xref: /llvm-project/llvm/test/CodeGen/AMDGPU/operand-folding.ll (revision 9e9907f1cfa424366fba58d9520f9305b537cec9)
1; RUN: llc -mtriple=amdgcn -mcpu=tahiti -verify-machineinstrs < %s | FileCheck %s
2; RUN: llc -mtriple=amdgcn -mcpu=tahiti -verify-machineinstrs -early-live-intervals < %s | FileCheck %s
3
4; CHECK-LABEL: {{^}}fold_sgpr:
5; CHECK: v_add_i32_e32 v{{[0-9]+}}, vcc, s
6define amdgpu_kernel void @fold_sgpr(ptr addrspace(1) %out, i32 %fold) #1 {
7entry:
8  %tmp0 = icmp ne i32 %fold, 0
9  br i1 %tmp0, label %if, label %endif
10
11if:
12  %id = call i32 @llvm.amdgcn.workitem.id.x()
13  %offset = add i32 %fold, %id
14  %tmp1 = getelementptr i32, ptr addrspace(1) %out, i32 %offset
15  store i32 0, ptr addrspace(1) %tmp1
16  br label %endif
17
18endif:
19  ret void
20}
21
22; CHECK-LABEL: {{^}}fold_imm:
23; CHECK: v_or_b32_e32 v{{[0-9]+}}, 5
24define amdgpu_kernel void @fold_imm(ptr addrspace(1) %out, i32 %cmp) #1 {
25entry:
26  %fold = add i32 3, 2
27  %tmp0 = icmp ne i32 %cmp, 0
28  br i1 %tmp0, label %if, label %endif
29
30if:
31  %id = call i32 @llvm.amdgcn.workitem.id.x()
32  %val = or i32 %id, %fold
33  store i32 %val, ptr addrspace(1) %out
34  br label %endif
35
36endif:
37  ret void
38}
39
40; CHECK-LABEL: {{^}}fold_64bit_constant_add:
41; CHECK-NOT: s_mov_b64
42; FIXME: It would be better if we could use v_add here and drop the extra
43; v_mov_b32 instructions.
44; CHECK-DAG: s_add_u32 [[LO:s[0-9]+]], s{{[0-9]+}}, 1
45; CHECK-DAG: s_addc_u32 [[HI:s[0-9]+]], s{{[0-9]+}}, 0
46; CHECK-DAG: v_mov_b32_e32 v[[VLO:[0-9]+]], [[LO]]
47; CHECK-DAG: v_mov_b32_e32 v[[VHI:[0-9]+]], [[HI]]
48; CHECK: buffer_store_dwordx2 v[[[VLO]]:[[VHI]]],
49
50define amdgpu_kernel void @fold_64bit_constant_add(ptr addrspace(1) %out, i32 %cmp, i64 %val) #1 {
51entry:
52  %tmp0 = add i64 %val, 1
53  store i64 %tmp0, ptr addrspace(1) %out
54  ret void
55}
56
57; Inline constants should always be folded.
58
59; CHECK-LABEL: {{^}}vector_inline:
60; CHECK: v_xor_b32_e32 v{{[0-9]+}}, 5, v{{[0-9]+}}
61; CHECK: v_xor_b32_e32 v{{[0-9]+}}, 5, v{{[0-9]+}}
62; CHECK: v_xor_b32_e32 v{{[0-9]+}}, 5, v{{[0-9]+}}
63; CHECK: v_xor_b32_e32 v{{[0-9]+}}, 5, v{{[0-9]+}}
64
65define amdgpu_kernel void @vector_inline(ptr addrspace(1) %out) #1 {
66entry:
67  %tmp0 = call i32 @llvm.amdgcn.workitem.id.x()
68  %tmp1 = add i32 %tmp0, 1
69  %tmp2 = add i32 %tmp0, 2
70  %tmp3 = add i32 %tmp0, 3
71  %vec0 = insertelement <4 x i32> undef, i32 %tmp0, i32 0
72  %vec1 = insertelement <4 x i32> %vec0, i32 %tmp1, i32 1
73  %vec2 = insertelement <4 x i32> %vec1, i32 %tmp2, i32 2
74  %vec3 = insertelement <4 x i32> %vec2, i32 %tmp3, i32 3
75  %tmp4 = xor <4 x i32> <i32 5, i32 5, i32 5, i32 5>, %vec3
76  store <4 x i32> %tmp4, ptr addrspace(1) %out
77  ret void
