xref: /llvm-project/llvm/test/CodeGen/X86/amx_transpose_intrinsics.ll (revision fbb9d49506baa05a613ab88f983d31e0f838dbae) 
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+amx-bf16,+amx-fp16,+amx-complex,+amx-transpose | FileCheck %s
3; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+amx-bf16,+amx-fp16,+amx-complex,+amx-transpose,+egpr --show-mc-encoding | FileCheck %s --check-prefix=EGPR
4
5define void @test_amx(i32 %rv32, i64 %stride, i64 %rvalue, i8* %addr1, <4 x float> %xmm) #0 {
6; CHECK-LABEL: test_amx:
7; CHECK:       # %bb.0:
8; CHECK-NEXT:    t2rpntlvwz0 (%rcx,%rsi), %tmm0
9; CHECK-NEXT:    t2rpntlvwz0t1 (%rcx,%rsi), %tmm2
10; CHECK-NEXT:    t2rpntlvwz1 (%rcx,%rsi), %tmm0
11; CHECK-NEXT:    t2rpntlvwz1t1 (%rcx,%rsi), %tmm2
12; CHECK-NEXT:    ttransposed %tmm3, %tmm1
13; CHECK-NEXT:    ttdpbf16ps %tmm3, %tmm2, %tmm1
14; CHECK-NEXT:    ttdpfp16ps %tmm6, %tmm5, %tmm4
15; CHECK-NEXT:    ttcmmimfp16ps %tmm3, %tmm2, %tmm1
16; CHECK-NEXT:    ttcmmrlfp16ps %tmm3, %tmm2, %tmm1
17; CHECK-NEXT:    tconjtcmmimfp16ps %tmm3, %tmm2, %tmm1
18; CHECK-NEXT:    tconjtfp16 %tmm2, %tmm1
19; CHECK-NEXT:    retq
20;
21; EGPR-LABEL: test_amx:
22; EGPR:       # %bb.0:
23; EGPR-NEXT:    t2rpntlvwz0 (%rcx,%rsi), %tmm0 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x78,0x6e,0x04,0x31]
24; EGPR-NEXT:    t2rpntlvwz0t1 (%rcx,%rsi), %tmm2 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x78,0x6f,0x14,0x31]
25; EGPR-NEXT:    t2rpntlvwz1 (%rcx,%rsi), %tmm0 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0x6e,0x04,0x31]
26; EGPR-NEXT:    t2rpntlvwz1t1 (%rcx,%rsi), %tmm2 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0x6f,0x14,0x31]
27; EGPR-NEXT:    ttransposed %tmm3, %tmm1 # encoding: [0xc4,0xe2,0x7a,0x5f,0xcb]
28; EGPR-NEXT:    ttdpbf16ps %tmm3, %tmm2, %tmm1 # encoding: [0xc4,0xe2,0x62,0x6c,0xca]
29; EGPR-NEXT:    ttdpfp16ps %tmm6, %tmm5, %tmm4 # encoding: [0xc4,0xe2,0x4b,0x6c,0xe5]
30; EGPR-NEXT:    ttcmmimfp16ps %tmm3, %tmm2, %tmm1 # encoding: [0xc4,0xe2,0x63,0x6b,0xca]
31; EGPR-NEXT:    ttcmmrlfp16ps %tmm3, %tmm2, %tmm1 # encoding: [0xc4,0xe2,0x62,0x6b,0xca]
32; EGPR-NEXT:    tconjtcmmimfp16ps %tmm3, %tmm2, %tmm1 # encoding: [0xc4,0xe2,0x60,0x6b,0xca]
33; EGPR-NEXT:    tconjtfp16 %tmm2, %tmm1 # encoding: [0xc4,0xe2,0x79,0x6b,0xca]
34; EGPR-NEXT:    retq # encoding: [0xc3]
35  call void @llvm.x86.t2rpntlvwz0(i8 1, i8* %addr1, i64 %stride)
36  call void @llvm.x86.t2rpntlvwz0t1(i8 2, i8* %addr1, i64 %stride)
37  call void @llvm.x86.t2rpntlvwz1(i8 1, i8* %addr1, i64 %stride)
38  call void @llvm.x86.t2rpntlvwz1t1(i8 2, i8* %addr1, i64 %stride)
39  call void @llvm.x86.ttransposed(i8 1, i8 3)
40  call void @llvm.x86.ttdpbf16ps(i8 1, i8 2, i8 3)
41  call void @llvm.x86.ttdpfp16ps(i8 4, i8 5, i8 6)
42  call void @llvm.x86.ttcmmimfp16ps(i8 1, i8 2, i8 3)
43  call void @llvm.x86.ttcmmrlfp16ps(i8 1, i8 2, i8 3)
44  call void @llvm.x86.tconjtcmmimfp16ps(i8 1, i8 2, i8 3)
45  call void @llvm.x86.tconjtfp16(i8 1, i8 2)
46  ret void
47}
48
49declare void @llvm.x86.t2rpntlvwz0(i8 %tile1, i8* %addr1, i64 %stride)
50declare void @llvm.x86.t2rpntlvwz0t1(i8 %tile1, i8* %addr1, i64 %stride)
