xref: /llvm-project/llvm/test/CodeGen/X86/AMX/amx-across-func.ll (revision e6bf48d11047e970cb24554a01b65b566d6b5d22) 
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; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+amx-int8 -mattr=+avx512f -verify-machineinstrs -enable-ipra | FileCheck -check-prefix=IPRA %s
4; RUN: llc < %s -O0 -mtriple=x86_64-unknown-unknown -mattr=+amx-int8 -mattr=+avx512f -verify-machineinstrs | FileCheck -check-prefix=O0 %s
5
6@buf = dso_local global [3072 x i8] zeroinitializer, align 64
7
8define internal void @foo() {
9; CHECK-LABEL: foo:
10; CHECK:       # %bb.0: # %entry
11; CHECK-NEXT:    retq
12;
13; IPRA-LABEL: foo:
14; IPRA:       # %bb.0: # %entry
15; IPRA-NEXT:    retq
16;
17; O0-LABEL: foo:
18; O0:       # %bb.0: # %entry
19; O0-NEXT:    retq
20entry:
21  ret void
22}
23
24define dso_local void @test_api(i16 signext %0, i16 signext %1) nounwind {
25; CHECK-LABEL: test_api:
26; CHECK:       # %bb.0:
27; CHECK-NEXT:    pushq %rbp
28; CHECK-NEXT:    pushq %r15
29; CHECK-NEXT:    pushq %r14
30; CHECK-NEXT:    pushq %rbx
31; CHECK-NEXT:    subq $2120, %rsp # imm = 0x848
32; CHECK-NEXT:    movl %esi, %ebx
33; CHECK-NEXT:    movl %edi, %ebp
34; CHECK-NEXT:    vxorps %xmm0, %xmm0, %xmm0
35; CHECK-NEXT:    vmovups %zmm0, (%rsp)
36; CHECK-NEXT:    movb $1, (%rsp)
37; CHECK-NEXT:    movw $8, {{[0-9]+}}(%rsp)
38; CHECK-NEXT:    movb $8, {{[0-9]+}}(%rsp)
39; CHECK-NEXT:    movw %bx, {{[0-9]+}}(%rsp)
40; CHECK-NEXT:    movb %bpl, {{[0-9]+}}(%rsp)
41; CHECK-NEXT:    movw %bx, {{[0-9]+}}(%rsp)
42; CHECK-NEXT:    movb %bpl, {{[0-9]+}}(%rsp)
43; CHECK-NEXT:    ldtilecfg (%rsp)
44; CHECK-NEXT:    movl $buf, %eax
45; CHECK-NEXT:    movl $32, %r14d
46; CHECK-NEXT:    movw $8, %r15w
47; CHECK-NEXT:    tileloadd (%rax,%r14), %tmm1
48; CHECK-NEXT:    movabsq $64, %rax
49; CHECK-NEXT:    tilestored %tmm1, 1088(%rsp,%rax) # 1024-byte Folded Spill
50; CHECK-NEXT:    movl $buf+1024, %eax
51; CHECK-NEXT:    tileloadd (%rax,%r14), %tmm2
52; CHECK-NEXT:    movabsq $64, %rax
53; CHECK-NEXT:    tilestored %tmm2, 64(%rsp,%rax) # 1024-byte Folded Spill
54; CHECK-NEXT:    vzeroupper
55; CHECK-NEXT:    callq foo
56; CHECK-NEXT:    ldtilecfg (%rsp)
57; CHECK-NEXT:    movl $buf+2048, %eax
58; CHECK-NEXT:    tileloadd (%rax,%r14), %tmm0
59; CHECK-NEXT:    movabsq $64, %rcx
60; CHECK-NEXT:    tileloadd 1088(%rsp,%rcx), %tmm1 # 1024-byte Folded Reload
61; CHECK-NEXT:    tileloadd 64(%rsp,%rcx), %tmm2 # 1024-byte Folded Reload
62; CHECK-NEXT:    tdpbssd %tmm2, %tmm1, %tmm0
63; CHECK-NEXT:    tilestored %tmm0, (%rax,%r14)
64; CHECK-NEXT:    addq $2120, %rsp # imm = 0x848
65; CHECK-NEXT:    popq %rbx
66; CHECK-NEXT:    popq %r14
67; CHECK-NEXT:    popq %r15
68; CHECK-NEXT:    popq %rbp
69; CHECK-NEXT:    tilerelease
70; CHECK-NEXT:    retq
71;
72; IPRA-LABEL: test_api:
73; IPRA:       # %bb.0:
74; IPRA-NEXT:    subq $72, %rsp
75; IPRA-NEXT:    vxorps %xmm0, %xmm0, %xmm0
76; IPRA-NEXT:    vmovups %zmm0, {{[0-9]+}}(%rsp)
77; IPRA-NEXT:    movb $1, {{[0-9]+}}(%rsp)
78; IPRA-NEXT:    movw $8, {{[0-9]+}}(%rsp)
