48 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-tile,+amx-bf16,+avx512f, \3; RUN: -mattr=+amx-complex \4; RUN: -verify-machineinstrs | FileCheck %s5 6define void @test_amx(ptr %pointer, ptr %base, i64 %stride) {7; CHECK-LABEL: test_amx:8; CHECK: # %bb.0:9; CHECK-NEXT: vxorps %xmm0, %xmm0, %xmm010; CHECK-NEXT: vmovups %zmm0, -{{[0-9]+}}(%rsp)11; CHECK-NEXT: movb $1, -{{[0-9]+}}(%rsp)12; CHECK-NEXT: movb $8, -{{[0-9]+}}(%rsp)13; CHECK-NEXT: movw $8, -{{[0-9]+}}(%rsp)14; CHECK-NEXT: movb $8, -{{[0-9]+}}(%rsp)15; CHECK-NEXT: movw $8, -{{[0-9]+}}(%rsp)16; CHECK-NEXT: movb $8, -{{[0-9]+}}(%rsp)17; CHECK-NEXT: movw $8, -{{[0-9]+}}(%rsp)18; CHECK-NEXT: ldtilecfg -{{[0-9]+}}(%rsp)19; CHECK-NEXT: movw $8, %ax20; CHECK-NEXT: tileloadd (%rsi,%rdx), %tmm021; CHECK-NEXT: tilezero %tmm122; CHECK-NEXT: tilezero %tmm223; CHECK-NEXT: tcmmimfp16ps %tmm1, %tmm0, %tmm224; CHECK-NEXT: tcmmrlfp16ps %tmm1, %tmm0, %tmm225; CHECK-NEXT: tilestored %tmm2, (%rdi,%rdx)26; CHECK-NEXT: tilerelease27; CHECK-NEXT: vzeroupper28; CHECK-NEXT: retq29 30 %a = call x86_amx @llvm.x86.tileloadd64.internal(i16 8, i16 8, ptr %base, i64 %stride)31 %b = call x86_amx @llvm.x86.tilezero.internal(i16 8, i16 8)32 %c = call x86_amx @llvm.x86.tilezero.internal(i16 8, i16 8)33 34 %c1 = call x86_amx @llvm.x86.tcmmimfp16ps.internal(i16 8, i16 8, i16 8, x86_amx %c, x86_amx %a, x86_amx %b)35 %c2 = call x86_amx @llvm.x86.tcmmrlfp16ps.internal(i16 8, i16 8, i16 8, x86_amx %c1, x86_amx %a, x86_amx %b)36 37 call void @llvm.x86.tilestored64.internal(i16 8, i16 8, ptr %pointer, i64 %stride, x86_amx %c2)38 ret void39}40 41declare x86_amx @llvm.x86.tilezero.internal(i16, i16)42declare x86_amx @llvm.x86.tileloadd64.internal(i16, i16, ptr, i64)43declare x86_amx @llvm.x86.tileloaddt164.internal(i16, i16, ptr, i64)44declare void @llvm.x86.tilestored64.internal(i16, i16, ptr, i64, x86_amx)45 46declare x86_amx @llvm.x86.tcmmimfp16ps.internal(i16, i16, i16, x86_amx, x86_amx, x86_amx)47declare x86_amx @llvm.x86.tcmmrlfp16ps.internal(i16, i16, i16, x86_amx, x86_amx, x86_amx)48