51declare void @llvm.x86.t2rpntlvwz1(i8 %tile1, i8* %addr1, i64 %stride)
52declare void @llvm.x86.t2rpntlvwz1t1(i8 %tile1, i8* %addr1, i64 %stride)
53declare void @llvm.x86.ttransposed(i8 %tile0, i8 %tile1)
54declare void @llvm.x86.ttdpbf16ps(i8 %tile0, i8 %tile1, i8 %tile2)
55declare void @llvm.x86.ttdpfp16ps(i8 %tile0, i8 %tile1, i8 %tile2)
56declare void @llvm.x86.ttcmmimfp16ps(i8 %A, i8 %B, i8 %C)
57declare void @llvm.x86.ttcmmrlfp16ps(i8 %A, i8 %B, i8 %C)
58declare void @llvm.x86.tconjtcmmimfp16ps(i8 %A, i8 %B, i8 %C)
59declare void @llvm.x86.tconjtfp16(i8 %A, i8 %B)
60
61define void @test_amx2(i8* %pointer, i8* %base, i64 %stride) #0 {
62; CHECK-LABEL: test_amx2:
63; CHECK:       # %bb.0:
64; CHECK-NEXT:    pushq %rbp
65; CHECK-NEXT:    subq $2928, %rsp # imm = 0xB70
66; CHECK-NEXT:    vxorps %xmm0, %xmm0, %xmm0
67; CHECK-NEXT:    vmovups %zmm0, {{[0-9]+}}(%rsp)
68; CHECK-NEXT:    movb $1, {{[0-9]+}}(%rsp)
69; CHECK-NEXT:    movb $8, {{[0-9]+}}(%rsp)
70; CHECK-NEXT:    movw $8, {{[0-9]+}}(%rsp)
71; CHECK-NEXT:    movb $8, {{[0-9]+}}(%rsp)
72; CHECK-NEXT:    movw $8, {{[0-9]+}}(%rsp)
73; CHECK-NEXT:    movb $8, {{[0-9]+}}(%rsp)
74; CHECK-NEXT:    movw $8, {{[0-9]+}}(%rsp)
75; CHECK-NEXT:    movb $8, {{[0-9]+}}(%rsp)
76; CHECK-NEXT:    movw $8, {{[0-9]+}}(%rsp)
77; CHECK-NEXT:    ldtilecfg {{[0-9]+}}(%rsp)
78; CHECK-NEXT:    movw $8, %ax
79; CHECK-NEXT:    tileloadd (%rsi,%rdx), %tmm0
80; CHECK-NEXT:    tilezero %tmm1
81; CHECK-NEXT:    tilezero %tmm2
82; CHECK-NEXT:    ttdpbf16ps %tmm1, %tmm0, %tmm2
83; CHECK-NEXT:    ttdpfp16ps %tmm1, %tmm0, %tmm2
84; CHECK-NEXT:    ttcmmimfp16ps %tmm1, %tmm0, %tmm2
85; CHECK-NEXT:    ttcmmrlfp16ps %tmm1, %tmm0, %tmm2
86; CHECK-NEXT:    movabsq $64, %rbp
87; CHECK-NEXT:    tilestored %tmm2, 896(%rsp,%rbp) # 1024-byte Folded Spill
88; CHECK-NEXT:    tileloadd 896(%rsp,%rbp), %tmm3 # 1024-byte Folded Reload
89; CHECK-NEXT:    tconjtcmmimfp16ps %tmm1, %tmm0, %tmm3
90; CHECK-NEXT:    tconjtfp16 %tmm3, %tmm0
91; CHECK-NEXT:    tilestored %tmm2, (%rdi,%rdx)
92; CHECK-NEXT:    addq $2928, %rsp # imm = 0xB70
93; CHECK-NEXT:    popq %rbp
94; CHECK-NEXT:    tilerelease
95; CHECK-NEXT:    vzeroupper
96; CHECK-NEXT:    retq
97;
98; EGPR-LABEL: test_amx2:
99; EGPR:       # %bb.0:
100; EGPR-NEXT:    pushq %rbp # encoding: [0x55]
101; EGPR-NEXT:    subq $2928, %rsp # encoding: [0x48,0x81,0xec,0x70,0x0b,0x00,0x00]
102; EGPR-NEXT:    # imm = 0xB70
103; EGPR-NEXT:    vxorps %xmm0, %xmm0, %xmm0 # encoding: [0xc5,0xf8,0x57,0xc0]
104; EGPR-NEXT:    vmovups %zmm0, {{[0-9]+}}(%rsp) # encoding: [0x62,0xf1,0x7c,0x48,0x11,0x44,0x24,0x0d]
105; EGPR-NEXT:    movb $1, {{[0-9]+}}(%rsp) # encoding: [0xc6,0x84,0x24,0x40,0x03,0x00,0x00,0x01]
106; EGPR-NEXT:    movb $8, {{[0-9]+}}(%rsp) # encoding: [0xc6,0x84,0x24,0x70,0x03,0x00,0x00,0x08]
107; EGPR-NEXT:    movw $8, {{[0-9]+}}(%rsp) # encoding: [0x66,0xc7,0x84,0x24,0x50,0x03,0x00,0x00,0x08,0x00]
108; EGPR-NEXT:    movb $8, {{[0-9]+}}(%rsp) # encoding: [0xc6,0x84,0x24,0x71,0x03,0x00,0x00,0x08]
109; EGPR-NEXT:    movw $8, {{[0-9]+}}(%rsp) # encoding: [0x66,0xc7,0x84,0x24,0x52,0x03,0x00,0x00,0x08,0x00]
110; EGPR-NEXT:    movb $8, {{[0-9]+}}(%rsp) # encoding: [0xc6,0x84,0x24,0x72,0x03,0x00,0x00,0x08]