79; IPRA-NEXT:    movb $8, {{[0-9]+}}(%rsp)
80; IPRA-NEXT:    movw %si, {{[0-9]+}}(%rsp)
81; IPRA-NEXT:    movb %dil, {{[0-9]+}}(%rsp)
82; IPRA-NEXT:    movw %si, {{[0-9]+}}(%rsp)
83; IPRA-NEXT:    movb %dil, {{[0-9]+}}(%rsp)
84; IPRA-NEXT:    ldtilecfg {{[0-9]+}}(%rsp)
85; IPRA-NEXT:    movl $buf, %eax
86; IPRA-NEXT:    movl $32, %ecx
87; IPRA-NEXT:    movw $8, %dx
88; IPRA-NEXT:    tileloadd (%rax,%rcx), %tmm0
89; IPRA-NEXT:    movl $buf+1024, %eax
90; IPRA-NEXT:    tileloadd (%rax,%rcx), %tmm1
91; IPRA-NEXT:    callq foo
92; IPRA-NEXT:    movl $buf+2048, %eax
93; IPRA-NEXT:    tileloadd (%rax,%rcx), %tmm2
94; IPRA-NEXT:    tdpbssd %tmm1, %tmm0, %tmm2
95; IPRA-NEXT:    tilestored %tmm2, (%rax,%rcx)
96; IPRA-NEXT:    addq $72, %rsp
97; IPRA-NEXT:    tilerelease
98; IPRA-NEXT:    vzeroupper
99; IPRA-NEXT:    retq
100;
101; O0-LABEL: test_api:
102; O0:       # %bb.0:
103; O0-NEXT:    pushq %rbp
104; O0-NEXT:    movq %rsp, %rbp
105; O0-NEXT:    andq $-1024, %rsp # imm = 0xFC00
106; O0-NEXT:    subq $8192, %rsp # imm = 0x2000
107; O0-NEXT:    vxorps %xmm0, %xmm0, %xmm0
108; O0-NEXT:    vmovups %zmm0, {{[0-9]+}}(%rsp)
109; O0-NEXT:    movb $1, {{[0-9]+}}(%rsp)
110; O0-NEXT:    movw %si, %cx
111; O0-NEXT:    movw %cx, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
112; O0-NEXT:    movw %di, %ax
113; O0-NEXT:    movw %ax, {{[-0-9]+}}(%r{{[sb]}}p) # 2-byte Spill
114; O0-NEXT:    movl $buf, %esi
115; O0-NEXT:    movl $32, %edi
116; O0-NEXT:    movw $8, %dx
117; O0-NEXT:    # implicit-def: $al
118; O0-NEXT:    movb %al, {{[0-9]+}}(%rsp)
119; O0-NEXT:    movw %dx, {{[0-9]+}}(%rsp)
120; O0-NEXT:    ldtilecfg {{[0-9]+}}(%rsp)
121; O0-NEXT:    tileloadd (%rsi,%rdi), %tmm0
122; O0-NEXT:    movl $64, %edi
123; O0-NEXT:    leaq {{[0-9]+}}(%rsp), %rsi
124; O0-NEXT:    movw $8, %dx
125; O0-NEXT:    tilestored %tmm0, (%rsi,%rdi)
126; O0-NEXT:    movl $32, %esi
127; O0-NEXT:    movl $buf+1024, %edx
128; O0-NEXT:    movw $8, %ax
129; O0-NEXT:    # implicit-def: $al
130; O0-NEXT:    movb %al, {{[0-9]+}}(%rsp)
131; O0-NEXT:    movw %cx, {{[0-9]+}}(%rsp)
132; O0-NEXT:    ldtilecfg {{[0-9]+}}(%rsp)
133; O0-NEXT:    tileloadd (%rdx,%rsi), %tmm0
134; O0-NEXT:    movl $64, %esi
135; O0-NEXT:    leaq {{[0-9]+}}(%rsp), %rdx
136; O0-NEXT:    movw $8, %ax
137; O0-NEXT:    tilestored %tmm0, (%rdx,%rsi)
138; O0-NEXT:    vzeroupper
139; O0-NEXT:    callq foo
140; O0-NEXT:    movw {{[-0-9]+}}(%r{{[sb]}}p), %cx # 2-byte Reload
141; O0-NEXT:    movw {{[-0-9]+}}(%r{{[sb]}}p), %ax # 2-byte Reload
142; O0-NEXT:    # implicit-def: $al
143; O0-NEXT:    movb %al, {{[0-9]+}}(%rsp)
144; O0-NEXT:    movw %cx, {{[0-9]+}}(%rsp)
145; O0-NEXT:    ldtilecfg {{[0-9]+}}(%rsp)
146; O0-NEXT:    movl $32, %esi
147; O0-NEXT:    movl $buf+2048, %edx
148; O0-NEXT:    tileloadd (%rdx,%rsi), %tmm0
149; O0-NEXT:    movl $64, %esi
150; O0-NEXT:    leaq {{[0-9]+}}(%rsp), %rdx
151; O0-NEXT:    tilestored %tmm0, (%rdx,%rsi)
152; O0-NEXT:    movl $64, %edi
153; O0-NEXT:    leaq {{[0-9]+}}(%rsp), %rdx
154; O0-NEXT:    movw $8, %si
155; O0-NEXT:    # implicit-def: $al
