     std::vector<std::list<VectorInfo>::iterator> Res(Factor, Candidates.end());
1061 
1062     for (auto C = Candidates.begin(), E = Candidates.end(); C != E; C++) {
1063       if (C->VTy != C0->VTy)
1064         continue;
1065       if (C->BB != C0->BB)
1066         continue;
1067       if (C->PV != C0->PV)
1068         continue;
1069 
1070       // Check the current value matches any of factor - 1 remaining lines
1071       for (i = 1; i < Factor; i++) {
1072         if (C->EI[0].Ofs.isProvenEqualTo(C0->EI[0].Ofs + i * Size)) {
1073           Res[i] = C;
1074         }
1075       }
1076 
1077       for (i = 1; i < Factor; i++) {
1078         if (Res[i] == Candidates.end())
1079           break;
1080       }
1081       if (i == Factor) {
1082         Res[0] = C0;
1083         break;
1084       }
1085     }
1086 
1087     if (Res[0] != Candidates.end()) {
1088       // Move the result into the output
1089       for (unsigned i = 0; i < Factor; i++) {
1090         InterleavedLoad.splice(InterleavedLoad.end(), Candidates, Res[i]);
1091       }
1092 
1093       return true;
1094     }
1095   }
1096   return false;
1097 }
1098 
1099 LoadInst *
findFirstLoad(const std::set<LoadInst * > & LIs)1100 InterleavedLoadCombineImpl::findFirstLoad(const std::set<LoadInst *> &LIs) {
1101   assert(!LIs.empty() && "No load instructions given.");
1102 
1103   // All LIs are within the same BB. Select the first for a reference.
1104   BasicBlock *BB = (*LIs.begin())->getParent();
1105   BasicBlock::iterator FLI = llvm::find_if(
1106       *BB, [&LIs](Instruction &I) -> bool { return is_contained(LIs, &I); });
1107   assert(FLI != BB->end());
1108 
1109   return cast<LoadInst>(FLI);
1110 }
1111 
combine(std::list<VectorInfo> & InterleavedLoad,OptimizationRemarkEmitter & ORE)1112 bool InterleavedLoadCombineImpl::combine(std::list<VectorInfo> &InterleavedLoad,
1113                                          OptimizationRemarkEmitter &ORE) {
1114   LLVM_DEBUG(dbgs() << "Checking interleaved load\n");
1115 
1116   // The insertion point is the LoadInst which loads the first values. The
1117   // following tests are used to proof that the combined load can be inserted
1118   // just before InsertionPoint.
1119   LoadInst *InsertionPoint = InterleavedLoad.front().EI[0].LI;
1120 
1121   // Test if the offset is computed
1122   if (!InsertionPoint)
1123     return false;
1124 
1125   std::set<LoadInst *> LIs;
1126   std::set<Instruction *> Is;
1127   std::set<Instruction *> SVIs;
1128 
1129   InstructionCost InterleavedCost;
1130   InstructionCost InstructionCost = 0;
1131   const TTI::TargetCostKind CostKind = TTI::TCK_SizeAndLatency;
1132 
1133   // Get the interleave factor
1134   unsigned Factor = InterleavedLoad.size();
1135 
1136   // Merge all input sets used in analysis
1137   for (auto &VI : InterleavedLoad) {
1138     // Generate a set of all load instructions to be combined
1139     LIs.insert(VI.LIs.begin(), VI.LIs.end());
1140 
1141     // Generate a set of all instructions taking part in load
1142     // interleaved. This list excludes the instructions necessary for the
1143     // polynomial construction.
1144     Is.insert(VI.Is.begin(), VI.Is.end());
1145 
1146     // Generate the set of the final ShuffleVectorInst.
1147     SVIs.insert(VI.SVI);
1148   }
1149 
1150   // There is nothing to combine.
1151   if (LIs.size() < 2)
1152     return false;
1153 
1154   // Test if all participating instruction will be dead after the
1155   // transformation. If intermediate results are used, no performance gain can
1156   // be expected. Also sum the cost of the Instructions beeing left dead.
1157   for (const auto &I : Is) {
1158     // Compute the old cost
1159     InstructionCost += TTI.getInstructionCost(I, CostKind);
1160 
1161     // The final SVIs are allowed not to be dead, all uses will be replaced
1162     if (SVIs.find(I) != SVIs.end())
1163       continue;
1164 
1165     // If there are users outside the set to be eliminated, we abort the
1166     // transformation. No gain can be expected.
