44 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512bf16 | FileCheck %s --check-prefix=X643; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+avx512bf16 | FileCheck %s --check-prefix=X864 5define dso_local void @funbf16(ptr readonly %src, ptr writeonly %dst) {6; X64-LABEL: funbf16:7; X64: # %bb.0: # %entry8; X64-NEXT: vmovups (%rdi), %xmm09; X64-NEXT: vmovups %xmm0, (%rsi)10; X64-NEXT: vmovaps (%rdi), %xmm011; X64-NEXT: vmovaps %xmm0, (%rsi)12; X64-NEXT: vmovups (%rdi), %ymm013; X64-NEXT: vmovups %ymm0, (%rsi)14; X64-NEXT: vmovaps (%rdi), %ymm015; X64-NEXT: vmovaps %ymm0, (%rsi)16; X64-NEXT: vzeroupper17; X64-NEXT: retq18;19; X86-LABEL: funbf16:20; X86: # %bb.0: # %entry21; X86-NEXT: movl {{[0-9]+}}(%esp), %eax22; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx23; X86-NEXT: vmovups (%ecx), %xmm024; X86-NEXT: vmovups %xmm0, (%eax)25; X86-NEXT: vmovaps (%ecx), %xmm026; X86-NEXT: vmovaps %xmm0, (%eax)27; X86-NEXT: vmovups (%ecx), %ymm028; X86-NEXT: vmovups %ymm0, (%eax)29; X86-NEXT: vmovaps (%ecx), %ymm030; X86-NEXT: vmovaps %ymm0, (%eax)31; X86-NEXT: vzeroupper32; X86-NEXT: retl33entry:34 %0 = load <8 x bfloat>, ptr %src, align 135 store <8 x bfloat> %0, ptr %dst, align 136 %1 = load <8 x bfloat>, ptr %src, align 3237 store <8 x bfloat> %1, ptr %dst, align 3238 %2 = load <16 x bfloat>, ptr %src, align 139 store <16 x bfloat> %2, ptr %dst, align 140 %3 = load <16 x bfloat>, ptr %src, align 3241 store <16 x bfloat> %3, ptr %dst, align 3242 ret void43}44