156; O0-NEXT:    movb %al, {{[0-9]+}}(%rsp)
157; O0-NEXT:    movw %si, {{[0-9]+}}(%rsp)
158; O0-NEXT:    ldtilecfg {{[0-9]+}}(%rsp)
159; O0-NEXT:    tileloadd (%rdx,%rdi), %tmm0
160; O0-NEXT:    movabsq $64, %rdx
161; O0-NEXT:    tilestored %tmm0, 1024(%rsp,%rdx) # 1024-byte Folded Spill
162; O0-NEXT:    movl $64, %r8d
163; O0-NEXT:    leaq {{[0-9]+}}(%rsp), %rdi
164; O0-NEXT:    movw $8, %dx
165; O0-NEXT:    # implicit-def: $al
166; O0-NEXT:    movb %al, {{[0-9]+}}(%rsp)
167; O0-NEXT:    movw %cx, {{[0-9]+}}(%rsp)
168; O0-NEXT:    # implicit-def: $al
169; O0-NEXT:    movb %al, {{[0-9]+}}(%rsp)
170; O0-NEXT:    movw %cx, {{[0-9]+}}(%rsp)
171; O0-NEXT:    # implicit-def: $al
172; O0-NEXT:    movb %al, {{[0-9]+}}(%rsp)
173; O0-NEXT:    movw %si, {{[0-9]+}}(%rsp)
174; O0-NEXT:    # implicit-def: $al
175; O0-NEXT:    movb %al, {{[0-9]+}}(%rsp)
176; O0-NEXT:    movw %cx, {{[0-9]+}}(%rsp)
177; O0-NEXT:    # implicit-def: $dl
178; O0-NEXT:    movb %dl, {{[0-9]+}}(%rsp)
179; O0-NEXT:    movw %cx, {{[0-9]+}}(%rsp)
180; O0-NEXT:    ldtilecfg {{[0-9]+}}(%rsp)
181; O0-NEXT:    tileloadd (%rdi,%r8), %tmm2
182; O0-NEXT:    movl $64, %edi
183; O0-NEXT:    leaq {{[0-9]+}}(%rsp), %rdx
184; O0-NEXT:    tileloadd (%rdx,%rdi), %tmm0
185; O0-NEXT:    movw $8, %dx
186; O0-NEXT:    movabsq $64, %rdi
187; O0-NEXT:    tileloadd 1024(%rsp,%rdi), %tmm1 # 1024-byte Folded Reload
188; O0-NEXT:    tdpbssd %tmm2, %tmm1, %tmm0
189; O0-NEXT:    movl $64, %esi
190; O0-NEXT:    leaq {{[0-9]+}}(%rsp), %rdx
191; O0-NEXT:    tilestored %tmm0, (%rdx,%rsi)
192; O0-NEXT:    movl $64, %esi
193; O0-NEXT:    leaq {{[0-9]+}}(%rsp), %rdx
194; O0-NEXT:    tileloadd (%rdx,%rsi), %tmm0
195; O0-NEXT:    movl $32, %esi
196; O0-NEXT:    movl $buf+2048, %edx
197; O0-NEXT:    tilestored %tmm0, (%rdx,%rsi)
198; O0-NEXT:    movq %rbp, %rsp
199; O0-NEXT:    popq %rbp
200; O0-NEXT:    tilerelease
201; O0-NEXT:    retq
202  %3 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 %0, i16 8, ptr @buf, i64 32)
203  %4 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 8, i16 %1, ptr getelementptr inbounds ([3072 x i8], ptr @buf, i64 0, i64 1024), i64 32)
204  call void @foo()
205  %5 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 %0, i16 %1, ptr getelementptr inbounds ([3072 x i8], ptr @buf, i64 0, i64 2048), i64 32)
206  %6 = tail call x86_amx @llvm.x86.tdpbssd.internal(i16 %0, i16 %1, i16 8, x86_amx %5, x86_amx %3, x86_amx %4)
207  tail call void @llvm.x86.tilestored64.internal(i16 %0, i16 %1, ptr getelementptr inbounds ([3072 x i8], ptr @buf, i64 0, i64 2048), i64 32, x86_amx %6)
208  ret void
209}
210
211define dso_local i32 @test_loop(i32 %0) nounwind {
212; CHECK-LABEL: test_loop:
213; CHECK:       # %bb.0:
214; CHECK-NEXT:    pushq %rbp
215; CHECK-NEXT:    pushq %r15
216; CHECK-NEXT:    pushq %r14
217; CHECK-NEXT:    pushq %r13
218; CHECK-NEXT:    pushq %r12
219; CHECK-NEXT:    pushq %rbx
220; CHECK-NEXT:    subq $1096, %rsp # imm = 0x448
221; CHECK-NEXT:    movl %edi, %ebx
222; CHECK-NEXT:    vxorps %xmm0, %xmm0, %xmm0
223; CHECK-NEXT:    vmovups %zmm0, (%rsp)