1167     for (auto *U : I->users()) {
1168       if (Is.find(dyn_cast<Instruction>(U)) == Is.end())
1169         return false;
1170     }
1171   }
1172 
1173   // We need to have a valid cost in order to proceed.
1174   if (!InstructionCost.isValid())
1175     return false;
1176 
1177   // We know that all LoadInst are within the same BB. This guarantees that
1178   // either everything or nothing is loaded.
1179   LoadInst *First = findFirstLoad(LIs);
1180 
1181   // To be safe that the loads can be combined, iterate over all loads and test
1182   // that the corresponding defining access dominates first LI. This guarantees
1183   // that there are no aliasing stores in between the loads.
1184   auto FMA = MSSA.getMemoryAccess(First);
1185   for (auto *LI : LIs) {
1186     auto MADef = MSSA.getMemoryAccess(LI)->getDefiningAccess();
1187     if (!MSSA.dominates(MADef, FMA))
1188       return false;
1189   }
1190   assert(!LIs.empty() && "There are no LoadInst to combine");
1191 
1192   // It is necessary that insertion point dominates all final ShuffleVectorInst.
1193   for (auto &VI : InterleavedLoad) {
1194     if (!DT.dominates(InsertionPoint, VI.SVI))
1195       return false;
1196   }
1197 
1198   // All checks are done. Add instructions detectable by InterleavedAccessPass
1199   // The old instruction will are left dead.
1200   IRBuilder<> Builder(InsertionPoint);
1201   Type *ETy = InterleavedLoad.front().SVI->getType()->getElementType();
1202   unsigned ElementsPerSVI =
1203       cast<FixedVectorType>(InterleavedLoad.front().SVI->getType())
1204           ->getNumElements();
1205   FixedVectorType *ILTy = FixedVectorType::get(ETy, Factor * ElementsPerSVI);
1206 
1207   auto Indices = llvm::to_vector<4>(llvm::seq<unsigned>(0, Factor));
1208   InterleavedCost = TTI.getInterleavedMemoryOpCost(
1209       Instruction::Load, ILTy, Factor, Indices, InsertionPoint->getAlign(),
1210       InsertionPoint->getPointerAddressSpace(), CostKind);
1211 
1212   if (InterleavedCost >= InstructionCost) {
1213     return false;
1214   }
1215 
1216   // Create a pointer cast for the wide load.
1217   auto CI = Builder.CreatePointerCast(InsertionPoint->getOperand(0),
1218                                       ILTy->getPointerTo(),
1219                                       "interleaved.wide.ptrcast");
1220 
1221   // Create the wide load and update the MemorySSA.
1222   auto LI = Builder.CreateAlignedLoad(ILTy, CI, InsertionPoint->getAlign(),
1223                                       "interleaved.wide.load");
1224   auto MSSAU = MemorySSAUpdater(&MSSA);
1225   MemoryUse *MSSALoad = cast<MemoryUse>(MSSAU.createMemoryAccessBefore(
1226       LI, nullptr, MSSA.getMemoryAccess(InsertionPoint)));
1227   MSSAU.insertUse(MSSALoad, /*RenameUses=*/ true);
1228 
1229   // Create the final SVIs and replace all uses.
1230   int i = 0;
1231   for (auto &VI : InterleavedLoad) {
1232     SmallVector<int, 4> Mask;
1233     for (unsigned j = 0; j < ElementsPerSVI; j++)
1234       Mask.push_back(i + j * Factor);
1235 
1236     Builder.SetInsertPoint(VI.SVI);
1237     auto SVI = Builder.CreateShuffleVector(LI, Mask, "interleaved.shuffle");
1238     VI.SVI->replaceAllUsesWith(SVI);
1239     i++;
1240   }
1241 
1242   NumInterleavedLoadCombine++;
1243   ORE.emit([&]() {
1244     return OptimizationRemark(DEBUG_TYPE, "Combined Interleaved Load", LI)
1245            << "Load interleaved combined with factor "
1246            << ore::NV("Factor", Factor);
1247   });
1248 
1249   return true;
1250 }
1251 
run()1252 bool InterleavedLoadCombineImpl::run() {
1253   OptimizationRemarkEmitter ORE(&F);
1254   bool changed = false;
1255   unsigned MaxFactor = TLI.getMaxSupportedInterleaveFactor();
1256 
1257   auto &DL = F.getParent()->getDataLayout();
1258 
1259   // Start with the highest factor to avoid combining and recombining.
