xref: /llvm-project/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.sdot8.ll (revision 9e9907f1cfa424366fba58d9520f9305b537cec9)
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc -global-isel -mtriple=amdgcn -mcpu=gfx906 -verify-machineinstrs < %s | FileCheck --check-prefix=GFX906 %s
3; RUN: llc -global-isel -mtriple=amdgcn -mcpu=gfx1011 -verify-machineinstrs < %s | FileCheck --check-prefix=GFX10 %s
4; RUN: llc -global-isel -mtriple=amdgcn -mcpu=gfx1012 -verify-machineinstrs < %s | FileCheck --check-prefix=GFX10 %s
5
6define i32 @v_sdot8(i32 %a, i32 %b, i32 %c) {
7; GFX906-LABEL: v_sdot8:
8; GFX906:       ; %bb.0:
9; GFX906-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
10; GFX906-NEXT:    v_dot8_i32_i4 v0, v0, v1, v2
11; GFX906-NEXT:    s_setpc_b64 s[30:31]
12;
13; GFX10-LABEL: v_sdot8:
14; GFX10:       ; %bb.0:
15; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
16; GFX10-NEXT:    v_dot8_i32_i4 v0, v0, v1, v2
17; GFX10-NEXT:    s_setpc_b64 s[30:31]
18  %r = call i32 @llvm.amdgcn.sdot8(i32 %a, i32 %b, i32 %c, i1 false)
19  ret i32 %r
20}
21
22define i32 @v_sdot8_clamp(i32 %a, i32 %b, i32 %c) {
23; GFX906-LABEL: v_sdot8_clamp:
24; GFX906:       ; %bb.0:
25; GFX906-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
26; GFX906-NEXT:    v_dot8_i32_i4 v0, v0, v1, v2 clamp
27; GFX906-NEXT:    s_setpc_b64 s[30:31]
28;
29; GFX10-LABEL: v_sdot8_clamp:
30; GFX10:       ; %bb.0:
31; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
32; GFX10-NEXT:    v_dot8_i32_i4 v0, v0, v1, v2 clamp
33; GFX10-NEXT:    s_setpc_b64 s[30:31]
34  %r = call i32 @llvm.amdgcn.sdot8(i32 %a, i32 %b, i32 %c, i1 true)
35  ret i32 %r
36}
37
38; FIXME: Fix argument do not let these casts expand
39; define i32 @v_sdot8_cast_v8i4(<8 x i4> %a, <8 x i4> %b, i32 %c) {
40;   %a.cast = bitcast <8 x i4> %a to i32
41;   %b.cast = bitcast <8 x i4> %b to i32
42;   %r = call i32 @llvm.amdgcn.sdot8(i32 %a.cast, i32 %b.cast, i32 %c, i1 false)
43;   ret i32 %r
44; }
45
46define i32 @v_sdot8_fnegf32_a(float %a, i32 %b, i32 %c) {
47; GFX906-LABEL: v_sdot8_fnegf32_a:
48; GFX906:       ; %bb.0:
49; GFX906-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
50; GFX906-NEXT:    v_xor_b32_e32 v0, 0x80000000, v0
51; GFX906-NEXT:    v_dot8_i32_i4 v0, v0, v1, v2
52; GFX906-NEXT:    s_setpc_b64 s[30:31]
53;
54; GFX10-LABEL: v_sdot8_fnegf32_a:
55; GFX10:       ; %bb.0:
56; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
57; GFX10-NEXT:    v_xor_b32_e32 v0, 0x80000000, v0
58; GFX10-NEXT:    v_dot8_i32_i4 v0, v0, v1, v2
59; GFX10-NEXT:    s_setpc_b64 s[30:31]
60  %neg.a = fneg float %a
61  %cast.neg.a = bitcast float %neg.a to i32
62  %r = call i32 @llvm.amdgcn.sdot8(i32 %cast.neg.a, i32 %b, i32 %c, i1 false)
63  ret i32 %r
64}
65
66define i32 @v_sdot8_fnegv2f16_a(<2 x half> %a, i32 %b, i32 %c) {
67; GFX906-LABEL: v_sdot8_fnegv2f16_a:
68; GFX906:       ; %bb.0:
69; GFX906-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
70; GFX906-NEXT:    v_xor_b32_e32 v0, 0x80008000, v0
71; GFX906-NEXT:    v_dot8_i32_i4 v0, v0, v1, v2
72; GFX906-NEXT:    s_setpc_b64 s[30:31]
73;
74; GFX10-LABEL: v_sdot8_fnegv2f16_a:
75; GFX10:       ; %bb.0:
76; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
77; GFX10-NEXT:    v_xor_b32_e32 v0, 0x80008000, v0
78; GFX10-NEXT:    v_dot8_i32_i4 v0, v0, v1, v2
79; GFX10-NEXT:    s_setpc_b64 s[30:31]
80  %neg.a = fneg <2 x half> %a
81  %cast.neg.a = bitcast <2 x half> %neg.a to i32
82  %r = call i32 @llvm.amdgcn.sdot8(i32 %cast.neg.a, i32 %b, i32 %c, i1 false)
83  ret i32 %r
84}
85
86declare i32 @llvm.amdgcn.sdot8(i32, i32, i32, i1 immarg) #0
87
88attributes #0 = { nounwind readnone speculatable }
89