1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc -mtriple=riscv32 -mattr=+d -verify-machineinstrs < %s \
3; RUN:   | FileCheck -check-prefix=RV32IFD %s
4; RUN: llc -mtriple=riscv64 -mattr=+d -verify-machineinstrs < %s \
5; RUN:   | FileCheck -check-prefix=RV64IFD %s
6
7define double @fld(double *%a) nounwind {
8; RV32IFD-LABEL: fld:
9; RV32IFD:       # %bb.0:
10; RV32IFD-NEXT:    addi sp, sp, -16
11; RV32IFD-NEXT:    fld ft0, 0(a0)
12; RV32IFD-NEXT:    fld ft1, 24(a0)
13; RV32IFD-NEXT:    fadd.d ft0, ft0, ft1
14; RV32IFD-NEXT:    fsd ft0, 8(sp)
15; RV32IFD-NEXT:    lw a0, 8(sp)
16; RV32IFD-NEXT:    lw a1, 12(sp)
17; RV32IFD-NEXT:    addi sp, sp, 16
18; RV32IFD-NEXT:    ret
19;
20; RV64IFD-LABEL: fld:
21; RV64IFD:       # %bb.0:
22; RV64IFD-NEXT:    fld ft0, 0(a0)
23; RV64IFD-NEXT:    fld ft1, 24(a0)
24; RV64IFD-NEXT:    fadd.d ft0, ft0, ft1
25; RV64IFD-NEXT:    fmv.x.d a0, ft0
26; RV64IFD-NEXT:    ret
27  %1 = load double, double* %a
28  %2 = getelementptr double, double* %a, i32 3
29  %3 = load double, double* %2
30; Use both loaded values in an FP op to ensure an fld is used, even for the
31; soft float ABI
32  %4 = fadd double %1, %3
33  ret double %4
34}
35
36define void @fsd(double *%a, double %b, double %c) nounwind {
37; RV32IFD-LABEL: fsd:
38; RV32IFD:       # %bb.0:
39; RV32IFD-NEXT:    addi sp, sp, -16
40; RV32IFD-NEXT:    sw a3, 8(sp)
41; RV32IFD-NEXT:    sw a4, 12(sp)
42; RV32IFD-NEXT:    fld ft0, 8(sp)
43; RV32IFD-NEXT:    sw a1, 8(sp)
44; RV32IFD-NEXT:    sw a2, 12(sp)
45; RV32IFD-NEXT:    fld ft1, 8(sp)
46; RV32IFD-NEXT:    fadd.d ft0, ft1, ft0
47; RV32IFD-NEXT:    fsd ft0, 0(a0)
48; RV32IFD-NEXT:    fsd ft0, 64(a0)
49; RV32IFD-NEXT:    addi sp, sp, 16
50; RV32IFD-NEXT:    ret
51;
52; RV64IFD-LABEL: fsd:
53; RV64IFD:       # %bb.0:
54; RV64IFD-NEXT:    fmv.d.x ft0, a2
55; RV64IFD-NEXT:    fmv.d.x ft1, a1
56; RV64IFD-NEXT:    fadd.d ft0, ft1, ft0
57; RV64IFD-NEXT:    fsd ft0, 0(a0)
58; RV64IFD-NEXT:    fsd ft0, 64(a0)
59; RV64IFD-NEXT:    ret
60; Use %b and %c in an FP op to ensure floating point registers are used, even
61; for the soft float ABI
62  %1 = fadd double %b, %c
63  store double %1, double* %a
64  %2 = getelementptr double, double* %a, i32 8
65  store double %1, double* %2
66  ret void
67}
68
69; Check load and store to a global
70@G = global double 0.0
71
72define double @fld_fsd_global(double %a, double %b) nounwind {
73; RV32IFD-LABEL: fld_fsd_global:
74; RV32IFD:       # %bb.0:
75; RV32IFD-NEXT:    addi sp, sp, -16
76; RV32IFD-NEXT:    sw a2, 8(sp)
77; RV32IFD-NEXT:    sw a3, 12(sp)
78; RV32IFD-NEXT:    fld ft0, 8(sp)
79; RV32IFD-NEXT:    sw a0, 8(sp)
80; RV32IFD-NEXT:    sw a1, 12(sp)
81; RV32IFD-NEXT:    fld ft1, 8(sp)
82; RV32IFD-NEXT:    fadd.d ft0, ft1, ft0
83; RV32IFD-NEXT:    lui a0, %hi(G)
84; RV32IFD-NEXT:    fld ft1, %lo(G)(a0)
85; RV32IFD-NEXT:    fsd ft0, %lo(G)(a0)
86; RV32IFD-NEXT:    addi a0, a0, %lo(G)
87; RV32IFD-NEXT:    fld ft1, 72(a0)
88; RV32IFD-NEXT:    fsd ft0, 72(a0)