111; EGPR-NEXT:    movw $8, {{[0-9]+}}(%rsp) # encoding: [0x66,0xc7,0x84,0x24,0x54,0x03,0x00,0x00,0x08,0x00]
112; EGPR-NEXT:    movb $8, {{[0-9]+}}(%rsp) # encoding: [0xc6,0x84,0x24,0x73,0x03,0x00,0x00,0x08]
113; EGPR-NEXT:    movw $8, {{[0-9]+}}(%rsp) # encoding: [0x66,0xc7,0x84,0x24,0x56,0x03,0x00,0x00,0x08,0x00]
114; EGPR-NEXT:    ldtilecfg {{[0-9]+}}(%rsp) # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x78,0x49,0x84,0x24,0x40,0x03,0x00,0x00]
115; EGPR-NEXT:    movw $8, %ax # encoding: [0x66,0xb8,0x08,0x00]
116; EGPR-NEXT:    tileloadd (%rsi,%rdx), %tmm0 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7b,0x4b,0x04,0x16]
117; EGPR-NEXT:    tilezero %tmm1 # encoding: [0xc4,0xe2,0x7b,0x49,0xc8]
118; EGPR-NEXT:    tilezero %tmm2 # encoding: [0xc4,0xe2,0x7b,0x49,0xd0]
119; EGPR-NEXT:    ttdpbf16ps %tmm1, %tmm0, %tmm2 # encoding: [0xc4,0xe2,0x72,0x6c,0xd0]
120; EGPR-NEXT:    ttdpfp16ps %tmm1, %tmm0, %tmm2 # encoding: [0xc4,0xe2,0x73,0x6c,0xd0]
121; EGPR-NEXT:    ttcmmimfp16ps %tmm1, %tmm0, %tmm2 # encoding: [0xc4,0xe2,0x73,0x6b,0xd0]
122; EGPR-NEXT:    ttcmmrlfp16ps %tmm1, %tmm0, %tmm2 # encoding: [0xc4,0xe2,0x72,0x6b,0xd0]
123; EGPR-NEXT:    movabsq $64, %rbp # encoding: [0x48,0xbd,0x40,0x00,0x00,0x00,0x00,0x00,0x00,0x00]
124; EGPR-NEXT:    tilestored %tmm2, 896(%rsp,%rbp) # 1024-byte Folded Spill
125; EGPR-NEXT:    # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7a,0x4b,0x94,0x2c,0x80,0x03,0x00,0x00]
126; EGPR-NEXT:    tileloadd 896(%rsp,%rbp), %tmm3 # 1024-byte Folded Reload
127; EGPR-NEXT:    # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7b,0x4b,0x9c,0x2c,0x80,0x03,0x00,0x00]
128; EGPR-NEXT:    tconjtcmmimfp16ps %tmm1, %tmm0, %tmm3 # encoding: [0xc4,0xe2,0x70,0x6b,0xd8]
129; EGPR-NEXT:    tconjtfp16 %tmm3, %tmm0 # encoding: [0xc4,0xe2,0x79,0x6b,0xc3]
130; EGPR-NEXT:    tilestored %tmm2, (%rdi,%rdx) # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7a,0x4b,0x14,0x17]
131; EGPR-NEXT:    addq $2928, %rsp # encoding: [0x48,0x81,0xc4,0x70,0x0b,0x00,0x00]
132; EGPR-NEXT:    # imm = 0xB70
133; EGPR-NEXT:    popq %rbp # encoding: [0x5d]
134; EGPR-NEXT:    tilerelease # encoding: [0xc4,0xe2,0x78,0x49,0xc0]
135; EGPR-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
136; EGPR-NEXT:    retq # encoding: [0xc3]
137
138  %a = call x86_amx @llvm.x86.tileloadd64.internal(i16 8, i16 8, i8* %base, i64 %stride)
139  %b = call x86_amx @llvm.x86.tilezero.internal(i16 8, i16 8)
140  %c = call x86_amx @llvm.x86.tilezero.internal(i16 8, i16 8)
141  %c1 = call x86_amx @llvm.x86.ttdpbf16ps.internal(i16 8, i16 8, i16 8, x86_amx %c, x86_amx %a, x86_amx %b)
142  %c2 = call x86_amx @llvm.x86.ttdpfp16ps.internal(i16 8, i16 8, i16 8, x86_amx %c1, x86_amx %a, x86_amx %b)
143  %c3 = call x86_amx @llvm.x86.ttcmmimfp16ps.internal(i16 8, i16 8, i16 8, x86_amx %c2, x86_amx %a, x86_amx %b)
144  %c4 = call x86_amx @llvm.x86.ttcmmrlfp16ps.internal(i16 8, i16 8, i16 8, x86_amx %c3, x86_amx %a, x86_amx %b)
145  %c5 = call x86_amx @llvm.x86.tconjtcmmimfp16ps.internal(i16 8, i16 8, i16 8, x86_amx %c4, x86_amx %a, x86_amx %b)
146  %c6 = call x86_amx @llvm.x86.tconjtfp16.internal(i16 8, i16 8, x86_amx %c5)
147