224; CHECK-NEXT:    movb $1, (%rsp)
225; CHECK-NEXT:    movb $8, {{[0-9]+}}(%rsp)
226; CHECK-NEXT:    movw $8, {{[0-9]+}}(%rsp)
227; CHECK-NEXT:    vzeroupper
228; CHECK-NEXT:    callq foo
229; CHECK-NEXT:    ldtilecfg (%rsp)
230; CHECK-NEXT:    testl %ebx, %ebx
231; CHECK-NEXT:    jg .LBB2_4
232; CHECK-NEXT:  # %bb.1: # %.preheader
233; CHECK-NEXT:    movl $7, %ebp
234; CHECK-NEXT:    movl $buf, %r14d
235; CHECK-NEXT:    movl $32, %r15d
236; CHECK-NEXT:    movw $8, %r12w
237; CHECK-NEXT:    movl $buf+2048, %r13d
238; CHECK-NEXT:    .p2align 4
239; CHECK-NEXT:  .LBB2_2: # =>This Inner Loop Header: Depth=1
240; CHECK-NEXT:    tileloadd (%r14,%r15), %tmm0
241; CHECK-NEXT:    movabsq $64, %rax
242; CHECK-NEXT:    tilestored %tmm0, 64(%rsp,%rax) # 1024-byte Folded Spill
243; CHECK-NEXT:    callq foo
244; CHECK-NEXT:    ldtilecfg (%rsp)
245; CHECK-NEXT:    movabsq $64, %rax
246; CHECK-NEXT:    tileloadd 64(%rsp,%rax), %tmm0 # 1024-byte Folded Reload
247; CHECK-NEXT:    tilestored %tmm0, (%r13,%r15)
248; CHECK-NEXT:    callq foo
249; CHECK-NEXT:    ldtilecfg (%rsp)
250; CHECK-NEXT:    decl %ebp
251; CHECK-NEXT:    cmpl $7, %ebp
252; CHECK-NEXT:    jne .LBB2_2
253; CHECK-NEXT:  # %bb.3:
254; CHECK-NEXT:    cmpl $3, %ebx
255; CHECK-NEXT:    jne .LBB2_4
256; CHECK-NEXT:  # %bb.6:
257; CHECK-NEXT:    testl %ebp, %ebp
258; CHECK-NEXT:    jne .LBB2_5
259; CHECK-NEXT:  # %bb.7:
260; CHECK-NEXT:    incl %ebx
261; CHECK-NEXT:    jmp .LBB2_8
262; CHECK-NEXT:  .LBB2_4:
263; CHECK-NEXT:    callq foo
264; CHECK-NEXT:    ldtilecfg (%rsp)
265; CHECK-NEXT:    movl $32, %eax
266; CHECK-NEXT:    movl $buf+1024, %ecx
267; CHECK-NEXT:    movw $8, %dx
268; CHECK-NEXT:    tileloadd (%rcx,%rax), %tmm0
269; CHECK-NEXT:    tilestored %tmm0, (%rcx,%rax)
270; CHECK-NEXT:  .LBB2_5:
271; CHECK-NEXT:    decl %ebx
272; CHECK-NEXT:  .LBB2_8:
273; CHECK-NEXT:    movl %ebx, %eax
274; CHECK-NEXT:    addq $1096, %rsp # imm = 0x448
275; CHECK-NEXT:    popq %rbx
276; CHECK-NEXT:    popq %r12
277; CHECK-NEXT:    popq %r13
278; CHECK-NEXT:    popq %r14
279; CHECK-NEXT:    popq %r15
280; CHECK-NEXT:    popq %rbp
281; CHECK-NEXT:    tilerelease
282; CHECK-NEXT:    retq
283;
284; IPRA-LABEL: test_loop:
285; IPRA:       # %bb.0:
286; IPRA-NEXT:    subq $72, %rsp
287; IPRA-NEXT:    movl %edi, %eax
288; IPRA-NEXT:    vxorps %xmm0, %xmm0, %xmm0
289; IPRA-NEXT:    vmovups %zmm0, {{[0-9]+}}(%rsp)
290; IPRA-NEXT:    movb $1, {{[0-9]+}}(%rsp)
291; IPRA-NEXT:    movb $8, {{[0-9]+}}(%rsp)
292; IPRA-NEXT:    movw $8, {{[0-9]+}}(%rsp)
293; IPRA-NEXT:    ldtilecfg {{[0-9]+}}(%rsp)
294; IPRA-NEXT:    callq foo
295; IPRA-NEXT:    testl %edi, %edi
296; IPRA-NEXT:    jg .LBB2_4
297; IPRA-NEXT:  # %bb.1: # %.preheader
298; IPRA-NEXT:    movl $7, %ecx
299; IPRA-NEXT:    movl $buf, %edx
300; IPRA-NEXT:    movl $32, %esi
301; IPRA-NEXT:    movw $8, %di
302; IPRA-NEXT:    movl $buf+2048, %r8d
303; IPRA-NEXT:    .p2align 4
304; IPRA-NEXT:  .LBB2_2: # =>This Inner Loop Header: Depth=1
305; IPRA-NEXT:    tileloadd (%rdx,%rsi), %tmm0
306; IPRA-NEXT:    callq foo
307; IPRA-NEXT:    tilestored %tmm0, (%r8,%rsi)
