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 $8192, %rsp # imm = 0x2000
14; AVX512-NEXT:    vxorps %xmm0, %xmm0, %xmm0
15; AVX512-NEXT:    vmovups %zmm0, {{[0-9]+}}(%rsp)
16; AVX512-NEXT:    movb $1, {{[0-9]+}}(%rsp)
17; AVX512-NEXT:    movw %dx, %ax
18; AVX512-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
19; AVX512-NEXT:    movw %si, %ax
20; AVX512-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
21; AVX512-NEXT:    leaq {{[0-9]+}}(%rsp), %rax
22; AVX512-NEXT:    movq %rax, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
23; AVX512-NEXT:    leaq {{[0-9]+}}(%rsp), %rax
24; AVX512-NEXT:    movq %rax, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
25; AVX512-NEXT:    leaq {{[0-9]+}}(%rsp), %rax
26; AVX512-NEXT:    movq %rax, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
27; AVX512-NEXT:    leaq {{[0-9]+}}(%rsp), %rax
28; AVX512-NEXT:    movq %rax, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
29; AVX512-NEXT:    cmpl $0, %edi
30; AVX512-NEXT:    je .LBB0_2
31; AVX512-NEXT:  # %bb.1: # %if.then
32; AVX512-NEXT:    movw {{[-0-9]+}}(%r{{[sb]}}p), %ax # 2-byte Reload
33; AVX512-NEXT:    movw {{[-0-9]+}}(%r{{[sb]}}p), %cx # 2-byte Reload
34; AVX512-NEXT:    movq {{[-0-9]+}}(%r{{[sb]}}p), %rdx # 8-byte Reload
35; AVX512-NEXT:    movq {{[-0-9]+}}(%r{{[sb]}}p), %rdi # 8-byte Reload
36; AVX512-NEXT:    movq {{[-0-9]+}}(%r{{[sb]}}p), %r8 # 8-byte Reload
37; AVX512-NEXT:    movl $buf, %r9d
38; AVX512-NEXT:    movl $32, %r10d
39; AVX512-NEXT:    movw $8, %si
40; AVX512-NEXT:    # implicit-def: $al
41; AVX512-NEXT:    movb %al, {{[0-9]+}}(%rsp)
42; AVX512-NEXT:    movw %si, {{[0-9]+}}(%rsp)
43; AVX512-NEXT:    ldtilecfg {{[0-9]+}}(%rsp)
44; AVX512-NEXT:    tileloadd (%r9,%r10), %tmm0
45; AVX512-NEXT:    movl $64, %r9d
46; AVX512-NEXT:    movw $8, %si
47; AVX512-NEXT:    tilestored %tmm0, (%r8,%r9)
48; AVX512-NEXT:    movl $buf, %r8d
49; AVX512-NEXT:    movl $32, %r9d
50; AVX512-NEXT:    movw $8, %si
51; AVX512-NEXT:    # implicit-def: $al
52; AVX512-NEXT:    movb %al, {{[0-9]+}}(%rsp)
53; AVX512-NEXT:    movw %cx, {{[0-9]+}}(%rsp)
54; AVX512-NEXT:    # implicit-def: $sil
55; AVX512-NEXT:    movb %sil, {{[0-9]+}}(%rsp)
56; AVX512-NEXT:    movw %cx, {{[0-9]+}}(%rsp)
57; AVX512-NEXT:    ldtilecfg {{[0-9]+}}(%rsp)
58; AVX512-NEXT:    tileloadd (%r8,%r9), %tmm0
59; AVX512-NEXT:    movl $64, %r8d
60; AVX512-NEXT:    movw $8, %si
61; AVX512-NEXT:    tilestored %tmm0, (%rdi,%r8)
62; AVX512-NEXT:    movl $buf, %esi
63; AVX512-NEXT:    movl $32, %edi
64; AVX512-NEXT:    tileloadd (%rsi,%rdi), %tmm0
65; AVX512-NEXT:    movl $64, %esi
66; AVX512-NEXT:    tilestored %tmm0, (%rdx,%rsi)
67; AVX512-NEXT:    jmp .LBB0_3
68; AVX512-NEXT:  .LBB0_2: # %if.else