148  call void @llvm.x86.tilestored64.internal(i16 8, i16 8, i8* %pointer, i64 %stride, x86_amx %c4)
149  ret void
150}
151
152define void @test_amx3(i8* %pointer, i8* %base, i64 %stride) #0 {
153; CHECK-LABEL: test_amx3:
154; CHECK:       # %bb.0:
155; CHECK-NEXT:    vxorps %xmm0, %xmm0, %xmm0
156; CHECK-NEXT:    vmovups %zmm0, -{{[0-9]+}}(%rsp)
157; CHECK-NEXT:    movb $1, -{{[0-9]+}}(%rsp)
158; CHECK-NEXT:    movb $8, -{{[0-9]+}}(%rsp)
159; CHECK-NEXT:    movw $8, -{{[0-9]+}}(%rsp)
160; CHECK-NEXT:    movb $8, -{{[0-9]+}}(%rsp)
161; CHECK-NEXT:    movw $8, -{{[0-9]+}}(%rsp)
162; CHECK-NEXT:    movb $0, -{{[0-9]+}}(%rsp)
163; CHECK-NEXT:    movw $0, -{{[0-9]+}}(%rsp)
164; CHECK-NEXT:    ldtilecfg -{{[0-9]+}}(%rsp)
165; CHECK-NEXT:    xorl %eax, %eax
166; CHECK-NEXT:    movw $8, %cx
167; CHECK-NEXT:    t2rpntlvwz0 (%rsi,%rdx), %tmm4
168; CHECK-NEXT:    t2rpntlvwz0t1 (%rsi,%rdx), %tmm4
169; CHECK-NEXT:    t2rpntlvwz1 (%rsi,%rdx), %tmm4
170; CHECK-NEXT:    t2rpntlvwz1t1 (%rsi,%rdx), %tmm4
171; CHECK-NEXT:    ttransposed %tmm4, %tmm0
172; CHECK-NEXT:    tilestored %tmm0, (%rdi,%rdx)
173; CHECK-NEXT:    tilerelease
174; CHECK-NEXT:    vzeroupper
175; CHECK-NEXT:    retq
176;
177; EGPR-LABEL: test_amx3:
178; EGPR:       # %bb.0:
179; EGPR-NEXT:    vxorps %xmm0, %xmm0, %xmm0 # encoding: [0xc5,0xf8,0x57,0xc0]
180; EGPR-NEXT:    vmovups %zmm0, -{{[0-9]+}}(%rsp) # encoding: [0x62,0xf1,0x7c,0x48,0x11,0x44,0x24,0xff]
181; EGPR-NEXT:    movb $1, -{{[0-9]+}}(%rsp) # encoding: [0xc6,0x44,0x24,0xc0,0x01]
182; EGPR-NEXT:    movb $8, -{{[0-9]+}}(%rsp) # encoding: [0xc6,0x44,0x24,0xf0,0x08]
183; EGPR-NEXT:    movw $8, -{{[0-9]+}}(%rsp) # encoding: [0x66,0xc7,0x44,0x24,0xd0,0x08,0x00]
184; EGPR-NEXT:    movb $8, -{{[0-9]+}}(%rsp) # encoding: [0xc6,0x44,0x24,0xf4,0x08]
185; EGPR-NEXT:    movw $8, -{{[0-9]+}}(%rsp) # encoding: [0x66,0xc7,0x44,0x24,0xd8,0x08,0x00]
186; EGPR-NEXT:    movb $0, -{{[0-9]+}}(%rsp) # encoding: [0xc6,0x44,0x24,0xf5,0x00]
187; EGPR-NEXT:    movw $0, -{{[0-9]+}}(%rsp) # encoding: [0x66,0xc7,0x44,0x24,0xda,0x00,0x00]
188; EGPR-NEXT:    ldtilecfg -{{[0-9]+}}(%rsp) # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x78,0x49,0x44,0x24,0xc0]
189; EGPR-NEXT:    xorl %eax, %eax # encoding: [0x31,0xc0]
190; EGPR-NEXT:    movw $8, %cx # encoding: [0x66,0xb9,0x08,0x00]
191; EGPR-NEXT:    t2rpntlvwz0 (%rsi,%rdx), %tmm4 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x78,0x6e,0x24,0x16]
192; EGPR-NEXT:    t2rpntlvwz0t1 (%rsi,%rdx), %tmm4 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x78,0x6f,0x24,0x16]
193; EGPR-NEXT:    t2rpntlvwz1 (%rsi,%rdx), %tmm4 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0x6e,0x24,0x16]
194; EGPR-NEXT:    t2rpntlvwz1t1 (%rsi,%rdx), %tmm4 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0x6f,0x24,0x16]
195; EGPR-NEXT:    ttransposed %tmm4, %tmm0 # encoding: [0xc4,0xe2,0x7a,0x5f,0xc4]
196; EGPR-NEXT:    tilestored %tmm0, (%rdi,%rdx) # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7a,0x4b,0x04,0x17]
197; EGPR-NEXT:    tilerelease # encoding: [0xc4,0xe2,0x78,0x49,0xc0]
198; EGPR-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
199; EGPR-NEXT:    retq # encoding: [0xc3]