308; IPRA-NEXT:    callq foo
309; IPRA-NEXT:    decl %ecx
310; IPRA-NEXT:    cmpl $7, %ecx
311; IPRA-NEXT:    jne .LBB2_2
312; IPRA-NEXT:  # %bb.3:
313; IPRA-NEXT:    cmpl $3, %eax
314; IPRA-NEXT:    jne .LBB2_4
315; IPRA-NEXT:  # %bb.6:
316; IPRA-NEXT:    testl %ecx, %ecx
317; IPRA-NEXT:    jne .LBB2_5
318; IPRA-NEXT:  # %bb.7:
319; IPRA-NEXT:    incl %eax
320; IPRA-NEXT:    jmp .LBB2_8
321; IPRA-NEXT:  .LBB2_4:
322; IPRA-NEXT:    callq foo
323; IPRA-NEXT:    movl $32, %ecx
324; IPRA-NEXT:    movl $buf+1024, %edx
325; IPRA-NEXT:    movw $8, %si
326; IPRA-NEXT:    tileloadd (%rdx,%rcx), %tmm0
327; IPRA-NEXT:    tilestored %tmm0, (%rdx,%rcx)
328; IPRA-NEXT:  .LBB2_5:
329; IPRA-NEXT:    decl %eax
330; IPRA-NEXT:  .LBB2_8:
331; IPRA-NEXT:    addq $72, %rsp
332; IPRA-NEXT:    tilerelease
333; IPRA-NEXT:    vzeroupper
334; IPRA-NEXT:    retq
335;
336; O0-LABEL: test_loop:
337; O0:       # %bb.0:
338; O0-NEXT:    pushq %rbp
339; O0-NEXT:    movq %rsp, %rbp
340; O0-NEXT:    andq $-1024, %rsp # imm = 0xFC00
341; O0-NEXT:    subq $4096, %rsp # imm = 0x1000
342; O0-NEXT:    vxorps %xmm0, %xmm0, %xmm0
343; O0-NEXT:    vmovups %zmm0, {{[0-9]+}}(%rsp)
344; O0-NEXT:    movb $1, {{[0-9]+}}(%rsp)
345; O0-NEXT:    movl %edi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
346; O0-NEXT:    vzeroupper
347; O0-NEXT:    callq foo
348; O0-NEXT:  # %bb.1:
349; O0-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %ecx # 4-byte Reload
350; O0-NEXT:    xorl %eax, %eax
351; O0-NEXT:    cmpl $0, %ecx
352; O0-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
353; O0-NEXT:    jg .LBB2_4
354; O0-NEXT:    jmp .LBB2_3
355; O0-NEXT:  .LBB2_2:
356; O0-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 4-byte Reload
357; O0-NEXT:    cmpl $3, %eax
358; O0-NEXT:    je .LBB2_5
359; O0-NEXT:    jmp .LBB2_4
360; O0-NEXT:  .LBB2_3: # =>This Inner Loop Header: Depth=1
361; O0-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 4-byte Reload
362; O0-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
363; O0-NEXT:    movl $buf, %ecx
364; O0-NEXT:    movl $32, %edx
365; O0-NEXT:    movw $8, %ax
366; O0-NEXT:    # implicit-def: $al
367; O0-NEXT:    movb %al, {{[0-9]+}}(%rsp)
368; O0-NEXT:    movw %ax, {{[0-9]+}}(%rsp)
369; O0-NEXT:    ldtilecfg {{[0-9]+}}(%rsp)
370; O0-NEXT:    tileloadd (%rcx,%rdx), %tmm0
371; O0-NEXT:    movl $64, %edx
372; O0-NEXT:    leaq {{[0-9]+}}(%rsp), %rcx
373; O0-NEXT:    movw $8, %ax
374; O0-NEXT:    tilestored %tmm0, (%rcx,%rdx)
375; O0-NEXT:    callq foo
376; O0-NEXT:    movl $64, %edx
377; O0-NEXT:    leaq {{[0-9]+}}(%rsp), %rcx
378; O0-NEXT:    movw $8, %ax
379; O0-NEXT:    # implicit-def: $al
380; O0-NEXT:    movb %al, {{[0-9]+}}(%rsp)
381; O0-NEXT:    movw %ax, {{[0-9]+}}(%rsp)
382; O0-NEXT:    ldtilecfg {{[0-9]+}}(%rsp)
383; O0-NEXT:    tileloadd (%rcx,%rdx), %tmm0
384; O0-NEXT:    movl $32, %edx
385; O0-NEXT:    movl $buf+2048, %ecx
386; O0-NEXT:    movw $8, %ax
387; O0-NEXT:    tilestored %tmm0, (%rcx,%rdx)
388; O0-NEXT:    callq foo
389; O0-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 4-byte Reload