69; AVX512-NEXT:    movw {{[-0-9]+}}(%r{{[sb]}}p), %ax # 2-byte Reload
70; AVX512-NEXT:    movw {{[-0-9]+}}(%r{{[sb]}}p), %cx # 2-byte Reload
71; AVX512-NEXT:    movq {{[-0-9]+}}(%r{{[sb]}}p), %rdx # 8-byte Reload
72; AVX512-NEXT:    movq {{[-0-9]+}}(%r{{[sb]}}p), %rdi # 8-byte Reload
73; AVX512-NEXT:    movq {{[-0-9]+}}(%r{{[sb]}}p), %r8 # 8-byte Reload
74; AVX512-NEXT:    movl $buf2, %r9d
75; AVX512-NEXT:    movl $32, %r10d
76; AVX512-NEXT:    movw $8, %si
77; AVX512-NEXT:    # implicit-def: $al
78; AVX512-NEXT:    movb %al, {{[0-9]+}}(%rsp)
79; AVX512-NEXT:    movw %si, {{[0-9]+}}(%rsp)
80; AVX512-NEXT:    ldtilecfg {{[0-9]+}}(%rsp)
81; AVX512-NEXT:    tileloadd (%r9,%r10), %tmm0
82; AVX512-NEXT:    movl $64, %r9d
83; AVX512-NEXT:    movw $8, %si
84; AVX512-NEXT:    tilestored %tmm0, (%r8,%r9)
85; AVX512-NEXT:    movl $buf2, %r8d
86; AVX512-NEXT:    movl $32, %r9d
87; AVX512-NEXT:    movw $8, %si
88; AVX512-NEXT:    # implicit-def: $al
89; AVX512-NEXT:    movb %al, {{[0-9]+}}(%rsp)
90; AVX512-NEXT:    movw %cx, {{[0-9]+}}(%rsp)
91; AVX512-NEXT:    # implicit-def: $sil
92; AVX512-NEXT:    movb %sil, {{[0-9]+}}(%rsp)
93; AVX512-NEXT:    movw %cx, {{[0-9]+}}(%rsp)
94; AVX512-NEXT:    ldtilecfg {{[0-9]+}}(%rsp)
95; AVX512-NEXT:    tileloadd (%r8,%r9), %tmm0
96; AVX512-NEXT:    movl $64, %r8d
97; AVX512-NEXT:    movw $8, %si
98; AVX512-NEXT:    tilestored %tmm0, (%rdi,%r8)
99; AVX512-NEXT:    movl $buf2, %esi
100; AVX512-NEXT:    movl $32, %edi
101; AVX512-NEXT:    tileloadd (%rsi,%rdi), %tmm0
102; AVX512-NEXT:    movl $64, %esi
103; AVX512-NEXT:    tilestored %tmm0, (%rdx,%rsi)
104; AVX512-NEXT:  .LBB0_3: # %if.end
105; AVX512-NEXT:    movw {{[-0-9]+}}(%r{{[sb]}}p), %ax # 2-byte Reload
106; AVX512-NEXT:    movw {{[-0-9]+}}(%r{{[sb]}}p), %cx # 2-byte Reload
107; AVX512-NEXT:    movq {{[-0-9]+}}(%r{{[sb]}}p), %rdx # 8-byte Reload
108; AVX512-NEXT:    movq {{[-0-9]+}}(%r{{[sb]}}p), %rsi # 8-byte Reload
109; AVX512-NEXT:    movq {{[-0-9]+}}(%r{{[sb]}}p), %r9 # 8-byte Reload
110; AVX512-NEXT:    movq {{[-0-9]+}}(%r{{[sb]}}p), %r8 # 8-byte Reload
111; AVX512-NEXT:    movl $64, %r10d
112; AVX512-NEXT:    movw $8, %di
113; AVX512-NEXT:    # implicit-def: $al
114; AVX512-NEXT:    movb %al, {{[0-9]+}}(%rsp)
115; AVX512-NEXT:    movw %di, {{[0-9]+}}(%rsp)
116; AVX512-NEXT:    ldtilecfg {{[0-9]+}}(%rsp)
117; AVX512-NEXT:    tileloadd (%r8,%r10), %tmm0
118; AVX512-NEXT:    movabsq $64, %r8
119; AVX512-NEXT:    tilestored %tmm0, 1024(%rsp,%r8) # 1024-byte Folded Spill
120; AVX512-NEXT:    movl $64, %r10d
121; AVX512-NEXT:    movw $8, %r8w
122; AVX512-NEXT:    # implicit-def: $al
123; AVX512-NEXT:    movb %al, {{[0-9]+}}(%rsp)
124; AVX512-NEXT:    movw %cx, {{[0-9]+}}(%rsp)
