1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+amx-int8 -mattr=+avx512f -verify-machineinstrs | FileCheck %s
3@buf = dso_local global [3072 x i8] zeroinitializer, align 16
4
5define dso_local void @test1(i16 signext %0, i16 signext %1) local_unnamed_addr {
6; CHECK-LABEL: test1:
7; CHECK:       # %bb.0:
8; CHECK-NEXT:    vpxord %zmm0, %zmm0, %zmm0
9; CHECK-NEXT:    vmovdqu64 %zmm0, -{{[0-9]+}}(%rsp)
10; CHECK-NEXT:    movb $1, -{{[0-9]+}}(%rsp)
11; CHECK-NEXT:    movb %dil, -{{[0-9]+}}(%rsp)
12; CHECK-NEXT:    movw $8, -{{[0-9]+}}(%rsp)
13; CHECK-NEXT:    movb $8, -{{[0-9]+}}(%rsp)
14; CHECK-NEXT:    movw %si, -{{[0-9]+}}(%rsp)
15; CHECK-NEXT:    movb %dil, -{{[0-9]+}}(%rsp)
16; CHECK-NEXT:    movw %si, -{{[0-9]+}}(%rsp)
17; CHECK-NEXT:    ldtilecfg -{{[0-9]+}}(%rsp)
18; CHECK-NEXT:    movl $buf, %eax
19; CHECK-NEXT:    movl $32, %ecx
20; CHECK-NEXT:    movw $8, %dx
21; CHECK-NEXT:    tileloadd (%rax,%rcx), %tmm0
22; CHECK-NEXT:    movl $buf+1024, %eax
23; CHECK-NEXT:    tileloadd (%rax,%rcx), %tmm1
24; CHECK-NEXT:    movl $buf+2048, %eax
25; CHECK-NEXT:    tileloadd (%rax,%rcx), %tmm2
26; CHECK-NEXT:    tdpbssd %tmm1, %tmm0, %tmm2
27; CHECK-NEXT:    tilestored %tmm2, (%rax,%rcx)
28; CHECK-NEXT:    tilerelease
29; CHECK-NEXT:    vzeroupper
30; CHECK-NEXT:    jmp foo # TAILCALL
31  %3 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 %0, i16 8, i8* getelementptr inbounds ([3072 x i8], [3072 x i8]* @buf, i64 0, i64 0), i64 32)
32  %4 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 8, i16 %1, i8* getelementptr inbounds ([3072 x i8], [3072 x i8]* @buf, i64 0, i64 1024), i64 32)
33  %5 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 %0, i16 %1, i8* getelementptr inbounds ([3072 x i8], [3072 x i8]* @buf, i64 0, i64 2048), i64 32)
34  %6 = tail call x86_amx @llvm.x86.tdpbssd.internal(i16 %0, i16 %1, i16 8, x86_amx %5, x86_amx %3, x86_amx %4)
35  tail call void @llvm.x86.tilestored64.internal(i16 %0, i16 %1, i8* getelementptr inbounds ([3072 x i8], [3072 x i8]* @buf, i64 0, i64 2048), i64 32, x86_amx %6)
36  tail call void @foo()
37  ret void
38}
39
40define dso_local void @test2(i16 signext %0, i16 signext %1) local_unnamed_addr {
41; CHECK-LABEL: test2:
42; CHECK:       # %bb.0:
43; CHECK-NEXT:    pushq %rbp
44; CHECK-NEXT:    .cfi_def_cfa_offset 16
45; CHECK-NEXT:    pushq %rbx
46; CHECK-NEXT:    .cfi_def_cfa_offset 24
47; CHECK-NEXT:    subq $72, %rsp
48; CHECK-NEXT:    .cfi_def_cfa_offset 96
49; CHECK-NEXT:    .cfi_offset %rbx, -24
50; CHECK-NEXT:    .cfi_offset %rbp, -16
51; CHECK-NEXT:    movl %esi, %ebx
52; CHECK-NEXT:    movl %edi, %ebp
53; CHECK-NEXT:    vpxord %zmm0, %zmm0, %zmm0
54; CHECK-NEXT:    vmovdqu64 %zmm0, {{[0-9]+}}(%rsp)
55; CHECK-NEXT:    movb $1, {{[0-9]+}}(%rsp)
56; CHECK-NEXT:    movb $8, {{[0-9]+}}(%rsp)
57; CHECK-NEXT:    movw %bx, {{[0-9]+}}(%rsp)
58; CHECK-NEXT:    movb %bpl, {{[0-9]+}}(%rsp)
59; CHECK-NEXT:    movw %bx, {{[0-9]+}}(%rsp)
60; CHECK-NEXT:    movb %bpl, {{[0-9]+}}(%rsp)
61; CHECK-NEXT:    movw $8, {{[0-9]+}}(%rsp)
62; CHECK-NEXT:    movb $8, {{[0-9]+}}(%rsp)
63; CHECK-NEXT:    movw %bx, {{[0-9]+}}(%rsp)
64; CHECK-NEXT:    movb %bpl, {{[0-9]+}}(%rsp)
65; CHECK-NEXT:    movw $8, {{[0-9]+}}(%rsp)
66; CHECK-NEXT:    ldtilecfg {{[0-9]+}}(%rsp)
67; CHECK-NEXT:    vzeroupper
68; CHECK-NEXT:    callq foo