390; O0-NEXT:    addl $1, %eax
391; O0-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
392; O0-NEXT:    cmpl $0, %eax
393; O0-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
394; O0-NEXT:    je .LBB2_2
395; O0-NEXT:    jmp .LBB2_3
396; O0-NEXT:  .LBB2_4:
397; O0-NEXT:    callq foo
398; O0-NEXT:    movl $32, %edx
399; O0-NEXT:    movl $buf+1024, %ecx
400; O0-NEXT:    movw $8, %ax
401; O0-NEXT:    # implicit-def: $al
402; O0-NEXT:    movb %al, {{[0-9]+}}(%rsp)
403; O0-NEXT:    movw %ax, {{[0-9]+}}(%rsp)
404; O0-NEXT:    ldtilecfg {{[0-9]+}}(%rsp)
405; O0-NEXT:    tileloadd (%rcx,%rdx), %tmm0
406; O0-NEXT:    movl $64, %edx
407; O0-NEXT:    leaq {{[0-9]+}}(%rsp), %rcx
408; O0-NEXT:    movw $8, %ax
409; O0-NEXT:    tilestored %tmm0, (%rcx,%rdx)
410; O0-NEXT:    movl $64, %edx
411; O0-NEXT:    leaq {{[0-9]+}}(%rsp), %rcx
412; O0-NEXT:    movw $8, %ax
413; O0-NEXT:    # implicit-def: $al
414; O0-NEXT:    movb %al, {{[0-9]+}}(%rsp)
415; O0-NEXT:    movw %ax, {{[0-9]+}}(%rsp)
416; O0-NEXT:    ldtilecfg {{[0-9]+}}(%rsp)
417; O0-NEXT:    tileloadd (%rcx,%rdx), %tmm0
418; O0-NEXT:    movl $32, %edx
419; O0-NEXT:    movl $buf+1024, %ecx
420; O0-NEXT:    movw $8, %ax
421; O0-NEXT:    tilestored %tmm0, (%rcx,%rdx)
422; O0-NEXT:    jmp .LBB2_7
423; O0-NEXT:  .LBB2_5:
424; O0-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 4-byte Reload
425; O0-NEXT:    cmpl $7, %eax
426; O0-NEXT:    jne .LBB2_7
427; O0-NEXT:  # %bb.6:
428; O0-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 4-byte Reload
429; O0-NEXT:    addl $1, %eax
430; O0-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
431; O0-NEXT:    jmp .LBB2_8
432; O0-NEXT:  .LBB2_7:
433; O0-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 4-byte Reload
434; O0-NEXT:    subl $1, %eax
435; O0-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
436; O0-NEXT:  .LBB2_8:
437; O0-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 4-byte Reload
438; O0-NEXT:    movq %rbp, %rsp
439; O0-NEXT:    popq %rbp
440; O0-NEXT:    tilerelease
441; O0-NEXT:    retq
442  call void @foo()
443  br label %2
4442:
445  %3 = icmp sgt i32 %0, 0
446  br i1 %3, label %11, label %6
4474:
448  %5 = icmp eq i32 %0, 3
449  br i1 %5, label %13, label %11
4506:
451  %7 = phi i32 [ %9, %6 ], [ 0, %2 ]
452  %8 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 8, i16 8, ptr @buf, i64 32)
453  call void @foo()
454  tail call void @llvm.x86.tilestored64.internal(i16 8, i16 8, ptr getelementptr inbounds ([3072 x i8], ptr @buf, i64 0, i64 2048), i64 32, x86_amx %8)
455  call void @foo()
456  %9 = add i32 %7, 1
457  %10 = icmp eq i32 %9, 0
458  br i1 %10, label %4, label %6
45911:
460  call void @foo()
461  %12 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 8, i16 8, ptr getelementptr inbounds ([3072 x i8], ptr @buf, i64 0, i64 1024), i64 32)
462  tail call void @llvm.x86.tilestored64.internal(i16 8, i16 8, ptr getelementptr inbounds ([3072 x i8], ptr @buf, i64 0, i64 1024), i64 32, x86_amx %12)
463  br label %17
46413:
465  %14 = icmp eq i32 %9, 7