78}
79
80; Immediates with one use should be folded
81; CHECK-LABEL: {{^}}imm_one_use:
82; CHECK: v_xor_b32_e32 v{{[0-9]+}}, 0x64, v{{[0-9]+}}
83
84define amdgpu_kernel void @imm_one_use(ptr addrspace(1) %out) #1 {
85entry:
86  %tmp0 = call i32 @llvm.amdgcn.workitem.id.x()
87  %tmp1 = xor i32 %tmp0, 100
88  store i32 %tmp1, ptr addrspace(1) %out
89  ret void
90}
91; CHECK-LABEL: {{^}}vector_imm:
92; CHECK: v_xor_b32_e32 v{{[0-9]}}, 0x64, v{{[0-9]}}
93; CHECK: v_xor_b32_e32 v{{[0-9]}}, 0x64, v{{[0-9]}}
94; CHECK: v_xor_b32_e32 v{{[0-9]}}, 0x64, v{{[0-9]}}
95; CHECK: v_xor_b32_e32 v{{[0-9]}}, 0x64, v{{[0-9]}}
96
97define amdgpu_kernel void @vector_imm(ptr addrspace(1) %out) #1 {
98entry:
99  %tmp0 = call i32 @llvm.amdgcn.workitem.id.x()
100  %tmp1 = add i32 %tmp0, 1
101  %tmp2 = add i32 %tmp0, 2
102  %tmp3 = add i32 %tmp0, 3
103  %vec0 = insertelement <4 x i32> undef, i32 %tmp0, i32 0
104  %vec1 = insertelement <4 x i32> %vec0, i32 %tmp1, i32 1
105  %vec2 = insertelement <4 x i32> %vec1, i32 %tmp2, i32 2
106  %vec3 = insertelement <4 x i32> %vec2, i32 %tmp3, i32 3
107  %tmp4 = xor <4 x i32> <i32 100, i32 100, i32 100, i32 100>, %vec3
108  store <4 x i32> %tmp4, ptr addrspace(1) %out
109  ret void
110}
111
112; A subregister use operand should not be tied.
113; CHECK-LABEL: {{^}}no_fold_tied_subregister:
114; CHECK: buffer_load_dwordx2 v[[[LO:[0-9]+]]:[[HI:[0-9]+]]]
115; CHECK: v_madmk_f32 v[[RES:[0-9]+]], v[[HI]], 0x41200000, v[[LO]]
116; CHECK: buffer_store_dword v[[RES]]
117define amdgpu_kernel void @no_fold_tied_subregister() #1 {
118  %tmp1 = load volatile <2 x float>, ptr addrspace(1) undef
119  %tmp2 = extractelement <2 x float> %tmp1, i32 0
120  %tmp3 = extractelement <2 x float> %tmp1, i32 1
121  %tmp4 = fmul float %tmp3, 10.0
122  %tmp5 = fadd float %tmp4, %tmp2
123  store volatile float %tmp5, ptr addrspace(1) undef
124  ret void
125}
126
127; There should be exact one folding on the same operand.
128; CHECK-LABEL: {{^}}no_extra_fold_on_same_opnd
129; CHECK-NOT: %bb.1:
130; CHECK: v_xor_b32_e32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}
131define void @no_extra_fold_on_same_opnd() #1 {
132entry:
133  %s0 = load i32, ptr addrspace(5) undef, align 4
134  %s0.i64= zext i32 %s0 to i64
135  br label %for.body.i.i
136
137for.body.i.i:
138  %s1 = load i32, ptr addrspace(1) undef, align 8
139  %s1.i64 = sext i32 %s1 to i64
140  %xor = xor i64 %s1.i64, %s0.i64
141  %flag = icmp ult i64 %xor, 8
142  br i1 %flag, label %if.then, label %if.else
143
144if.then:
145  unreachable
146
147if.else:
148  unreachable
149}
150
151declare i32 @llvm.amdgcn.workitem.id.x() #0
152
153attributes #0 = { nounwind readnone }
154attributes #1 = { nounwind "denormal-fp-math-f32"="preserve-sign,preserve-sign" }
155