1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc < %s -O0 -mtriple=x86_64-unknown-unknown -mattr=+amx-int8 -mattr=+avx512f | FileCheck %s --check-prefix=AVX512
3
4@buf = dso_local global [1024 x i8] zeroinitializer, align 16
5@buf2 = dso_local global [1024 x i8] zeroinitializer, align 16
6
7define dso_local void @test_api(i32 %cond, i16 signext %row, i16 signext %col) nounwind {
8; AVX512-LABEL: test_api:
9; AVX512:       # %bb.0: # %entry
10; AVX512-NEXT:    pushq %rbp
11; AVX512-NEXT:    movq %rsp, %rbp
12; AVX512-NEXT:    andq $-1024, %rsp # imm = 0xFC00
13; AVX512-NEXT:    subq $6144, %rsp # imm = 0x1800
14; AVX512-NEXT:    movw %dx, %ax
15; AVX512-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
16; AVX512-NEXT:    movw %si, %ax
17; AVX512-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
18; AVX512-NEXT:    leaq {{[0-9]+}}(%rsp), %rax
19; AVX512-NEXT:    movq %rax, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
20; AVX512-NEXT:    leaq {{[0-9]+}}(%rsp), %rax
21; AVX512-NEXT:    movq %rax, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
22; AVX512-NEXT:    leaq {{[0-9]+}}(%rsp), %rax
23; AVX512-NEXT:    movq %rax, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
24; AVX512-NEXT:    leaq {{[0-9]+}}(%rsp), %rax
25; AVX512-NEXT:    movq %rax, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
26; AVX512-NEXT:    cmpl $0, %edi
27; AVX512-NEXT:    je .LBB0_2
28; AVX512-NEXT:  # %bb.1: # %if.then
29; AVX512-NEXT:    movw {{[-0-9]+}}(%r{{[sb]}}p), %ax # 2-byte Reload
30; AVX512-NEXT:    movw {{[-0-9]+}}(%r{{[sb]}}p), %cx # 2-byte Reload
31; AVX512-NEXT:    movq {{[-0-9]+}}(%r{{[sb]}}p), %rdx # 8-byte Reload
32; AVX512-NEXT:    movq {{[-0-9]+}}(%r{{[sb]}}p), %rdi # 8-byte Reload
33; AVX512-NEXT:    movq {{[-0-9]+}}(%r{{[sb]}}p), %r11 # 8-byte Reload
34; AVX512-NEXT:    vxorps %xmm0, %xmm0, %xmm0
35; AVX512-NEXT:    vmovdqu64 %zmm0, {{[0-9]+}}(%rsp)
36; AVX512-NEXT:    movb $1, {{[0-9]+}}(%rsp)
37; AVX512-NEXT:    movb %al, %sil
38; AVX512-NEXT:    movb %sil, {{[0-9]+}}(%rsp)
39; AVX512-NEXT:    movw $8, {{[0-9]+}}(%rsp)
40; AVX512-NEXT:    ldtilecfg {{[0-9]+}}(%rsp)
41; AVX512-NEXT:    movl $buf, %r9d
42; AVX512-NEXT:    movl $32, %r10d
43; AVX512-NEXT:    movw $8, %si
44; AVX512-NEXT:    tileloadd (%r9,%r10), %tmm0
45; AVX512-NEXT:    movl $64, %r8d
46; AVX512-NEXT:    tilestored %tmm0, (%r11,%r8)
47; AVX512-NEXT:    vmovdqu64 %zmm0, {{[0-9]+}}(%rsp)
48; AVX512-NEXT:    movb $1, {{[0-9]+}}(%rsp)
49; AVX512-NEXT:    movb $8, {{[0-9]+}}(%rsp)
50; AVX512-NEXT:    movw %cx, {{[0-9]+}}(%rsp)
51; AVX512-NEXT:    ldtilecfg {{[0-9]+}}(%rsp)
52; AVX512-NEXT:    tileloadd (%r9,%r10), %tmm0
53; AVX512-NEXT:    tilestored %tmm0, (%rdi,%r8)
54; AVX512-NEXT:    leaq {{[0-9]+}}(%rsp), %rsi
55; AVX512-NEXT:    vmovdqu64 %zmm0, {{[0-9]+}}(%rsp)
56; AVX512-NEXT:    movb $1, {{[0-9]+}}(%rsp)
57; AVX512-NEXT:    movb %al, %dil
58; AVX512-NEXT:    movb %dil, {{[0-9]+}}(%rsp)
59; AVX512-NEXT:    movw %cx, {{[0-9]+}}(%rsp)
60; AVX512-NEXT:    ldtilecfg (%rsi)
61; AVX512-NEXT:    movl $buf, %esi