466  br i1 %14, label %15, label %17
46715:
468  %16 = add i32 %0, 1
469  br label %19
47017:
471  %18 = sub i32 %0, 1
472  br label %19
47319:
474  %20 = phi i32 [ %16, %15 ], [ %18, %17 ]
475  ret i32 %20
476}
477
478define dso_local void @test_loop2(i32 %0) nounwind {
479; CHECK-LABEL: test_loop2:
480; CHECK:       # %bb.0:
481; CHECK-NEXT:    pushq %rbp
482; CHECK-NEXT:    pushq %r15
483; CHECK-NEXT:    pushq %r14
484; CHECK-NEXT:    pushq %r12
485; CHECK-NEXT:    pushq %rbx
486; CHECK-NEXT:    subq $1088, %rsp # imm = 0x440
487; CHECK-NEXT:    movl %edi, %ebx
488; CHECK-NEXT:    vxorps %xmm0, %xmm0, %xmm0
489; CHECK-NEXT:    vmovups %zmm0, (%rsp)
490; CHECK-NEXT:    movb $1, (%rsp)
491; CHECK-NEXT:    movb $8, {{[0-9]+}}(%rsp)
492; CHECK-NEXT:    movw $8, {{[0-9]+}}(%rsp)
493; CHECK-NEXT:    movl $buf, %r14d
494; CHECK-NEXT:    movl $32, %r15d
495; CHECK-NEXT:    movw $8, %bp
496; CHECK-NEXT:    movl $buf+2048, %r12d
497; CHECK-NEXT:    .p2align 4
498; CHECK-NEXT:  .LBB3_1: # =>This Inner Loop Header: Depth=1
499; CHECK-NEXT:    vzeroupper
500; CHECK-NEXT:    callq foo
501; CHECK-NEXT:    ldtilecfg (%rsp)
502; CHECK-NEXT:    testl %ebx, %ebx
503; CHECK-NEXT:    jle .LBB3_3
504; CHECK-NEXT:  # %bb.2: # in Loop: Header=BB3_1 Depth=1
505; CHECK-NEXT:    tileloadd (%r14,%r15), %tmm0
506; CHECK-NEXT:    movabsq $64, %rax
507; CHECK-NEXT:    tilestored %tmm0, 64(%rsp,%rax) # 1024-byte Folded Spill
508; CHECK-NEXT:    callq foo
509; CHECK-NEXT:    ldtilecfg (%rsp)
510; CHECK-NEXT:    movabsq $64, %rax
511; CHECK-NEXT:    tileloadd 64(%rsp,%rax), %tmm0 # 1024-byte Folded Reload
512; CHECK-NEXT:    tilestored %tmm0, (%r12,%r15)
513; CHECK-NEXT:    callq foo
514; CHECK-NEXT:    jmp .LBB3_1
515; CHECK-NEXT:  .LBB3_3:
516; CHECK-NEXT:    addq $1088, %rsp # imm = 0x440
517; CHECK-NEXT:    popq %rbx
518; CHECK-NEXT:    popq %r12
519; CHECK-NEXT:    popq %r14
520; CHECK-NEXT:    popq %r15
521; CHECK-NEXT:    popq %rbp
522; CHECK-NEXT:    tilerelease
523; CHECK-NEXT:    retq
524;
525; IPRA-LABEL: test_loop2:
526; IPRA:       # %bb.0:
527; IPRA-NEXT:    subq $72, %rsp
528; IPRA-NEXT:    vxorps %xmm0, %xmm0, %xmm0
529; IPRA-NEXT:    vmovups %zmm0, {{[0-9]+}}(%rsp)
530; IPRA-NEXT:    movb $1, {{[0-9]+}}(%rsp)
531; IPRA-NEXT:    movb $8, {{[0-9]+}}(%rsp)
532; IPRA-NEXT:    movw $8, {{[0-9]+}}(%rsp)
533; IPRA-NEXT:    ldtilecfg {{[0-9]+}}(%rsp)
534; IPRA-NEXT:    movl $buf, %eax
535; IPRA-NEXT:    movl $32, %ecx
536; IPRA-NEXT:    movw $8, %dx
537; IPRA-NEXT:    movl $buf+2048, %esi
538; IPRA-NEXT:    .p2align 4
539; IPRA-NEXT:  .LBB3_1: # =>This Inner Loop Header: Depth=1
540; IPRA-NEXT:    callq foo
541; IPRA-NEXT:    testl %edi, %edi
542; IPRA-NEXT:    jle .LBB3_3
543; IPRA-NEXT:  # %bb.2: # in Loop: Header=BB3_1 Depth=1
544; IPRA-NEXT:    tileloadd (%rax,%rcx), %tmm0
545; IPRA-NEXT:    callq foo
546; IPRA-NEXT:    tilestored %tmm0, (%rsi,%rcx)
547; IPRA-NEXT:    callq foo
548; IPRA-NEXT:    jmp .LBB3_1
549; IPRA-NEXT:  .LBB3_3:
550; IPRA-NEXT:    addq $72, %rsp
551; IPRA-NEXT:    tilerelease
552; IPRA-NEXT:    vzeroupper
