1 ; XUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=SI %s
2 ; RUN: llc -march=amdgcn -mcpu=tonga -mattr=+fp64-fp16-denormals,-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VI -check-prefix=VI-DENORM %s
3 ; RUN: llc -march=amdgcn -mcpu=tonga -mattr=-fp64-fp16-denormals,-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VI -check-prefix=VI-FLUSH %s
6 ; Make sure (fmul (fadd x, x), c) -> (fmul x, (fmul 2.0, c)) doesn't
7 ; make add an instruction if the fadd has more than one use.
9 declare half @llvm.fabs.f16(half) #1
10 declare float @llvm.fabs.f32(float) #1
12 ; GCN-LABEL: {{^}}multiple_fadd_use_test_f32:
13 ; SI: v_max_legacy_f32_e64 [[A16:v[0-9]+]],
14 ; SI: v_add_f32_e32 [[A17:v[0-9]+]], [[A16]], [[A16]]
15 ; SI: v_mul_f32_e32 [[A18:v[0-9]+]], [[A17]], [[A17]]
16 ; SI: v_mad_f32 [[A20:v[0-9]+]], -[[A18]], [[A17]], 1.0
17 ; SI: buffer_store_dword [[A20]]
19 ; VI: v_add_f32_e64 v{{[0-9]+}}, s{{[0-9]+}}, -1.0
20 ; VI: v_add_f32_e64 v{{[0-9]+}}, s{{[0-9]+}}, -1.0
21 ; VI: v_cmp_gt_f32_e64 vcc, |v{{[0-9]+}}|, |v{{[0-9]+}}|
22 ; VI: v_cndmask_b32_e32
23 ; VI: v_add_f32_e64 v{{[0-9]+}}, |v{{[0-9]+}}|, |v{{[0-9]+}}|
24 ; VI: v_mul_f32_e32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}
25 ; VI: v_mad_f32 v{{[0-9]+}}, -v{{[0-9]+}}, v{{[0-9]+}}, 1.0
26 define amdgpu_kernel void @multiple_fadd_use_test_f32(float addrspace(1)* %out, float %x, float %y, float %z) #0 {
27 %a11 = fadd fast float %y, -1.0
28 %a12 = call float @llvm.fabs.f32(float %a11)
29 %a13 = fadd fast float %x, -1.0
30 %a14 = call float @llvm.fabs.f32(float %a13)
31 %a15 = fcmp ogt float %a12, %a14
32 %a16 = select i1 %a15, float %a12, float %a14
33 %a17 = fmul fast float %a16, 2.0
34 %a18 = fmul fast float %a17, %a17
35 %a19 = fmul fast float %a18, %a17
36 %a20 = fsub fast float 1.0, %a19
37 store float %a20, float addrspace(1)* %out
41 ; GCN-LABEL: {{^}}multiple_use_fadd_fmac_f32:
42 ; GCN-DAG: v_add_f32_e64 [[MUL2:v[0-9]+]], [[X:s[0-9]+]], s{{[0-9]+}}
43 ; GCN-DAG: v_mac_f32_e64 [[MAD:v[0-9]+]], [[X]], 2.0
44 ; GCN-DAG: buffer_store_dword [[MUL2]]
45 ; GCN-DAG: buffer_store_dword [[MAD]]
47 define amdgpu_kernel void @multiple_use_fadd_fmac_f32(float addrspace(1)* %out, float %x, float %y) #0 {
48 %out.gep.1 = getelementptr float, float addrspace(1)* %out, i32 1
49 %mul2 = fmul fast float %x, 2.0
50 %mad = fadd fast float %mul2, %y
51 store volatile float %mul2, float addrspace(1)* %out
52 store volatile float %mad, float addrspace(1)* %out.gep.1
56 ; GCN-LABEL: {{^}}multiple_use_fadd_fmad_f32:
57 ; GCN-DAG: v_add_f32_e64 [[MUL2:v[0-9]+]], |[[X:s[0-9]+]]|, |s{{[0-9]+}}|
58 ; GCN-DAG: v_mad_f32 [[MAD:v[0-9]+]], |[[X]]|, 2.0, v{{[0-9]+}}
59 ; GCN-DAG: buffer_store_dword [[MUL2]]
60 ; GCN-DAG: buffer_store_dword [[MAD]]
62 define amdgpu_kernel void @multiple_use_fadd_fmad_f32(float addrspace(1)* %out, float %x, float %y) #0 {
63 %out.gep.1 = getelementptr float, float addrspace(1)* %out, i32 1
64 %x.abs = call float @llvm.fabs.f32(float %x)
65 %mul2 = fmul fast float %x.abs, 2.0
66 %mad = fadd fast float %mul2, %y
67 store volatile float %mul2, float addrspace(1)* %out
68 store volatile float %mad, float addrspace(1)* %out.gep.1
72 ; GCN-LABEL: {{^}}multiple_use_fadd_multi_fmad_f32:
73 ; GCN: v_mad_f32 {{v[0-9]+}}, |[[X:s[0-9]+]]|, 2.0, v{{[0-9]+}}
74 ; GCN: v_mad_f32 {{v[0-9]+}}, |[[X]]|, 2.0, v{{[0-9]+}}
75 define amdgpu_kernel void @multiple_use_fadd_multi_fmad_f32(float addrspace(1)* %out, float %x, float %y, float %z) #0 {
76 %out.gep.1 = getelementptr float, float addrspace(1)* %out, i32 1
77 %x.abs = call float @llvm.fabs.f32(float %x)
78 %mul2 = fmul fast float %x.abs, 2.0
79 %mad0 = fadd fast float %mul2, %y
80 %mad1 = fadd fast float %mul2, %z
81 store volatile float %mad0, float addrspace(1)* %out
82 store volatile float %mad1, float addrspace(1)* %out.gep.1
86 ; GCN-LABEL: {{^}}fmul_x2_xn2_f32:
87 ; GCN: v_mul_f32_e64 [[TMP0:v[0-9]+]], [[X:s[0-9]+]], -4.0
88 ; GCN: v_mul_f32_e32 [[RESULT:v[0-9]+]], [[X]], [[TMP0]]
89 ; GCN: buffer_store_dword [[RESULT]]
90 define amdgpu_kernel void @fmul_x2_xn2_f32(float addrspace(1)* %out, float %x, float %y) #0 {
91 %out.gep.1 = getelementptr float, float addrspace(1)* %out, i32 1
92 %mul2 = fmul fast float %x, 2.0
93 %muln2 = fmul fast float %x, -2.0
94 %mul = fmul fast float %mul2, %muln2
95 store volatile float %mul, float addrspace(1)* %out
99 ; GCN-LABEL: {{^}}fmul_x2_xn3_f32:
100 ; GCN: v_mov_b32_e32 [[K:v[0-9]+]], 0xc0c00000
101 ; GCN: v_mul_f32_e32 [[TMP0:v[0-9]+]], [[X:s[0-9]+]], [[K]]
102 ; GCN: v_mul_f32_e32 [[RESULT:v[0-9]+]], [[X]], [[TMP0]]
103 ; GCN: buffer_store_dword [[RESULT]]
104 define amdgpu_kernel void @fmul_x2_xn3_f32(float addrspace(1)* %out, float %x, float %y) #0 {
105 %out.gep.1 = getelementptr float, float addrspace(1)* %out, i32 1
106 %mul2 = fmul fast float %x, 2.0
107 %muln2 = fmul fast float %x, -3.0
108 %mul = fmul fast float %mul2, %muln2
109 store volatile float %mul, float addrspace(1)* %out
113 ; GCN-LABEL: {{^}}multiple_fadd_use_test_f16:
114 ; VI: v_add_f16_e64 v{{[0-9]+}}, s{{[0-9]+}}, -1.0
115 ; VI: v_add_f16_e64 v{{[0-9]+}}, s{{[0-9]+}}, -1.0