200  %1 = call { x86_amx, x86_amx } @llvm.x86.t2rpntlvwz0.internal(i16 8, i16 8, i16 0, i8* %base, i64 %stride)
201  %2 = call { x86_amx, x86_amx } @llvm.x86.t2rpntlvwz0t1.internal(i16 8, i16 8, i16 0, i8* %base, i64 %stride)
202  %3 = call { x86_amx, x86_amx } @llvm.x86.t2rpntlvwz1.internal(i16 8, i16 8, i16 0, i8* %base, i64 %stride)
203  %4 = call { x86_amx, x86_amx } @llvm.x86.t2rpntlvwz1t1.internal(i16 8, i16 8, i16 0, i8* %base, i64 %stride)
204  %5 = extractvalue { x86_amx, x86_amx } %4, 0
205  %6 = call x86_amx @llvm.x86.ttransposed.internal(i16 8, i16 8, x86_amx %5)
206  call void @llvm.x86.tilestored64.internal(i16 8, i16 8, i8* %pointer, i64 %stride, x86_amx %6)
207  ret void
208}
209
210define void @test_amx_spill(i8* %pointer, i8* %base, i64 %stride) #0 {
211; CHECK-LABEL: test_amx_spill:
212; CHECK:       # %bb.0:
213; CHECK-NEXT:    subq $6088, %rsp # imm = 0x17C8
214; CHECK-NEXT:    vxorps %xmm0, %xmm0, %xmm0
215; CHECK-NEXT:    vmovups %zmm0, -{{[0-9]+}}(%rsp)
216; CHECK-NEXT:    movb $1, -{{[0-9]+}}(%rsp)
217; CHECK-NEXT:    movb $8, -{{[0-9]+}}(%rsp)
218; CHECK-NEXT:    movw $8, -{{[0-9]+}}(%rsp)
219; CHECK-NEXT:    movb $8, -{{[0-9]+}}(%rsp)
220; CHECK-NEXT:    movw $8, -{{[0-9]+}}(%rsp)
221; CHECK-NEXT:    movb $8, -{{[0-9]+}}(%rsp)
222; CHECK-NEXT:    movw $8, -{{[0-9]+}}(%rsp)
223; CHECK-NEXT:    movb $8, -{{[0-9]+}}(%rsp)
224; CHECK-NEXT:    movw $8, -{{[0-9]+}}(%rsp)
225; CHECK-NEXT:    movb $8, -{{[0-9]+}}(%rsp)
226; CHECK-NEXT:    movw $8, -{{[0-9]+}}(%rsp)
227; CHECK-NEXT:    ldtilecfg -{{[0-9]+}}(%rsp)
228; CHECK-NEXT:    movw $8, %ax
229; CHECK-NEXT:    tileloadd (%rsi,%rdx), %tmm0
230; CHECK-NEXT:    t2rpntlvwz0 (%rsi,%rdx), %tmm4
231; CHECK-NEXT:    t2rpntlvwz0t1 (%rsi,%rdx), %tmm6
232; CHECK-NEXT:    movabsq $64, %rcx
233; CHECK-NEXT:    tilestored %tmm6, 4032(%rsp,%rcx) # 1024-byte Folded Spill
234; CHECK-NEXT:    tilestored %tmm7, 5056(%rsp,%rcx) # 1024-byte Folded Spill
235; CHECK-NEXT:    t2rpntlvwz1 (%rsi,%rdx), %tmm6
236; CHECK-NEXT:    tilestored %tmm6, 1984(%rsp,%rcx) # 1024-byte Folded Spill
237; CHECK-NEXT:    tilestored %tmm7, 3008(%rsp,%rcx) # 1024-byte Folded Spill
238; CHECK-NEXT:    t2rpntlvwz1t1 (%rsi,%rdx), %tmm6
239; CHECK-NEXT:    tilestored %tmm6, -64(%rsp,%rcx) # 1024-byte Folded Spill
240; CHECK-NEXT:    tilestored %tmm7, 960(%rsp,%rcx) # 1024-byte Folded Spill
241; CHECK-NEXT:    t2rpntlvwz0 (%rsi,%rdx), %tmm6
242; CHECK-NEXT:    tilestored %tmm4, (%rsi,%rdx)
243; CHECK-NEXT:    tilestored %tmm5, (%rsi,%rdx)
244; CHECK-NEXT:    tileloadd 4032(%rsp,%rcx), %tmm4 # 1024-byte Folded Reload
245; CHECK-NEXT:    tileloadd 5056(%rsp,%rcx), %tmm5 # 1024-byte Folded Reload
246; CHECK-NEXT:    tilestored %tmm4, (%rsi,%rdx)
247; CHECK-NEXT:    tilestored %tmm5, (%rsi,%rdx)
248; CHECK-NEXT:    tileloadd 1984(%rsp,%rcx), %tmm4 # 1024-byte Folded Reload
249; CHECK-NEXT:    tileloadd 3008(%rsp,%rcx), %tmm5 # 1024-byte Folded Reload
250; CHECK-NEXT:    tilestored %tmm4, (%rsi,%rdx)
251; CHECK-NEXT:    tilestored %tmm5, (%rsi,%rdx)
252; CHECK-NEXT:    tileloadd -64(%rsp,%rcx), %tmm4 # 1024-byte Folded Reload
253; CHECK-NEXT:    tileloadd 960(%rsp,%rcx), %tmm5 # 1024-byte Folded Reload
