194 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+amx-int8 -mattr=+avx512f | FileCheck %s --check-prefix=AVX5123; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+amx-int8 -mattr=+avx2 | FileCheck %s --check-prefix=AVX24; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+amx-int8 | FileCheck %s --check-prefix=SSE25; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+amx-int8 -mattr=+avx512f -O0 | FileCheck %s --check-prefix=AVX512-O06; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+amx-int8 -mattr=+avx2 -O0 | FileCheck %s --check-prefix=AVX2-O07; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+amx-int8 -O0 | FileCheck %s --check-prefix=SSE2-O08 9define void @foo(ptr %buf) nounwind {10; AVX512-LABEL: foo:11; AVX512: # %bb.0: # %entry12; AVX512-NEXT: vxorps %xmm0, %xmm0, %xmm013; AVX512-NEXT: vmovups %zmm0, -{{[0-9]+}}(%rsp)14; AVX512-NEXT: movb $1, -{{[0-9]+}}(%rsp)15; AVX512-NEXT: movb $8, -{{[0-9]+}}(%rsp)16; AVX512-NEXT: movw $32, -{{[0-9]+}}(%rsp)17; AVX512-NEXT: ldtilecfg -{{[0-9]+}}(%rsp)18; AVX512-NEXT: movw $32, %ax19; AVX512-NEXT: movw $8, %cx20; AVX512-NEXT: tilezero %tmm021; AVX512-NEXT: movl $1024, %edx # imm = 0x40022; AVX512-NEXT: tilestored %tmm0, (%rdi,%rdx)23; AVX512-NEXT: tilerelease24; AVX512-NEXT: vzeroupper25; AVX512-NEXT: retq26;27; AVX2-LABEL: foo:28; AVX2: # %bb.0: # %entry29; AVX2-NEXT: vxorps %xmm0, %xmm0, %xmm030; AVX2-NEXT: vmovups %ymm0, -{{[0-9]+}}(%rsp)31; AVX2-NEXT: vmovups %ymm0, -{{[0-9]+}}(%rsp)32; AVX2-NEXT: movb $1, -{{[0-9]+}}(%rsp)33; AVX2-NEXT: movb $8, -{{[0-9]+}}(%rsp)34; AVX2-NEXT: movw $32, -{{[0-9]+}}(%rsp)35; AVX2-NEXT: ldtilecfg -{{[0-9]+}}(%rsp)36; AVX2-NEXT: movw $32, %ax37; AVX2-NEXT: movw $8, %cx38; AVX2-NEXT: tilezero %tmm039; AVX2-NEXT: movl $1024, %edx # imm = 0x40040; AVX2-NEXT: tilestored %tmm0, (%rdi,%rdx)41; AVX2-NEXT: tilerelease42; AVX2-NEXT: vzeroupper43; AVX2-NEXT: retq44;45; SSE2-LABEL: foo:46; SSE2: # %bb.0: # %entry47; SSE2-NEXT: xorps %xmm0, %xmm048; SSE2-NEXT: movups %xmm0, -{{[0-9]+}}(%rsp)49; SSE2-NEXT: movups %xmm0, -{{[0-9]+}}(%rsp)50; SSE2-NEXT: movups %xmm0, -{{[0-9]+}}(%rsp)51; SSE2-NEXT: movups %xmm0, -{{[0-9]+}}(%rsp)52; SSE2-NEXT: movb $1, -{{[0-9]+}}(%rsp)53; SSE2-NEXT: movb $8, -{{[0-9]+}}(%rsp)54; SSE2-NEXT: movw $32, -{{[0-9]+}}(%rsp)55; SSE2-NEXT: ldtilecfg -{{[0-9]+}}(%rsp)56; SSE2-NEXT: movw $32, %ax57; SSE2-NEXT: movw $8, %cx58; SSE2-NEXT: tilezero %tmm059; SSE2-NEXT: movl $1024, %edx # imm = 0x40060; SSE2-NEXT: tilestored %tmm0, (%rdi,%rdx)61; SSE2-NEXT: tilerelease62; SSE2-NEXT: retq63;64; AVX512-O0-LABEL: foo:65; AVX512-O0: # %bb.0: # %entry66; AVX512-O0-NEXT: pushq %rbp67; AVX512-O0-NEXT: movq %rsp, %rbp68; AVX512-O0-NEXT: andq $-1024, %rsp # imm = 0xFC0069; AVX512-O0-NEXT: subq $3072, %rsp # imm = 0xC0070; AVX512-O0-NEXT: vxorps %xmm0, %xmm0, %xmm071; AVX512-O0-NEXT: vmovups %zmm0, {{[0-9]+}}(%rsp)72; AVX512-O0-NEXT: movb $1, {{[0-9]+}}(%rsp)73; AVX512-O0-NEXT: movw $32, %cx74; AVX512-O0-NEXT: movw $8, %ax75; AVX512-O0-NEXT: # implicit-def: $al76; AVX512-O0-NEXT: movb %al, {{[0-9]+}}(%rsp)77; AVX512-O0-NEXT: movw %cx, {{[0-9]+}}(%rsp)78; AVX512-O0-NEXT: ldtilecfg {{[0-9]+}}(%rsp)79; AVX512-O0-NEXT: tilezero %tmm080; AVX512-O0-NEXT: movl $64, %esi81; AVX512-O0-NEXT: leaq {{[0-9]+}}(%rsp), %rdx82; AVX512-O0-NEXT: movw $32, %cx83; AVX512-O0-NEXT: movw $8, %ax84; AVX512-O0-NEXT: tilestored %tmm0, (%rdx,%rsi)85; AVX512-O0-NEXT: movl $64, %esi86; AVX512-O0-NEXT: leaq {{[0-9]+}}(%rsp), %rdx87; AVX512-O0-NEXT: movw $32, %cx88; AVX512-O0-NEXT: movw $8, %ax89; AVX512-O0-NEXT: # implicit-def: $al90; AVX512-O0-NEXT: movb %al, {{[0-9]+}}(%rsp)91; AVX512-O0-NEXT: movw %cx, {{[0-9]+}}(%rsp)92; AVX512-O0-NEXT: ldtilecfg {{[0-9]+}}(%rsp)93; AVX512-O0-NEXT: tileloadd (%rdx,%rsi), %tmm094; AVX512-O0-NEXT: movl $1024, %edx # imm = 0x40095; AVX512-O0-NEXT: movw $32, %cx96; AVX512-O0-NEXT: movw $8, %ax97; AVX512-O0-NEXT: tilestored %tmm0, (%rdi,%rdx)98; AVX512-O0-NEXT: movq %rbp, %rsp99; AVX512-O0-NEXT: popq %rbp100; AVX512-O0-NEXT: tilerelease101; AVX512-O0-NEXT: vzeroupper102; AVX512-O0-NEXT: retq103;104; AVX2-O0-LABEL: foo:105; AVX2-O0: # %bb.0: # %entry106; AVX2-O0-NEXT: pushq %rbp107; AVX2-O0-NEXT: movq %rsp, %rbp108; AVX2-O0-NEXT: andq $-1024, %rsp # imm = 0xFC00109; AVX2-O0-NEXT: subq $3072, %rsp # imm = 0xC00110; AVX2-O0-NEXT: vxorps %xmm0, %xmm0, %xmm0111; AVX2-O0-NEXT: vmovups %ymm0, {{[0-9]+}}(%rsp)112; AVX2-O0-NEXT: vmovups %ymm0, {{[0-9]+}}(%rsp)113; AVX2-O0-NEXT: movb $1, {{[0-9]+}}(%rsp)114; AVX2-O0-NEXT: movw $32, %cx115; AVX2-O0-NEXT: movw $8, %ax116; AVX2-O0-NEXT: # implicit-def: $al117; AVX2-O0-NEXT: movb %al, {{[0-9]+}}(%rsp)118; AVX2-O0-NEXT: movw %cx, {{[0-9]+}}(%rsp)119; AVX2-O0-NEXT: ldtilecfg {{[0-9]+}}(%rsp)120; AVX2-O0-NEXT: tilezero %tmm0121; AVX2-O0-NEXT: movl $64, %esi122; AVX2-O0-NEXT: leaq {{[0-9]+}}(%rsp), %rdx123; AVX2-O0-NEXT: movw $32, %cx124; AVX2-O0-NEXT: movw $8, %ax125; AVX2-O0-NEXT: tilestored %tmm0, (%rdx,%rsi)126; AVX2-O0-NEXT: movl $64, %esi127; AVX2-O0-NEXT: leaq {{[0-9]+}}(%rsp), %rdx128; AVX2-O0-NEXT: movw $32, %cx129; AVX2-O0-NEXT: movw $8, %ax130; AVX2-O0-NEXT: # implicit-def: $al131; AVX2-O0-NEXT: movb %al, {{[0-9]+}}(%rsp)132; AVX2-O0-NEXT: movw %cx, {{[0-9]+}}(%rsp)133; AVX2-O0-NEXT: ldtilecfg {{[0-9]+}}(%rsp)134; AVX2-O0-NEXT: tileloadd (%rdx,%rsi), %tmm0135; AVX2-O0-NEXT: movl $1024, %edx # imm = 0x400136; AVX2-O0-NEXT: movw $32, %cx137; AVX2-O0-NEXT: movw $8, %ax138; AVX2-O0-NEXT: tilestored %tmm0, (%rdi,%rdx)139; AVX2-O0-NEXT: movq %rbp, %rsp140; AVX2-O0-NEXT: popq %rbp141; AVX2-O0-NEXT: tilerelease142; AVX2-O0-NEXT: vzeroupper143; AVX2-O0-NEXT: retq144;145; SSE2-O0-LABEL: foo:146; SSE2-O0: # %bb.0: # %entry147; SSE2-O0-NEXT: pushq %rbp148; SSE2-O0-NEXT: movq %rsp, %rbp149; SSE2-O0-NEXT: andq $-1024, %rsp # imm = 0xFC00150; SSE2-O0-NEXT: subq $3072, %rsp # imm = 0xC00151; SSE2-O0-NEXT: xorps %xmm0, %xmm0152; SSE2-O0-NEXT: movups %xmm0, {{[0-9]+}}(%rsp)153; SSE2-O0-NEXT: movups %xmm0, {{[0-9]+}}(%rsp)154; SSE2-O0-NEXT: movups %xmm0, {{[0-9]+}}(%rsp)155; SSE2-O0-NEXT: movups %xmm0, {{[0-9]+}}(%rsp)156; SSE2-O0-NEXT: movb $1, {{[0-9]+}}(%rsp)157; SSE2-O0-NEXT: movw $32, %cx158; SSE2-O0-NEXT: movw $8, %ax159; SSE2-O0-NEXT: # implicit-def: $al160; SSE2-O0-NEXT: movb %al, {{[0-9]+}}(%rsp)161; SSE2-O0-NEXT: movw %cx, {{[0-9]+}}(%rsp)162; SSE2-O0-NEXT: ldtilecfg {{[0-9]+}}(%rsp)163; SSE2-O0-NEXT: tilezero %tmm0164; SSE2-O0-NEXT: movl $64, %esi165; SSE2-O0-NEXT: leaq {{[0-9]+}}(%rsp), %rdx166; SSE2-O0-NEXT: movw $32, %cx167; SSE2-O0-NEXT: movw $8, %ax168; SSE2-O0-NEXT: tilestored %tmm0, (%rdx,%rsi)169; SSE2-O0-NEXT: movl $64, %esi170; SSE2-O0-NEXT: leaq {{[0-9]+}}(%rsp), %rdx171; SSE2-O0-NEXT: movw $32, %cx172; SSE2-O0-NEXT: movw $8, %ax173; SSE2-O0-NEXT: # implicit-def: $al174; SSE2-O0-NEXT: movb %al, {{[0-9]+}}(%rsp)175; SSE2-O0-NEXT: movw %cx, {{[0-9]+}}(%rsp)176; SSE2-O0-NEXT: ldtilecfg {{[0-9]+}}(%rsp)177; SSE2-O0-NEXT: tileloadd (%rdx,%rsi), %tmm0178; SSE2-O0-NEXT: movl $1024, %edx # imm = 0x400179; SSE2-O0-NEXT: movw $32, %cx180; SSE2-O0-NEXT: movw $8, %ax181; SSE2-O0-NEXT: tilestored %tmm0, (%rdi,%rdx)182; SSE2-O0-NEXT: movq %rbp, %rsp183; SSE2-O0-NEXT: popq %rbp184; SSE2-O0-NEXT: tilerelease185; SSE2-O0-NEXT: retq186entry:187 %t = call x86_amx @llvm.x86.tilezero.internal(i16 8, i16 32)188 call void @llvm.x86.tilestored64.internal(i16 8, i16 32, ptr %buf, i64 1024, x86_amx %t)189 ret void190}191 192declare x86_amx @llvm.x86.tilezero.internal(i16, i16)193declare void @llvm.x86.tilestored64.internal(i16, i16, ptr, i64, x86_amx)194