1; RUN: llc -march=amdgcn -mcpu=SI -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=SI %s
2; XUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VI %s
3
4; FIXME: Enable for VI.
5
6declare i32 @llvm.r600.read.tidig.x() nounwind readnone
7declare void @llvm.AMDGPU.barrier.global() nounwind noduplicate
8declare float @llvm.AMDGPU.div.fmas.f32(float, float, float, i1) nounwind readnone
9declare double @llvm.AMDGPU.div.fmas.f64(double, double, double, i1) nounwind readnone
10
11; GCN-LABEL: {{^}}test_div_fmas_f32:
12; SI-DAG: s_load_dword [[SA:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xb
13; SI-DAG: s_load_dword [[SC:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xd
14; SI-DAG: s_load_dword [[SB:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xc
15; VI-DAG: s_load_dword [[SA:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0x2c
16; VI-DAG: s_load_dword [[SC:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0x34
17; VI-DAG: s_load_dword [[SB:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0x30
18; GCN-DAG: v_mov_b32_e32 [[VC:v[0-9]+]], [[SC]]
19; GCN-DAG: v_mov_b32_e32 [[VB:v[0-9]+]], [[SB]]
20; GCN-DAG: v_mov_b32_e32 [[VA:v[0-9]+]], [[SA]]
21; GCN: v_div_fmas_f32 [[RESULT:v[0-9]+]], [[VB]], [[VA]], [[VC]]
22; GCN: buffer_store_dword [[RESULT]],
23; GCN: s_endpgm
24define void @test_div_fmas_f32(float addrspace(1)* %out, float %a, float %b, float %c, i1 %d) nounwind {
25  %result = call float @llvm.AMDGPU.div.fmas.f32(float %a, float %b, float %c, i1 %d) nounwind readnone
26  store float %result, float addrspace(1)* %out, align 4
27  ret void
28}
29
30; GCN-LABEL: {{^}}test_div_fmas_f32_inline_imm_0:
31; SI-DAG: s_load_dword [[SC:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xd
32; SI-DAG: s_load_dword [[SB:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xc
33; SI-DAG: v_mov_b32_e32 [[VC:v[0-9]+]], [[SC]]
34; SI-DAG: v_mov_b32_e32 [[VB:v[0-9]+]], [[SB]]
35; SI: v_div_fmas_f32 [[RESULT:v[0-9]+]], 1.0, [[VB]], [[VC]]
36; SI: buffer_store_dword [[RESULT]],
37; SI: s_endpgm
38define void @test_div_fmas_f32_inline_imm_0(float addrspace(1)* %out, float %a, float %b, float %c, i1 %d) nounwind {
39  %result = call float @llvm.AMDGPU.div.fmas.f32(float 1.0, float %b, float %c, i1 %d) nounwind readnone
40  store float %result, float addrspace(1)* %out, align 4
41  ret void
42}
43
44; GCN-LABEL: {{^}}test_div_fmas_f32_inline_imm_1:
45; SI-DAG: s_load_dword [[SA:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xb
46; SI-DAG: s_load_dword [[SC:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xd
47; SI-DAG: v_mov_b32_e32 [[VC:v[0-9]+]], [[SC]]
48; SI-DAG: v_mov_b32_e32 [[VA:v[0-9]+]], [[SA]]
49; SI: v_div_fmas_f32 [[RESULT:v[0-9]+]], 1.0, [[VA]], [[VC]]
50; SI: buffer_store_dword [[RESULT]],
51; SI: s_endpgm
52define void @test_div_fmas_f32_inline_imm_1(float addrspace(1)* %out, float %a, float %b, float %c, i1 %d) nounwind {
53  %result = call float @llvm.AMDGPU.div.fmas.f32(float %a, float 1.0, float %c, i1 %d) nounwind readnone
54  store float %result, float addrspace(1)* %out, align 4
55  ret void
56}
57
58; GCN-LABEL: {{^}}test_div_fmas_f32_inline_imm_2:
59; SI-DAG: s_load_dword [[SA:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xb
60; SI-DAG: s_load_dword [[SB:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xc
61; SI-DAG: v_mov_b32_e32 [[VA:v[0-9]+]], [[SA]]
62; SI-DAG: v_mov_b32_e32 [[VB:v[0-9]+]], [[SB]]
63; SI: v_div_fmas_f32 [[RESULT:v[0-9]+]], [[VA]], [[VB]], 1.0
64; SI: buffer_store_dword [[RESULT]],
65; SI: s_endpgm
66define void @test_div_fmas_f32_inline_imm_2(float addrspace(1)* %out, float %a, float %b, float %c, i1 %d) nounwind {
67  %result = call float @llvm.AMDGPU.div.fmas.f32(float %a, float %b, float 1.0, i1 %d) nounwind readnone
68  store float %result, float addrspace(1)* %out, align 4
69  ret void
70}
71
72; GCN-LABEL: {{^}}test_div_fmas_f64:
73; GCN: v_div_fmas_f64
74define void @test_div_fmas_f64(double addrspace(1)* %out, double %a, double %b, double %c, i1 %d) nounwind {
75  %result = call double @llvm.AMDGPU.div.fmas.f64(double %a, double %b, double %c, i1 %d) nounwind readnone
76  store double %result, double addrspace(1)* %out, align 8
77  ret void
78}
79
80; GCN-LABEL: {{^}}test_div_fmas_f32_cond_to_vcc:
81; SI: v_cmp_eq_i32_e64 vcc, 0, s{{[0-9]+}}
82; SI: v_div_fmas_f32 {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}
