• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
2
3; Make sure (fmul (fadd x, x), c) -> (fmul x, (fmul 2.0, c)) doesn't
4; make add an instruction if the fadd has more than one use.
5
6declare float @llvm.fabs.f32(float) #1
7
8; GCN-LABEL: {{^}}multiple_fadd_use_test:
9; GCN: v_max_legacy_f32_e64 [[A16:v[0-9]+]],
10; GCN: v_add_f32_e32 [[A17:v[0-9]+]], [[A16]], [[A16]]
11; GCN: v_mul_f32_e32 [[A18:v[0-9]+]], [[A17]], [[A17]]
12; GCN: v_mad_f32 [[A20:v[0-9]+]], -[[A18]], [[A17]], 1.0
13; GCN: buffer_store_dword [[A20]]
14define void @multiple_fadd_use_test(float addrspace(1)* %out, float %x, float %y, float %z) #0 {
15  %a11 = fadd fast float %y, -1.0
16  %a12 = call float @llvm.fabs.f32(float %a11)
17  %a13 = fadd fast float %x, -1.0
18  %a14 = call float @llvm.fabs.f32(float %a13)
19  %a15 = fcmp ogt float %a12, %a14
20  %a16 = select i1 %a15, float %a12, float %a14
21  %a17 = fmul fast float %a16, 2.0
22  %a18 = fmul fast float %a17, %a17
23  %a19 = fmul fast float %a18, %a17
24  %a20 = fsub fast float 1.0, %a19
25  store float %a20, float addrspace(1)* %out
26  ret void
27}
28
29; GCN-LABEL: {{^}}multiple_use_fadd_fmac
30; GCN-DAG: v_add_f32_e64 [[MUL2:v[0-9]+]], [[X:s[0-9]+]], s{{[0-9]+}}
31; GCN-DAG: v_mac_f32_e64 [[MAD:v[0-9]+]], 2.0, [[X]]
32; GCN-DAG: buffer_store_dword [[MUL2]]
33; GCN-DAG: buffer_store_dword [[MAD]]
34; GCN: s_endpgm
35define void @multiple_use_fadd_fmac(float addrspace(1)* %out, float %x, float %y) #0 {
36  %out.gep.1 = getelementptr float, float addrspace(1)* %out, i32 1
37  %mul2 = fmul fast float %x, 2.0
38  %mad = fadd fast float %mul2, %y
39  store volatile float %mul2, float addrspace(1)* %out
40  store volatile float %mad, float addrspace(1)* %out.gep.1
41  ret void
42}
43
44; GCN-LABEL: {{^}}multiple_use_fadd_fmad:
45; GCN-DAG: v_add_f32_e64 [[MUL2:v[0-9]+]], |[[X:s[0-9]+]]|, |s{{[0-9]+}}|
46; GCN-DAG: v_mad_f32 [[MAD:v[0-9]+]], 2.0, |[[X]]|, v{{[0-9]+}}
47; GCN-DAG: buffer_store_dword [[MUL2]]
48; GCN-DAG: buffer_store_dword [[MAD]]
49; GCN: s_endpgm
50define void @multiple_use_fadd_fmad(float addrspace(1)* %out, float %x, float %y) #0 {
51  %out.gep.1 = getelementptr float, float addrspace(1)* %out, i32 1
52  %x.abs = call float @llvm.fabs.f32(float %x)
53  %mul2 = fmul fast float %x.abs, 2.0
54  %mad = fadd fast float %mul2, %y
55  store volatile float %mul2, float addrspace(1)* %out
56  store volatile float %mad, float addrspace(1)* %out.gep.1
57  ret void
58}
59
60; GCN-LABEL: {{^}}multiple_use_fadd_multi_fmad:
61; GCN: v_mad_f32 {{v[0-9]+}}, 2.0, |[[X:s[0-9]+]]|, v{{[0-9]+}}
62; GCN: v_mad_f32 {{v[0-9]+}}, 2.0, |[[X]]|, v{{[0-9]+}}
63define void @multiple_use_fadd_multi_fmad(float addrspace(1)* %out, float %x, float %y, float %z) #0 {
64  %out.gep.1 = getelementptr float, float addrspace(1)* %out, i32 1
65  %x.abs = call float @llvm.fabs.f32(float %x)
66  %mul2 = fmul fast float %x.abs, 2.0
67  %mad0 = fadd fast float %mul2, %y
68  %mad1 = fadd fast float %mul2, %z
69  store volatile float %mad0, float addrspace(1)* %out
70  store volatile float %mad1, float addrspace(1)* %out.gep.1
71  ret void
72}
73
74; GCN-LABEL: {{^}}fmul_x2_xn2:
75; GCN: v_mul_f32_e64 [[TMP0:v[0-9]+]], -4.0, [[X:s[0-9]+]]
76; GCN: v_mul_f32_e32 [[RESULT:v[0-9]+]], [[X]], [[TMP0]]
77; GCN: buffer_store_dword [[RESULT]]
78define void @fmul_x2_xn2(float addrspace(1)* %out, float %x, float %y) #0 {
79  %out.gep.1 = getelementptr float, float addrspace(1)* %out, i32 1
80  %mul2 = fmul fast float %x, 2.0
81  %muln2 = fmul fast float %x, -2.0
82  %mul = fmul fast float %mul2, %muln2
83  store volatile float %mul, float addrspace(1)* %out
84  ret void
85}
86
87; GCN-LABEL: {{^}}fmul_x2_xn3:
88; GCN: v_mov_b32_e32 [[K:v[0-9]+]], 0xc0c00000
89; GCN: v_mul_f32_e32 [[TMP0:v[0-9]+]], [[X:s[0-9]+]], [[K]]
90; GCN: v_mul_f32_e32 [[RESULT:v[0-9]+]], [[X]], [[TMP0]]
91; GCN: buffer_store_dword [[RESULT]]
92define void @fmul_x2_xn3(float addrspace(1)* %out, float %x, float %y) #0 {
93  %out.gep.1 = getelementptr float, float addrspace(1)* %out, i32 1
94  %mul2 = fmul fast float %x, 2.0
95  %muln2 = fmul fast float %x, -3.0
96  %mul = fmul fast float %mul2, %muln2
97  store volatile float %mul, float addrspace(1)* %out
98  ret void
99}
100
101attributes #0 = { nounwind "unsafe-fp-math"="true" }
102attributes #1 = { nounwind readnone }
103