62; AVX512-NEXT:    movl $32, %edi
63; AVX512-NEXT:    tileloadd (%rsi,%rdi), %tmm0
64; AVX512-NEXT:    movl $64, %esi
65; AVX512-NEXT:    tilestored %tmm0, (%rdx,%rsi)
66; AVX512-NEXT:    jmp .LBB0_3
67; AVX512-NEXT:  .LBB0_2: # %if.else
68; AVX512-NEXT:    movw {{[-0-9]+}}(%r{{[sb]}}p), %ax # 2-byte Reload
69; AVX512-NEXT:    movw {{[-0-9]+}}(%r{{[sb]}}p), %cx # 2-byte Reload
70; AVX512-NEXT:    movq {{[-0-9]+}}(%r{{[sb]}}p), %rdx # 8-byte Reload
71; AVX512-NEXT:    movq {{[-0-9]+}}(%r{{[sb]}}p), %rdi # 8-byte Reload
72; AVX512-NEXT:    movq {{[-0-9]+}}(%r{{[sb]}}p), %r11 # 8-byte Reload
73; AVX512-NEXT:    vxorps %xmm0, %xmm0, %xmm0
74; AVX512-NEXT:    vmovdqu64 %zmm0, {{[0-9]+}}(%rsp)
75; AVX512-NEXT:    movb $1, {{[0-9]+}}(%rsp)
76; AVX512-NEXT:    movb %al, %sil
77; AVX512-NEXT:    movb %sil, {{[0-9]+}}(%rsp)
78; AVX512-NEXT:    movw $8, {{[0-9]+}}(%rsp)
79; AVX512-NEXT:    ldtilecfg {{[0-9]+}}(%rsp)
80; AVX512-NEXT:    movl $buf2, %r9d
81; AVX512-NEXT:    movl $32, %r10d
82; AVX512-NEXT:    movw $8, %si
83; AVX512-NEXT:    tileloadd (%r9,%r10), %tmm0
84; AVX512-NEXT:    movl $64, %r8d
85; AVX512-NEXT:    tilestored %tmm0, (%r11,%r8)
86; AVX512-NEXT:    vmovdqu64 %zmm0, {{[0-9]+}}(%rsp)
87; AVX512-NEXT:    movb $1, {{[0-9]+}}(%rsp)
88; AVX512-NEXT:    movb $8, {{[0-9]+}}(%rsp)
89; AVX512-NEXT:    movw %cx, {{[0-9]+}}(%rsp)
90; AVX512-NEXT:    ldtilecfg {{[0-9]+}}(%rsp)
91; AVX512-NEXT:    tileloadd (%r9,%r10), %tmm0
92; AVX512-NEXT:    tilestored %tmm0, (%rdi,%r8)
93; AVX512-NEXT:    leaq {{[0-9]+}}(%rsp), %rsi
94; AVX512-NEXT:    vmovdqu64 %zmm0, {{[0-9]+}}(%rsp)
95; AVX512-NEXT:    movb $1, {{[0-9]+}}(%rsp)
96; AVX512-NEXT:    movb %al, %dil
97; AVX512-NEXT:    movb %dil, {{[0-9]+}}(%rsp)
98; AVX512-NEXT:    movw %cx, {{[0-9]+}}(%rsp)
99; AVX512-NEXT:    ldtilecfg (%rsi)
100; AVX512-NEXT:    movl $buf2, %esi
101; AVX512-NEXT:    movl $32, %edi
102; AVX512-NEXT:    tileloadd (%rsi,%rdi), %tmm0
103; AVX512-NEXT:    movl $64, %esi
104; AVX512-NEXT:    tilestored %tmm0, (%rdx,%rsi)
105; AVX512-NEXT:  .LBB0_3: # %if.end
106; AVX512-NEXT:    movw {{[-0-9]+}}(%r{{[sb]}}p), %ax # 2-byte Reload
107; AVX512-NEXT:    movw {{[-0-9]+}}(%r{{[sb]}}p), %cx # 2-byte Reload
108; AVX512-NEXT:    movq {{[-0-9]+}}(%r{{[sb]}}p), %rdx # 8-byte Reload
109; AVX512-NEXT:    movq {{[-0-9]+}}(%r{{[sb]}}p), %r8 # 8-byte Reload
110; AVX512-NEXT:    movq {{[-0-9]+}}(%r{{[sb]}}p), %r9 # 8-byte Reload
111; AVX512-NEXT:    movq {{[-0-9]+}}(%r{{[sb]}}p), %r10 # 8-byte Reload
112; AVX512-NEXT:    vxorps %xmm0, %xmm0, %xmm0
113; AVX512-NEXT:    vmovdqu64 %zmm0, {{[0-9]+}}(%rsp)
114; AVX512-NEXT:    movb $1, {{[0-9]+}}(%rsp)
115; AVX512-NEXT:    movb %al, %sil
116; AVX512-NEXT:    movb %sil, {{[0-9]+}}(%rsp)
117; AVX512-NEXT:    movw %cx, {{[0-9]+}}(%rsp)
118; AVX512-NEXT:    movb %sil, {{[0-9]+}}(%rsp)
119; AVX512-NEXT:    movw $8, {{[0-9]+}}(%rsp)
120; AVX512-NEXT:    movb $8, {{[0-9]+}}(%rsp)
121; AVX512-NEXT:    movw %cx, {{[0-9]+}}(%rsp)
122; AVX512-NEXT:    movb %sil, {{[0-9]+}}(%rsp)
123; AVX512-NEXT:    movw %cx, {{[0-9]+}}(%rsp)