89; RV32IFD-NEXT:    fsd ft0, 8(sp)
90; RV32IFD-NEXT:    lw a0, 8(sp)
91; RV32IFD-NEXT:    lw a1, 12(sp)
92; RV32IFD-NEXT:    addi sp, sp, 16
93; RV32IFD-NEXT:    ret
94;
95; RV64IFD-LABEL: fld_fsd_global:
96; RV64IFD:       # %bb.0:
97; RV64IFD-NEXT:    fmv.d.x ft0, a1
98; RV64IFD-NEXT:    fmv.d.x ft1, a0
99; RV64IFD-NEXT:    fadd.d ft0, ft1, ft0
100; RV64IFD-NEXT:    lui a0, %hi(G)
101; RV64IFD-NEXT:    fld ft1, %lo(G)(a0)
102; RV64IFD-NEXT:    fsd ft0, %lo(G)(a0)
103; RV64IFD-NEXT:    addi a1, a0, %lo(G)
104; RV64IFD-NEXT:    fld ft1, 72(a1)
105; RV64IFD-NEXT:    fmv.x.d a0, ft0
106; RV64IFD-NEXT:    fsd ft0, 72(a1)
107; RV64IFD-NEXT:    ret
108; Use %a and %b in an FP op to ensure floating point registers are used, even
109; for the soft float ABI
110  %1 = fadd double %a, %b
111  %2 = load volatile double, double* @G
112  store double %1, double* @G
113  %3 = getelementptr double, double* @G, i32 9
114  %4 = load volatile double, double* %3
115  store double %1, double* %3
116  ret double %1
117}
118
119; Ensure that 1 is added to the high 20 bits if bit 11 of the low part is 1
120define double @fld_fsd_constant(double %a) nounwind {
121; RV32IFD-LABEL: fld_fsd_constant:
122; RV32IFD:       # %bb.0:
123; RV32IFD-NEXT:    addi sp, sp, -16
124; RV32IFD-NEXT:    sw a0, 8(sp)
125; RV32IFD-NEXT:    sw a1, 12(sp)
126; RV32IFD-NEXT:    fld ft0, 8(sp)
127; RV32IFD-NEXT:    lui a0, 912092
128; RV32IFD-NEXT:    fld ft1, -273(a0)
129; RV32IFD-NEXT:    fadd.d ft0, ft0, ft1
130; RV32IFD-NEXT:    fsd ft0, -273(a0)
131; RV32IFD-NEXT:    fsd ft0, 8(sp)
132; RV32IFD-NEXT:    lw a0, 8(sp)
133; RV32IFD-NEXT:    lw a1, 12(sp)
134; RV32IFD-NEXT:    addi sp, sp, 16
135; RV32IFD-NEXT:    ret
136;
137; RV64IFD-LABEL: fld_fsd_constant:
138; RV64IFD:       # %bb.0:
139; RV64IFD-NEXT:    lui a1, 56
140; RV64IFD-NEXT:    addiw a1, a1, -1353
141; RV64IFD-NEXT:    slli a1, a1, 14
142; RV64IFD-NEXT:    fld ft0, -273(a1)
143; RV64IFD-NEXT:    fmv.d.x ft1, a0
144; RV64IFD-NEXT:    fadd.d ft0, ft1, ft0
145; RV64IFD-NEXT:    fmv.x.d a0, ft0
146; RV64IFD-NEXT:    fsd ft0, -273(a1)
147; RV64IFD-NEXT:    ret
148  %1 = inttoptr i32 3735928559 to double*
149  %2 = load volatile double, double* %1
150  %3 = fadd double %a, %2
151  store double %3, double* %1
152  ret double %3
153}
154
155declare void @notdead(i8*)
156
157define double @fld_stack(double %a) nounwind {
158; RV32IFD-LABEL: fld_stack:
159; RV32IFD:       # %bb.0:
160; RV32IFD-NEXT:    addi sp, sp, -32
161; RV32IFD-NEXT:    sw ra, 28(sp)
162; RV32IFD-NEXT:    sw a0, 8(sp)
163; RV32IFD-NEXT:    sw a1, 12(sp)
164; RV32IFD-NEXT:    fld ft0, 8(sp)
165; RV32IFD-NEXT:    fsd ft0, 0(sp)
166; RV32IFD-NEXT:    addi a0, sp, 16
167; RV32IFD-NEXT:    call notdead
168; RV32IFD-NEXT:    fld ft0, 16(sp)
169; RV32IFD-NEXT:    fld ft1, 0(sp)
170; RV32IFD-NEXT:    fadd.d ft0, ft0, ft1
171; RV32IFD-NEXT:    fsd ft0, 8(sp)
172; RV32IFD-NEXT:    lw a0, 8(sp)
173; RV32IFD-NEXT:    lw a1, 12(sp)