83define void @test_div_fmas_f32_cond_to_vcc(float addrspace(1)* %out, float %a, float %b, float %c, i32 %i) nounwind {
84  %cmp = icmp eq i32 %i, 0
85  %result = call float @llvm.AMDGPU.div.fmas.f32(float %a, float %b, float %c, i1 %cmp) nounwind readnone
86  store float %result, float addrspace(1)* %out, align 4
87  ret void
88}
89
90; GCN-LABEL: {{^}}test_div_fmas_f32_imm_false_cond_to_vcc:
91; SI: s_mov_b64 vcc, 0
92; SI: v_div_fmas_f32 {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}
93define void @test_div_fmas_f32_imm_false_cond_to_vcc(float addrspace(1)* %out, float %a, float %b, float %c) nounwind {
94  %result = call float @llvm.AMDGPU.div.fmas.f32(float %a, float %b, float %c, i1 false) nounwind readnone
95  store float %result, float addrspace(1)* %out, align 4
96  ret void
97}
98
99; GCN-LABEL: {{^}}test_div_fmas_f32_imm_true_cond_to_vcc:
100; SI: s_mov_b64 vcc, -1
101; SI: v_div_fmas_f32 {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}
102define void @test_div_fmas_f32_imm_true_cond_to_vcc(float addrspace(1)* %out, float %a, float %b, float %c) nounwind {
103  %result = call float @llvm.AMDGPU.div.fmas.f32(float %a, float %b, float %c, i1 true) nounwind readnone
104  store float %result, float addrspace(1)* %out, align 4
105  ret void
106}
107
108; GCN-LABEL: {{^}}test_div_fmas_f32_logical_cond_to_vcc:
109; SI-DAG: buffer_load_dword [[A:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
110; SI-DAG: buffer_load_dword [[B:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:4{{$}}
111; SI-DAG: buffer_load_dword [[C:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
112
113; SI-DAG: v_cmp_eq_i32_e32 [[CMP0:vcc]], 0, v{{[0-9]+}}
114; SI-DAG: v_cmp_ne_i32_e64 [[CMP1:s\[[0-9]+:[0-9]+\]]], 0, s{{[0-9]+}}
115; SI: s_and_b64 vcc, [[CMP0]], [[CMP1]]
116; SI: v_div_fmas_f32 {{v[0-9]+}}, [[A]], [[B]], [[C]]
117; SI: s_endpgm
118define void @test_div_fmas_f32_logical_cond_to_vcc(float addrspace(1)* %out, float addrspace(1)* %in, i32 %d) nounwind {
119  %tid = call i32 @llvm.r600.read.tidig.x() nounwind readnone
120  %gep.a = getelementptr float, float addrspace(1)* %in, i32 %tid
121  %gep.b = getelementptr float, float addrspace(1)* %gep.a, i32 1
122  %gep.c = getelementptr float, float addrspace(1)* %gep.a, i32 2
123  %gep.out = getelementptr float, float addrspace(1)* %out, i32 2
124
125  %a = load float, float addrspace(1)* %gep.a
126  %b = load float, float addrspace(1)* %gep.b
127  %c = load float, float addrspace(1)* %gep.c
128
129  %cmp0 = icmp eq i32 %tid, 0
130  %cmp1 = icmp ne i32 %d, 0
131  %and = and i1 %cmp0, %cmp1
132
133  %result = call float @llvm.AMDGPU.div.fmas.f32(float %a, float %b, float %c, i1 %and) nounwind readnone
134  store float %result, float addrspace(1)* %gep.out, align 4
135  ret void
136}
137
138; GCN-LABEL: {{^}}test_div_fmas_f32_i1_phi_vcc:
139; SI: v_cmp_eq_i32_e32 vcc, 0, v{{[0-9]+}}
140; SI: s_and_saveexec_b64 [[SAVE:s\[[0-9]+:[0-9]+\]]], vcc
141; SI: s_xor_b64 [[SAVE]], exec, [[SAVE]]
142
143; SI: buffer_load_dword [[LOAD:v[0-9]+]]
144; SI: v_cmp_ne_i32_e32 vcc, 0, [[LOAD]]
145; SI: v_cndmask_b32_e64 {{v[0-9]+}}, 0, -1, vcc
146
147
148; SI: BB9_2:
149; SI: s_or_b64 exec, exec, [[SAVE]]
150; SI: v_cmp_ne_i32_e32 vcc, 0, v0
151; SI: v_div_fmas_f32 {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}
152; SI: buffer_store_dword
153; SI: s_endpgm
154define void @test_div_fmas_f32_i1_phi_vcc(float addrspace(1)* %out, float addrspace(1)* %in, i32 addrspace(1)* %dummy) nounwind {
155entry:
156  %tid = call i32 @llvm.r600.read.tidig.x() nounwind readnone
157  %gep.out = getelementptr float, float addrspace(1)* %out, i32 2
158  %gep.a = getelementptr float, float addrspace(1)* %in, i32 %tid
159  %gep.b = getelementptr float, float addrspace(1)* %gep.a, i32 1
160  %gep.c = getelementptr float, float addrspace(1)* %gep.a, i32 2
161
162  %a = load float, float addrspace(1)* %gep.a
163  %b = load float, float addrspace(1)* %gep.b
164  %c = load float, float addrspace(1)* %gep.c
165
166  %cmp0 = icmp eq i32 %tid, 0
167  br i1 %cmp0, label %bb, label %exit
168
169bb:
170  %val = load i32, i32 addrspace(1)* %dummy
171  %cmp1 = icmp ne i32 %val, 0
172  br label %exit
173
174exit:
175  %cond = phi i1 [false, %entry], [%cmp1, %bb]
176  %result = call float @llvm.AMDGPU.div.fmas.f32(float %a, float %b, float %c, i1 %cond) nounwind readnone
177  store float %result, float addrspace(1)* %gep.out, align 4
178  ret void
179}
180