116 ; VI: v_cmp_gt_f16_e64 vcc, |v{{[0-9]+}}|, |v{{[0-9]+}}|
117 ; VI: v_cndmask_b32_e32
118 ; VI: v_add_f16_e64 v{{[0-9]+}}, |v{{[0-9]+}}|, |v{{[0-9]+}}|
119 ; VI: v_mul_f16_e32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}
120 ; VI-FLUSH: v_mad_f16 v{{[0-9]+}}, -v{{[0-9]+}}, v{{[0-9]+}}, 1.0
121 ; VI-DENORM: v_fma_f16 v{{[0-9]+}}, -v{{[0-9]+}}, v{{[0-9]+}}, 1.0
122 define amdgpu_kernel void @multiple_fadd_use_test_f16(half addrspace(1)* %out, i16 zeroext %x.arg, i16 zeroext %y.arg, i16 zeroext %z.arg) #0 {
123 %x = bitcast i16 %x.arg to half
124 %y = bitcast i16 %y.arg to half
125 %z = bitcast i16 %z.arg to half
126 %a11 = fadd fast half %y, -1.0
127 %a12 = call half @llvm.fabs.f16(half %a11)
128 %a13 = fadd fast half %x, -1.0
129 %a14 = call half @llvm.fabs.f16(half %a13)
130 %a15 = fcmp ogt half %a12, %a14
131 %a16 = select i1 %a15, half %a12, half %a14
132 %a17 = fmul fast half %a16, 2.0
133 %a18 = fmul fast half %a17, %a17
134 %a19 = fmul fast half %a18, %a17
135 %a20 = fsub fast half 1.0, %a19
136 store half %a20, half addrspace(1)* %out
140 ; GCN-LABEL: {{^}}multiple_use_fadd_fmac_f16:
141 ; GCN-DAG: v_add_f16_e64 [[MUL2:v[0-9]+]], [[X:s[0-9]+]], s{{[0-9]+}}
143 ; VI-FLUSH-DAG: v_mac_f16_e64 [[MAD:v[0-9]+]], [[X]], 2.0
144 ; VI-DENORM-DAG: v_fma_f16 [[MAD:v[0-9]+]], [[X]], 2.0, v{{[0-9]+}}
146 ; GCN-DAG: buffer_store_short [[MUL2]]
147 ; GCN-DAG: buffer_store_short [[MAD]]
149 define amdgpu_kernel void @multiple_use_fadd_fmac_f16(half addrspace(1)* %out, i16 zeroext %x.arg, i16 zeroext %y.arg) #0 {
150 %x = bitcast i16 %x.arg to half
151 %y = bitcast i16 %y.arg to half
152 %out.gep.1 = getelementptr half, half addrspace(1)* %out, i32 1
153 %mul2 = fmul fast half %x, 2.0
154 %mad = fadd fast half %mul2, %y
155 store volatile half %mul2, half addrspace(1)* %out
156 store volatile half %mad, half addrspace(1)* %out.gep.1
160 ; GCN-LABEL: {{^}}multiple_use_fadd_fmad_f16:
161 ; GCN-DAG: v_add_f16_e64 [[MUL2:v[0-9]+]], |[[X:s[0-9]+]]|, |s{{[0-9]+}}|
163 ; VI-FLUSH-DAG: v_mad_f16 [[MAD:v[0-9]+]], |[[X]]|, 2.0, v{{[0-9]+}}
164 ; VI-DENORM-DAG: v_fma_f16 [[MAD:v[0-9]+]], |[[X]]|, 2.0, v{{[0-9]+}}
166 ; GCN-DAG: buffer_store_short [[MUL2]]
167 ; GCN-DAG: buffer_store_short [[MAD]]
169 define amdgpu_kernel void @multiple_use_fadd_fmad_f16(half addrspace(1)* %out, i16 zeroext %x.arg, i16 zeroext %y.arg) #0 {
170 %x = bitcast i16 %x.arg to half
171 %y = bitcast i16 %y.arg to half
172 %out.gep.1 = getelementptr half, half addrspace(1)* %out, i32 1
173 %x.abs = call half @llvm.fabs.f16(half %x)