553; IPRA-NEXT:    retq
554;
555; O0-LABEL: test_loop2:
556; O0:       # %bb.0:
557; O0-NEXT:    pushq %rbp
558; O0-NEXT:    movq %rsp, %rbp
559; O0-NEXT:    andq $-1024, %rsp # imm = 0xFC00
560; O0-NEXT:    subq $3072, %rsp # imm = 0xC00
561; O0-NEXT:    vxorps %xmm0, %xmm0, %xmm0
562; O0-NEXT:    vmovups %zmm0, {{[0-9]+}}(%rsp)
563; O0-NEXT:    movb $1, {{[0-9]+}}(%rsp)
564; O0-NEXT:    movl %edi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
565; O0-NEXT:    xorl %eax, %eax
566; O0-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
567; O0-NEXT:  .LBB3_1: # =>This Inner Loop Header: Depth=1
568; O0-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 4-byte Reload
569; O0-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
570; O0-NEXT:    vzeroupper
571; O0-NEXT:    callq foo
572; O0-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 4-byte Reload
573; O0-NEXT:    cmpl $0, %eax
574; O0-NEXT:    jle .LBB3_3
575; O0-NEXT:  # %bb.2: # in Loop: Header=BB3_1 Depth=1
576; O0-NEXT:    movl $buf, %ecx
577; O0-NEXT:    movl $32, %edx
578; O0-NEXT:    movw $8, %ax
579; O0-NEXT:    # implicit-def: $al
580; O0-NEXT:    movb %al, {{[0-9]+}}(%rsp)
581; O0-NEXT:    movw %ax, {{[0-9]+}}(%rsp)
582; O0-NEXT:    ldtilecfg {{[0-9]+}}(%rsp)
583; O0-NEXT:    tileloadd (%rcx,%rdx), %tmm0
584; O0-NEXT:    movl $64, %edx
585; O0-NEXT:    leaq {{[0-9]+}}(%rsp), %rcx
586; O0-NEXT:    movw $8, %ax
587; O0-NEXT:    tilestored %tmm0, (%rcx,%rdx)
588; O0-NEXT:    callq foo
589; O0-NEXT:    movl $64, %edx
590; O0-NEXT:    leaq {{[0-9]+}}(%rsp), %rcx
591; O0-NEXT:    movw $8, %ax
592; O0-NEXT:    # implicit-def: $al
593; O0-NEXT:    movb %al, {{[0-9]+}}(%rsp)
594; O0-NEXT:    movw %ax, {{[0-9]+}}(%rsp)
595; O0-NEXT:    ldtilecfg {{[0-9]+}}(%rsp)
596; O0-NEXT:    tileloadd (%rcx,%rdx), %tmm0
597; O0-NEXT:    movl $32, %edx
598; O0-NEXT:    movl $buf+2048, %ecx
599; O0-NEXT:    movw $8, %ax
600; O0-NEXT:    tilestored %tmm0, (%rcx,%rdx)
601; O0-NEXT:    callq foo
602; O0-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 4-byte Reload
603; O0-NEXT:    addl $1, %eax
604; O0-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
605; O0-NEXT:    jmp .LBB3_1
606; O0-NEXT:  .LBB3_3:
607; O0-NEXT:    movq %rbp, %rsp
608; O0-NEXT:    popq %rbp
609; O0-NEXT:    tilerelease
610; O0-NEXT:    retq
611  br label %2
6122:
613  %3 = phi i32 [ 0, %1 ], [ %7, %5 ]
614  call void @foo()
615  %4 = icmp sgt i32 %0, 0
616  br i1 %4, label %5, label %8
6175:
618  %6 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 8, i16 8, ptr @buf, i64 32)
619  call void @foo()
620  tail call void @llvm.x86.tilestored64.internal(i16 8, i16 8, ptr getelementptr inbounds ([3072 x i8], ptr @buf, i64 0, i64 2048), i64 32, x86_amx %6)
621  call void @foo()
622  %7 = add i32 %3, 1
623  br label %2
6248:
625  ret void
626}
627
628declare x86_amx @llvm.x86.tileloadd64.internal(i16, i16, ptr, i64)
629declare x86_amx @llvm.x86.tdpbssd.internal(i16, i16, i16, x86_amx, x86_amx, x86_amx)
630declare void @llvm.x86.tilestored64.internal(i16, i16, ptr, i64, x86_amx)
631