254; CHECK-NEXT:    tilestored %tmm4, (%rsi,%rdx)
255; CHECK-NEXT:    tilestored %tmm5, (%rsi,%rdx)
256; CHECK-NEXT:    tilestored %tmm6, (%rsi,%rdx)
257; CHECK-NEXT:    tilestored %tmm7, (%rsi,%rdx)
258; CHECK-NEXT:    addq $6088, %rsp # imm = 0x17C8
259; CHECK-NEXT:    tilerelease
260; CHECK-NEXT:    vzeroupper
261; CHECK-NEXT:    retq
262;
263; EGPR-LABEL: test_amx_spill:
264; EGPR:       # %bb.0:
265; EGPR-NEXT:    subq $6088, %rsp # encoding: [0x48,0x81,0xec,0xc8,0x17,0x00,0x00]
266; EGPR-NEXT:    # imm = 0x17C8
267; EGPR-NEXT:    vxorps %xmm0, %xmm0, %xmm0 # encoding: [0xc5,0xf8,0x57,0xc0]
268; EGPR-NEXT:    vmovups %zmm0, -{{[0-9]+}}(%rsp) # encoding: [0x62,0xf1,0x7c,0x48,0x11,0x44,0x24,0xfe]
269; EGPR-NEXT:    movb $1, -{{[0-9]+}}(%rsp) # encoding: [0xc6,0x44,0x24,0x80,0x01]
270; EGPR-NEXT:    movb $8, -{{[0-9]+}}(%rsp) # encoding: [0xc6,0x44,0x24,0xb0,0x08]
271; EGPR-NEXT:    movw $8, -{{[0-9]+}}(%rsp) # encoding: [0x66,0xc7,0x44,0x24,0x90,0x08,0x00]
272; EGPR-NEXT:    movb $8, -{{[0-9]+}}(%rsp) # encoding: [0xc6,0x44,0x24,0xb4,0x08]
273; EGPR-NEXT:    movw $8, -{{[0-9]+}}(%rsp) # encoding: [0x66,0xc7,0x44,0x24,0x98,0x08,0x00]
274; EGPR-NEXT:    movb $8, -{{[0-9]+}}(%rsp) # encoding: [0xc6,0x44,0x24,0xb5,0x08]
275; EGPR-NEXT:    movw $8, -{{[0-9]+}}(%rsp) # encoding: [0x66,0xc7,0x44,0x24,0x9a,0x08,0x00]
276; EGPR-NEXT:    movb $8, -{{[0-9]+}}(%rsp) # encoding: [0xc6,0x44,0x24,0xb6,0x08]
277; EGPR-NEXT:    movw $8, -{{[0-9]+}}(%rsp) # encoding: [0x66,0xc7,0x44,0x24,0x9c,0x08,0x00]
278; EGPR-NEXT:    movb $8, -{{[0-9]+}}(%rsp) # encoding: [0xc6,0x44,0x24,0xb7,0x08]
279; EGPR-NEXT:    movw $8, -{{[0-9]+}}(%rsp) # encoding: [0x66,0xc7,0x44,0x24,0x9e,0x08,0x00]
280; EGPR-NEXT:    ldtilecfg -{{[0-9]+}}(%rsp) # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x78,0x49,0x44,0x24,0x80]
281; EGPR-NEXT:    movw $8, %ax # encoding: [0x66,0xb8,0x08,0x00]
282; EGPR-NEXT:    tileloadd (%rsi,%rdx), %tmm0 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7b,0x4b,0x04,0x16]
283; EGPR-NEXT:    t2rpntlvwz0 (%rsi,%rdx), %tmm4 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x78,0x6e,0x24,0x16]
284; EGPR-NEXT:    t2rpntlvwz0t1 (%rsi,%rdx), %tmm6 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x78,0x6f,0x34,0x16]
285; EGPR-NEXT:    movabsq $64, %rcx # encoding: [0x48,0xb9,0x40,0x00,0x00,0x00,0x00,0x00,0x00,0x00]
286; EGPR-NEXT:    tilestored %tmm6, 4032(%rsp,%rcx) # 1024-byte Folded Spill
287; EGPR-NEXT:    # encoding: [0xc4,0xe2,0x7a,0x4b,0xb4,0x0c,0xc0,0x0f,0x00,0x00]
288; EGPR-NEXT:    tilestored %tmm7, 5056(%rsp,%rcx) # 1024-byte Folded Spill
289; EGPR-NEXT:    # encoding: [0xc4,0xe2,0x7a,0x4b,0xbc,0x0c,0xc0,0x13,0x00,0x00]
290; EGPR-NEXT:    t2rpntlvwz1 (%rsi,%rdx), %tmm6 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0x6e,0x34,0x16]
291; EGPR-NEXT:    tilestored %tmm6, 1984(%rsp,%rcx) # 1024-byte Folded Spill
292; EGPR-NEXT:    # encoding: [0xc4,0xe2,0x7a,0x4b,0xb4,0x0c,0xc0,0x07,0x00,0x00]
293; EGPR-NEXT:    tilestored %tmm7, 3008(%rsp,%rcx) # 1024-byte Folded Spill
294; EGPR-NEXT:    # encoding: [0xc4,0xe2,0x7a,0x4b,0xbc,0x0c,0xc0,0x0b,0x00,0x00]