174 %mul2 = fmul fast half %x.abs, 2.0
175 %mad = fadd fast half %mul2, %y
176 store volatile half %mul2, half addrspace(1)* %out
177 store volatile half %mad, half addrspace(1)* %out.gep.1
181 ; GCN-LABEL: {{^}}multiple_use_fadd_multi_fmad_f16:
182 ; VI-FLUSH: v_mad_f16 {{v[0-9]+}}, |[[X:s[0-9]+]]|, 2.0, v{{[0-9]+}}
183 ; VI-FLUSH: v_mad_f16 {{v[0-9]+}}, |[[X]]|, 2.0, v{{[0-9]+}}
185 ; VI-DENORM: v_fma_f16 {{v[0-9]+}}, |[[X:s[0-9]+]]|, 2.0, v{{[0-9]+}}
186 ; VI-DENORM: v_fma_f16 {{v[0-9]+}}, |[[X]]|, 2.0, v{{[0-9]+}}
188 define amdgpu_kernel void @multiple_use_fadd_multi_fmad_f16(half addrspace(1)* %out, i16 zeroext %x.arg, i16 zeroext %y.arg, i16 zeroext %z.arg) #0 {
189 %x = bitcast i16 %x.arg to half
190 %y = bitcast i16 %y.arg to half
191 %z = bitcast i16 %z.arg to half
192 %out.gep.1 = getelementptr half, half addrspace(1)* %out, i32 1
193 %x.abs = call half @llvm.fabs.f16(half %x)
194 %mul2 = fmul fast half %x.abs, 2.0
195 %mad0 = fadd fast half %mul2, %y
196 %mad1 = fadd fast half %mul2, %z
197 store volatile half %mad0, half addrspace(1)* %out
198 store volatile half %mad1, half addrspace(1)* %out.gep.1
202 ; GCN-LABEL: {{^}}fmul_x2_xn2_f16:
203 ; GCN: v_mul_f16_e64 [[TMP0:v[0-9]+]], [[X:s[0-9]+]], -4.0
204 ; GCN: v_mul_f16_e32 [[RESULT:v[0-9]+]], [[X]], [[TMP0]]
205 ; GCN: buffer_store_short [[RESULT]]
206 define amdgpu_kernel void @fmul_x2_xn2_f16(half addrspace(1)* %out, i16 zeroext %x.arg, i16 zeroext %y.arg) #0 {
207 %x = bitcast i16 %x.arg to half
208 %y = bitcast i16 %y.arg to half
209 %out.gep.1 = getelementptr half, half addrspace(1)* %out, i32 1
210 %mul2 = fmul fast half %x, 2.0
211 %muln2 = fmul fast half %x, -2.0
212 %mul = fmul fast half %mul2, %muln2
213 store volatile half %mul, half addrspace(1)* %out
217 ; GCN-LABEL: {{^}}fmul_x2_xn3_f16:
218 ; GCN: v_mov_b32_e32 [[K:v[0-9]+]], 0xc600
219 ; GCN: v_mul_f16_e32 [[TMP0:v[0-9]+]], [[X:s[0-9]+]], [[K]]
220 ; GCN: v_mul_f16_e32 [[RESULT:v[0-9]+]], [[X]], [[TMP0]]
221 ; GCN: buffer_store_short [[RESULT]]
222 define amdgpu_kernel void @fmul_x2_xn3_f16(half addrspace(1)* %out, i16 zeroext %x.arg, i16 zeroext %y.arg) #0 {
223 %x = bitcast i16 %x.arg to half
224 %y = bitcast i16 %y.arg to half
225 %out.gep.1 = getelementptr half, half addrspace(1)* %out, i32 1
226 %mul2 = fmul fast half %x, 2.0
227 %muln2 = fmul fast half %x, -3.0
228 %mul = fmul fast half %mul2, %muln2
229 store volatile half %mul, half addrspace(1)* %out
233 attributes #0 = { nounwind "unsafe-fp-math"="true" }
234 attributes #1 = { nounwind readnone }