124; AVX512-NEXT:    ldtilecfg {{[0-9]+}}(%rsp)
125; AVX512-NEXT:    movl $64, %esi
126; AVX512-NEXT:    movw $8, %di
127; AVX512-NEXT:    tileloadd (%r10,%rsi), %tmm1
128; AVX512-NEXT:    tileloadd (%r9,%rsi), %tmm2
129; AVX512-NEXT:    tileloadd (%r8,%rsi), %tmm0
130; AVX512-NEXT:    tdpbssd %tmm2, %tmm1, %tmm0
131; AVX512-NEXT:    tilestored %tmm0, (%rdx,%rsi)
132; AVX512-NEXT:    leaq {{[0-9]+}}(%rsp), %rsi
133; AVX512-NEXT:    vmovdqu64 %zmm0, {{[0-9]+}}(%rsp)
134; AVX512-NEXT:    movb $1, {{[0-9]+}}(%rsp)
135; AVX512-NEXT:    movb %al, %dil
136; AVX512-NEXT:    movb %dil, {{[0-9]+}}(%rsp)
137; AVX512-NEXT:    movw %cx, {{[0-9]+}}(%rsp)
138; AVX512-NEXT:    ldtilecfg (%rsi)
139; AVX512-NEXT:    movl $64, %esi
140; AVX512-NEXT:    tileloadd (%rdx,%rsi), %tmm0
141; AVX512-NEXT:    movl $buf, %edx
142; AVX512-NEXT:    movl $32, %esi
143; AVX512-NEXT:    tilestored %tmm0, (%rdx,%rsi)
144; AVX512-NEXT:    movq %rbp, %rsp
145; AVX512-NEXT:    popq %rbp
146; AVX512-NEXT:    tilerelease
147; AVX512-NEXT:    vzeroupper
148; AVX512-NEXT:    retq
149entry:
150  %tobool.not = icmp eq i32 %cond, 0
151  br i1 %tobool.not, label %if.else, label %if.then
152
153if.then:                                          ; preds = %entry
154  %0 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 %row, i16 8, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32)
155  %1 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 8, i16 %col, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32)
156  %2 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 %row, i16 %col, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32)
157  br label %if.end
158
159if.else:                                          ; preds = %entry
160  %3 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 %row, i16 8, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf2, i64 0, i64 0), i64 32)
161  %4 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 8, i16 %col, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf2, i64 0, i64 0), i64 32)
162  %5 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 %row, i16 %col, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf2, i64 0, i64 0), i64 32)
163  br label %if.end
164
165if.end:                                           ; preds = %if.else, %if.then
166  %a.sroa.1094.0.in = phi x86_amx [ %3, %if.else ], [ %0, %if.then ]
167  %b.sroa.1069.0.in = phi x86_amx [ %4, %if.else ], [ %1, %if.then ]
168  %c.sroa.1044.0.in = phi x86_amx [ %5, %if.else ], [ %2, %if.then ]
169  %6 = tail call x86_amx @llvm.x86.tdpbssd.internal(i16 %row, i16 %col, i16 8, x86_amx %c.sroa.1044.0.in, x86_amx %a.sroa.1094.0.in, x86_amx %b.sroa.1069.0.in)
170  tail call void @llvm.x86.tilestored64.internal(i16 %row, i16 %col, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32, x86_amx %6)
171  ret void
172}
173
174declare x86_amx @llvm.x86.tileloadd64.internal(i16, i16, i8*, i64)
175declare x86_amx @llvm.x86.tdpbssd.internal(i16, i16, i16, x86_amx, x86_amx, x86_amx)
176declare void @llvm.x86.tilestored64.internal(i16, i16, i8*, i64, x86_amx)
177