295; EGPR-NEXT:    t2rpntlvwz1t1 (%rsi,%rdx), %tmm6 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0x6f,0x34,0x16]
296; EGPR-NEXT:    tilestored %tmm6, -64(%rsp,%rcx) # 1024-byte Folded Spill
297; EGPR-NEXT:    # encoding: [0xc4,0xe2,0x7a,0x4b,0x74,0x0c,0xc0]
298; EGPR-NEXT:    tilestored %tmm7, 960(%rsp,%rcx) # 1024-byte Folded Spill
299; EGPR-NEXT:    # encoding: [0xc4,0xe2,0x7a,0x4b,0xbc,0x0c,0xc0,0x03,0x00,0x00]
300; EGPR-NEXT:    t2rpntlvwz0 (%rsi,%rdx), %tmm6 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x78,0x6e,0x34,0x16]
301; EGPR-NEXT:    tilestored %tmm4, (%rsi,%rdx) # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7a,0x4b,0x24,0x16]
302; EGPR-NEXT:    tilestored %tmm5, (%rsi,%rdx) # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7a,0x4b,0x2c,0x16]
303; EGPR-NEXT:    tileloadd 4032(%rsp,%rcx), %tmm4 # 1024-byte Folded Reload
304; EGPR-NEXT:    # encoding: [0xc4,0xe2,0x7b,0x4b,0xa4,0x0c,0xc0,0x0f,0x00,0x00]
305; EGPR-NEXT:    tileloadd 5056(%rsp,%rcx), %tmm5 # 1024-byte Folded Reload
306; EGPR-NEXT:    # encoding: [0xc4,0xe2,0x7b,0x4b,0xac,0x0c,0xc0,0x13,0x00,0x00]
307; EGPR-NEXT:    tilestored %tmm4, (%rsi,%rdx) # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7a,0x4b,0x24,0x16]
308; EGPR-NEXT:    tilestored %tmm5, (%rsi,%rdx) # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7a,0x4b,0x2c,0x16]
309; EGPR-NEXT:    tileloadd 1984(%rsp,%rcx), %tmm4 # 1024-byte Folded Reload
310; EGPR-NEXT:    # encoding: [0xc4,0xe2,0x7b,0x4b,0xa4,0x0c,0xc0,0x07,0x00,0x00]
311; EGPR-NEXT:    tileloadd 3008(%rsp,%rcx), %tmm5 # 1024-byte Folded Reload
312; EGPR-NEXT:    # encoding: [0xc4,0xe2,0x7b,0x4b,0xac,0x0c,0xc0,0x0b,0x00,0x00]
313; EGPR-NEXT:    tilestored %tmm4, (%rsi,%rdx) # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7a,0x4b,0x24,0x16]
314; EGPR-NEXT:    tilestored %tmm5, (%rsi,%rdx) # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7a,0x4b,0x2c,0x16]
315; EGPR-NEXT:    tileloadd -64(%rsp,%rcx), %tmm4 # 1024-byte Folded Reload
316; EGPR-NEXT:    # encoding: [0xc4,0xe2,0x7b,0x4b,0x64,0x0c,0xc0]
317; EGPR-NEXT:    tileloadd 960(%rsp,%rcx), %tmm5 # 1024-byte Folded Reload
318; EGPR-NEXT:    # encoding: [0xc4,0xe2,0x7b,0x4b,0xac,0x0c,0xc0,0x03,0x00,0x00]
319; EGPR-NEXT:    tilestored %tmm4, (%rsi,%rdx) # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7a,0x4b,0x24,0x16]
320; EGPR-NEXT:    tilestored %tmm5, (%rsi,%rdx) # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7a,0x4b,0x2c,0x16]
321; EGPR-NEXT:    tilestored %tmm6, (%rsi,%rdx) # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7a,0x4b,0x34,0x16]
322; EGPR-NEXT:    tilestored %tmm7, (%rsi,%rdx) # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7a,0x4b,0x3c,0x16]
323; EGPR-NEXT:    addq $6088, %rsp # encoding: [0x48,0x81,0xc4,0xc8,0x17,0x00,0x00]
324; EGPR-NEXT:    # imm = 0x17C8
325; EGPR-NEXT:    tilerelease # encoding: [0xc4,0xe2,0x78,0x49,0xc0]
326; EGPR-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
327; EGPR-NEXT:    retq # encoding: [0xc3]
328  %a = call x86_amx @llvm.x86.tileloadd64.internal(i16 8, i16 8, i8* %base, i64 %stride)
329  %b1 = call { x86_amx, x86_amx } @llvm.x86.t2rpntlvwz0.internal(i16 8, i16 8, i16 8, i8* %base, i64 %stride)