125; AVX512-NEXT:    # implicit-def: $al
126; AVX512-NEXT:    movb %al, {{[0-9]+}}(%rsp)
127; AVX512-NEXT:    movw %cx, {{[0-9]+}}(%rsp)
128; AVX512-NEXT:    # implicit-def: $al
129; AVX512-NEXT:    movb %al, {{[0-9]+}}(%rsp)
130; AVX512-NEXT:    movw %di, {{[0-9]+}}(%rsp)
131; AVX512-NEXT:    # implicit-def: $al
132; AVX512-NEXT:    movb %al, {{[0-9]+}}(%rsp)
133; AVX512-NEXT:    movw %cx, {{[0-9]+}}(%rsp)
134; AVX512-NEXT:    # implicit-def: $r8b
135; AVX512-NEXT:    movb %r8b, {{[0-9]+}}(%rsp)
136; AVX512-NEXT:    movw %cx, {{[0-9]+}}(%rsp)
137; AVX512-NEXT:    ldtilecfg {{[0-9]+}}(%rsp)
138; AVX512-NEXT:    tileloadd (%r9,%r10), %tmm2
139; AVX512-NEXT:    movl $64, %r8d
140; AVX512-NEXT:    tileloadd (%rsi,%r8), %tmm0
141; AVX512-NEXT:    movw $8, %si
142; AVX512-NEXT:    movabsq $64, %r8
143; AVX512-NEXT:    tileloadd 1024(%rsp,%r8), %tmm1 # 1024-byte Folded Reload
144; AVX512-NEXT:    tdpbssd %tmm2, %tmm1, %tmm0
145; AVX512-NEXT:    movl $64, %esi
146; AVX512-NEXT:    tilestored %tmm0, (%rdx,%rsi)
147; AVX512-NEXT:    movl $64, %esi
148; AVX512-NEXT:    tileloadd (%rdx,%rsi), %tmm0
149; AVX512-NEXT:    movl $buf, %edx
150; AVX512-NEXT:    movl $32, %esi
151; AVX512-NEXT:    tilestored %tmm0, (%rdx,%rsi)
152; AVX512-NEXT:    movq %rbp, %rsp
153; AVX512-NEXT:    popq %rbp
154; AVX512-NEXT:    tilerelease
155; AVX512-NEXT:    vzeroupper
156; AVX512-NEXT:    retq
157entry:
158  %tobool.not = icmp eq i32 %cond, 0
159  br i1 %tobool.not, label %if.else, label %if.then
160
161if.then:                                          ; preds = %entry
162  %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)
163  %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)
164  %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)
165  br label %if.end
166
167if.else:                                          ; preds = %entry
168  %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)
169  %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)
170  %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)
171  br label %if.end
172
173if.end:                                           ; preds = %if.else, %if.then
174  %a.sroa.1094.0.in = phi x86_amx [ %3, %if.else ], [ %0, %if.then ]
175  %b.sroa.1069.0.in = phi x86_amx [ %4, %if.else ], [ %1, %if.then ]
176  %c.sroa.1044.0.in = phi x86_amx [ %5, %if.else ], [ %2, %if.then ]
177  %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)
178  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)
179  ret void
180}
181
182declare x86_amx @llvm.x86.tileloadd64.internal(i16, i16, i8*, i64)
183declare x86_amx @llvm.x86.tdpbssd.internal(i16, i16, i16, x86_amx, x86_amx, x86_amx)
184declare void @llvm.x86.tilestored64.internal(i16, i16, i8*, i64, x86_amx)
185