1260   for (unsigned Factor = MaxFactor; Factor >= 2; Factor--) {
1261     std::list<VectorInfo> Candidates;
1262 
1263     for (BasicBlock &BB : F) {
1264       for (Instruction &I : BB) {
1265         if (auto SVI = dyn_cast<ShuffleVectorInst>(&I)) {
1266           // We don't support scalable vectors in this pass.
1267           if (isa<ScalableVectorType>(SVI->getType()))
1268             continue;
1269 
1270           Candidates.emplace_back(cast<FixedVectorType>(SVI->getType()));
1271 
1272           if (!VectorInfo::computeFromSVI(SVI, Candidates.back(), DL)) {
1273             Candidates.pop_back();
1274             continue;
1275           }
1276 
1277           if (!Candidates.back().isInterleaved(Factor, DL)) {
1278             Candidates.pop_back();
1279           }
1280         }
1281       }
1282     }
1283 
1284     std::list<VectorInfo> InterleavedLoad;
1285     while (findPattern(Candidates, InterleavedLoad, Factor, DL)) {
1286       if (combine(InterleavedLoad, ORE)) {
1287         changed = true;
1288       } else {
1289         // Remove the first element of the Interleaved Load but put the others
1290         // back on the list and continue searching
1291         Candidates.splice(Candidates.begin(), InterleavedLoad,
1292                           std::next(InterleavedLoad.begin()),
1293                           InterleavedLoad.end());
1294       }
1295       InterleavedLoad.clear();
1296     }
1297   }
1298 
1299   return changed;
1300 }
1301 
1302 namespace {
1303 /// This pass combines interleaved loads into a pattern detectable by
1304 /// InterleavedAccessPass.
1305 struct InterleavedLoadCombine : public FunctionPass {
1306   static char ID;
1307 
InterleavedLoadCombine__anonb47475680411::InterleavedLoadCombine1308   InterleavedLoadCombine() : FunctionPass(ID) {
1309     initializeInterleavedLoadCombinePass(*PassRegistry::getPassRegistry());
1310   }
1311 
getPassName__anonb47475680411::InterleavedLoadCombine1312   StringRef getPassName() const override {
1313     return "Interleaved Load Combine Pass";
1314   }
1315 
runOnFunction__anonb47475680411::InterleavedLoadCombine1316   bool runOnFunction(Function &F) override {
1317     if (DisableInterleavedLoadCombine)
1318       return false;
1319 
1320     auto *TPC = getAnalysisIfAvailable<TargetPassConfig>();
1321     if (!TPC)
1322       return false;
1323 
1324     LLVM_DEBUG(dbgs() << "*** " << getPassName() << ": " << F.getName()
1325                       << "\n");
1326 
1327     return InterleavedLoadCombineImpl(
1328                F, getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
1329                getAnalysis<MemorySSAWrapperPass>().getMSSA(),
1330                TPC->getTM<TargetMachine>())
1331         .run();
1332   }
1333 
getAnalysisUsage__anonb47475680411::InterleavedLoadCombine1334   void getAnalysisUsage(AnalysisUsage &AU) const override {
1335     AU.addRequired<MemorySSAWrapperPass>();
1336     AU.addRequired<DominatorTreeWrapperPass>();
1337     FunctionPass::getAnalysisUsage(AU);
1338   }
1339 
1340 private:
1341 };
1342 } // anonymous namespace
1343 
1344 char InterleavedLoadCombine::ID = 0;
1345 
1346 INITIALIZE_PASS_BEGIN(
1347     InterleavedLoadCombine, DEBUG_TYPE,
1348     "Combine interleaved loads into wide loads and shufflevector instructions",
1349     false, false)
INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)1350 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1351 INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)
1352 INITIALIZE_PASS_END(
1353     InterleavedLoadCombine, DEBUG_TYPE,
1354     "Combine interleaved loads into wide loads and shufflevector instructions",
1355     false, false)
1356 
1357 FunctionPass *
1358 llvm::createInterleavedLoadCombinePass() {
1359   auto P = new InterleavedLoadCombine();
1360   return P;
1361 }
1362