69; CHECK-NEXT:    ldtilecfg {{[0-9]+}}(%rsp)
70; CHECK-NEXT:    xorl %eax, %eax
71; CHECK-NEXT:    testb %al, %al
72; CHECK-NEXT:    jne .LBB1_3
73; CHECK-NEXT:  # %bb.1: # %if.true
74; CHECK-NEXT:    movw $8, %ax
75; CHECK-NEXT:    tilezero %tmm0
76; CHECK-NEXT:    movl $32, %ecx
77; CHECK-NEXT:    movl $buf+1024, %edx
78; CHECK-NEXT:    tileloadd (%rdx,%rcx), %tmm1
79; CHECK-NEXT:    movl $buf+2048, %edx
80; CHECK-NEXT:    tileloadd (%rdx,%rcx), %tmm2
81; CHECK-NEXT:    tdpbssd %tmm2, %tmm1, %tmm0
82; CHECK-NEXT:    tilestored %tmm0, (%rdx,%rcx)
83; CHECK-NEXT:    jmp .LBB1_2
84; CHECK-NEXT:  .LBB1_3: # %if.false
85; CHECK-NEXT:    movl $buf, %eax
86; CHECK-NEXT:    movl $32, %ecx
87; CHECK-NEXT:    movw $8, %dx
88; CHECK-NEXT:    tileloadd (%rax,%rcx), %tmm3
89; CHECK-NEXT:    movl $buf+1024, %eax
90; CHECK-NEXT:    tileloadd (%rax,%rcx), %tmm4
91; CHECK-NEXT:    movl $buf+2048, %eax
92; CHECK-NEXT:    tileloadd (%rax,%rcx), %tmm2
93; CHECK-NEXT:    tdpbssd %tmm2, %tmm4, %tmm3
94; CHECK-NEXT:    tilestored %tmm3, (%rax,%rcx)
95; CHECK-NEXT:  .LBB1_2: # %if.true
96; CHECK-NEXT:    addq $72, %rsp
97; CHECK-NEXT:    .cfi_def_cfa_offset 24
98; CHECK-NEXT:    popq %rbx
99; CHECK-NEXT:    .cfi_def_cfa_offset 16
100; CHECK-NEXT:    popq %rbp
101; CHECK-NEXT:    .cfi_def_cfa_offset 8
102; CHECK-NEXT:    tilerelease
103; CHECK-NEXT:    retq
104  call void @foo()
105  br i1 undef, label %if.true, label %if.false
106
107if.true:
108  %t1 = tail call x86_amx @llvm.x86.tilezero.internal(i16 %0, i16 8)
109  %t2 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 8, i16 %1, i8* getelementptr inbounds ([3072 x i8], [3072 x i8]* @buf, i64 0, i64 1024), i64 32)
110  %t3 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 %0, i16 %1, i8* getelementptr inbounds ([3072 x i8], [3072 x i8]* @buf, i64 0, i64 2048), i64 32)
111  %t4 = tail call x86_amx @llvm.x86.tdpbssd.internal(i16 %0, i16 %1, i16 8, x86_amx %t1, x86_amx %t2, x86_amx %t3)
112  tail call void @llvm.x86.tilestored64.internal(i16 %0, i16 %1, i8* getelementptr inbounds ([3072 x i8], [3072 x i8]* @buf, i64 0, i64 2048), i64 32, x86_amx %t4)
113  br label %exit
114
115if.false:
116  %t5 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 %0, i16 8, i8* getelementptr inbounds ([3072 x i8], [3072 x i8]* @buf, i64 0, i64 0), i64 32)
117  %t6 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 8, i16 %1, i8* getelementptr inbounds ([3072 x i8], [3072 x i8]* @buf, i64 0, i64 1024), i64 32)
118  %t7 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 %0, i16 %1, i8* getelementptr inbounds ([3072 x i8], [3072 x i8]* @buf, i64 0, i64 2048), i64 32)
119  %t8 = tail call x86_amx @llvm.x86.tdpbssd.internal(i16 %0, i16 %1, i16 8, x86_amx %t5, x86_amx %t6, x86_amx %t7)
120  tail call void @llvm.x86.tilestored64.internal(i16 %0, i16 %1, i8* getelementptr inbounds ([3072 x i8], [3072 x i8]* @buf, i64 0, i64 2048), i64 32, x86_amx %t8)
121  br label %exit
122
123exit:
124  ret void
125}
126
127declare dso_local void @foo() local_unnamed_addr
128declare x86_amx @llvm.x86.tilezero.internal(i16, i16)
129declare x86_amx @llvm.x86.tileloadd64.internal(i16, i16, i8*, i64)
130declare x86_amx @llvm.x86.tdpbssd.internal(i16, i16, i16, x86_amx, x86_amx, x86_amx)
131declare void @llvm.x86.tilestored64.internal(i16, i16, i8*, i64, x86_amx)
132