330  %b2 = call { x86_amx, x86_amx } @llvm.x86.t2rpntlvwz0t1.internal(i16 8, i16 8, i16 8, i8* %base, i64 %stride)
331  %b3 = call { x86_amx, x86_amx } @llvm.x86.t2rpntlvwz1.internal(i16 8, i16 8, i16 8, i8* %base, i64 %stride)
332  %b4 = call { x86_amx, x86_amx } @llvm.x86.t2rpntlvwz1t1.internal(i16 8, i16 8, i16 8, i8* %base, i64 %stride)
333  %b5 = call { x86_amx, x86_amx } @llvm.x86.t2rpntlvwz0.internal(i16 8, i16 8, i16 8, i8* %base, i64 %stride)
334  %e11 = extractvalue { x86_amx, x86_amx } %b1, 0
335  %e12 = extractvalue { x86_amx, x86_amx } %b1, 1
336  %e21 = extractvalue { x86_amx, x86_amx } %b2, 0
337  %e22 = extractvalue { x86_amx, x86_amx } %b2, 1
338  %e31 = extractvalue { x86_amx, x86_amx } %b3, 0
339  %e32 = extractvalue { x86_amx, x86_amx } %b3, 1
340  %e41 = extractvalue { x86_amx, x86_amx } %b4, 0
341  %e42 = extractvalue { x86_amx, x86_amx } %b4, 1
342  %e51 = extractvalue { x86_amx, x86_amx } %b5, 0
343  %e52 = extractvalue { x86_amx, x86_amx } %b5, 1
344  call void @llvm.x86.tilestored64.internal(i16 8, i16 8, i8* %base, i64 %stride, x86_amx %e11)
345  call void @llvm.x86.tilestored64.internal(i16 8, i16 8, i8* %base, i64 %stride, x86_amx %e12)
346  call void @llvm.x86.tilestored64.internal(i16 8, i16 8, i8* %base, i64 %stride, x86_amx %e21)
347  call void @llvm.x86.tilestored64.internal(i16 8, i16 8, i8* %base, i64 %stride, x86_amx %e22)
348  call void @llvm.x86.tilestored64.internal(i16 8, i16 8, i8* %base, i64 %stride, x86_amx %e31)
349  call void @llvm.x86.tilestored64.internal(i16 8, i16 8, i8* %base, i64 %stride, x86_amx %e32)
350  call void @llvm.x86.tilestored64.internal(i16 8, i16 8, i8* %base, i64 %stride, x86_amx %e41)
351  call void @llvm.x86.tilestored64.internal(i16 8, i16 8, i8* %base, i64 %stride, x86_amx %e42)
352  call void @llvm.x86.tilestored64.internal(i16 8, i16 8, i8* %base, i64 %stride, x86_amx %e51)
353  call void @llvm.x86.tilestored64.internal(i16 8, i16 8, i8* %base, i64 %stride, x86_amx %e52)
354  ret void
355}
356
357declare x86_amx @llvm.x86.tileloadd64.internal(i16, i16, i8*, i64)
358declare void @llvm.x86.tilestored64.internal(i16, i16, i8*, i64, x86_amx)
359declare { x86_amx, x86_amx } @llvm.x86.t2rpntlvwz0.internal(i16, i16, i16, i8*, i64)
360declare { x86_amx, x86_amx } @llvm.x86.t2rpntlvwz0t1.internal(i16, i16, i16, i8*, i64)
361declare { x86_amx, x86_amx } @llvm.x86.t2rpntlvwz1.internal(i16, i16, i16, i8*, i64)
362declare { x86_amx, x86_amx } @llvm.x86.t2rpntlvwz1t1.internal(i16, i16, i16, i8*, i64)
363declare x86_amx @llvm.x86.ttransposed.internal(i16, i16, x86_amx)
364declare x86_amx @llvm.x86.ttdpbf16ps.internal(i16, i16, i16, x86_amx, x86_amx, x86_amx)
365declare x86_amx @llvm.x86.ttdpfp16ps.internal(i16, i16, i16, x86_amx, x86_amx, x86_amx)
366declare x86_amx @llvm.x86.ttcmmimfp16ps.internal(i16, i16, i16, x86_amx, x86_amx, x86_amx)
367declare x86_amx @llvm.x86.ttcmmrlfp16ps.internal(i16, i16, i16, x86_amx, x86_amx, x86_amx)
368declare x86_amx @llvm.x86.tconjtcmmimfp16ps.internal(i16, i16, i16, x86_amx, x86_amx, x86_amx)
369declare x86_amx @llvm.x86.tconjtfp16.internal(i16, i16, x86_amx)
370
371attributes #0 = { nounwind }
372