174; RV32IFD-NEXT:    lw ra, 28(sp)
175; RV32IFD-NEXT:    addi sp, sp, 32
176; RV32IFD-NEXT:    ret
177;
178; RV64IFD-LABEL: fld_stack:
179; RV64IFD:       # %bb.0:
180; RV64IFD-NEXT:    addi sp, sp, -32
181; RV64IFD-NEXT:    sd ra, 24(sp)
182; RV64IFD-NEXT:    fmv.d.x ft0, a0
183; RV64IFD-NEXT:    fsd ft0, 8(sp)
184; RV64IFD-NEXT:    addi a0, sp, 16
185; RV64IFD-NEXT:    call notdead
186; RV64IFD-NEXT:    fld ft0, 16(sp)
187; RV64IFD-NEXT:    fld ft1, 8(sp)
188; RV64IFD-NEXT:    fadd.d ft0, ft0, ft1
189; RV64IFD-NEXT:    fmv.x.d a0, ft0
190; RV64IFD-NEXT:    ld ra, 24(sp)
191; RV64IFD-NEXT:    addi sp, sp, 32
192; RV64IFD-NEXT:    ret
193  %1 = alloca double, align 8
194  %2 = bitcast double* %1 to i8*
195  call void @notdead(i8* %2)
196  %3 = load double, double* %1
197  %4 = fadd double %3, %a ; force load in to FPR64
198  ret double %4
199}
200
201define void @fsd_stack(double %a, double %b) nounwind {
202; RV32IFD-LABEL: fsd_stack:
203; RV32IFD:       # %bb.0:
204; RV32IFD-NEXT:    addi sp, sp, -32
205; RV32IFD-NEXT:    sw ra, 28(sp)
206; RV32IFD-NEXT:    sw a2, 8(sp)
207; RV32IFD-NEXT:    sw a3, 12(sp)
208; RV32IFD-NEXT:    fld ft0, 8(sp)
209; RV32IFD-NEXT:    sw a0, 8(sp)
210; RV32IFD-NEXT:    sw a1, 12(sp)
211; RV32IFD-NEXT:    fld ft1, 8(sp)
212; RV32IFD-NEXT:    fadd.d ft0, ft1, ft0
213; RV32IFD-NEXT:    fsd ft0, 16(sp)
214; RV32IFD-NEXT:    addi a0, sp, 16
215; RV32IFD-NEXT:    call notdead
216; RV32IFD-NEXT:    lw ra, 28(sp)
217; RV32IFD-NEXT:    addi sp, sp, 32
218; RV32IFD-NEXT:    ret
219;
220; RV64IFD-LABEL: fsd_stack:
221; RV64IFD:       # %bb.0:
222; RV64IFD-NEXT:    addi sp, sp, -16
223; RV64IFD-NEXT:    sd ra, 8(sp)
224; RV64IFD-NEXT:    fmv.d.x ft0, a1
225; RV64IFD-NEXT:    fmv.d.x ft1, a0
226; RV64IFD-NEXT:    fadd.d ft0, ft1, ft0
227; RV64IFD-NEXT:    fsd ft0, 0(sp)
228; RV64IFD-NEXT:    mv a0, sp
229; RV64IFD-NEXT:    call notdead
230; RV64IFD-NEXT:    ld ra, 8(sp)
231; RV64IFD-NEXT:    addi sp, sp, 16
232; RV64IFD-NEXT:    ret
233  %1 = fadd double %a, %b ; force store from FPR64
234  %2 = alloca double, align 8
235  store double %1, double* %2
236  %3 = bitcast double* %2 to i8*
237  call void @notdead(i8* %3)
238  ret void
239}
240
241; Test selection of store<ST4[%a], trunc to f32>, ..
242define void @fsd_trunc(float* %a, double %b) nounwind noinline optnone {
243; RV32IFD-LABEL: fsd_trunc:
244; RV32IFD:       # %bb.0:
245; RV32IFD-NEXT:    addi sp, sp, -16
246; RV32IFD-NEXT:    sw a1, 8(sp)
247; RV32IFD-NEXT:    sw a2, 12(sp)
248; RV32IFD-NEXT:    fld ft0, 8(sp)
249; RV32IFD-NEXT:    fcvt.s.d ft0, ft0
250; RV32IFD-NEXT:    fsw ft0, 0(a0)
251; RV32IFD-NEXT:    addi sp, sp, 16
252; RV32IFD-NEXT:    ret
253;
254; RV64IFD-LABEL: fsd_trunc:
255; RV64IFD:       # %bb.0:
256; RV64IFD-NEXT:    fmv.d.x ft0, a1
257; RV64IFD-NEXT:    fcvt.s.d ft0, ft0
258; RV64IFD-NEXT:    fsw ft0, 0(a0)
259; RV64IFD-NEXT:    ret
260  %1 = fptrunc double %b to float
261  store float %1, float* %a, align 4
262  ret void
263}
264