156 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+sse2 | FileCheck %s --check-prefix=X863; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse2 | FileCheck %s --check-prefix=X644; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+f16c | FileCheck %s --check-prefix=F16C5; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+f16c -O0 | FileCheck %s --check-prefix=F16C-O06 7define <1 x half> @ir_fadd_v1f16(<1 x half> %arg0, <1 x half> %arg1) nounwind {8; X86-LABEL: ir_fadd_v1f16:9; X86: # %bb.0:10; X86-NEXT: subl $28, %esp11; X86-NEXT: movups %xmm1, {{[-0-9]+}}(%e{{[sb]}}p) # 16-byte Spill12; X86-NEXT: pextrw $0, %xmm0, %eax13; X86-NEXT: movw %ax, (%esp)14; X86-NEXT: calll __extendhfsf215; X86-NEXT: movdqu {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 # 16-byte Reload16; X86-NEXT: pextrw $0, %xmm0, %eax17; X86-NEXT: movw %ax, (%esp)18; X86-NEXT: fstps {{[0-9]+}}(%esp)19; X86-NEXT: calll __extendhfsf220; X86-NEXT: fstps {{[0-9]+}}(%esp)21; X86-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero22; X86-NEXT: addss {{[0-9]+}}(%esp), %xmm023; X86-NEXT: movss %xmm0, (%esp)24; X86-NEXT: calll __truncsfhf225; X86-NEXT: addl $28, %esp26; X86-NEXT: retl27;28; X64-LABEL: ir_fadd_v1f16:29; X64: # %bb.0:30; X64-NEXT: subq $40, %rsp31; X64-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill32; X64-NEXT: movaps %xmm1, %xmm033; X64-NEXT: callq __extendhfsf2@PLT34; X64-NEXT: movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill35; X64-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload36; X64-NEXT: callq __extendhfsf2@PLT37; X64-NEXT: addss {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 4-byte Folded Reload38; X64-NEXT: callq __truncsfhf2@PLT39; X64-NEXT: addq $40, %rsp40; X64-NEXT: retq41;42; F16C-LABEL: ir_fadd_v1f16:43; F16C: # %bb.0:44; F16C-NEXT: vcvtph2ps %xmm1, %xmm145; F16C-NEXT: vcvtph2ps %xmm0, %xmm046; F16C-NEXT: vaddss %xmm1, %xmm0, %xmm047; F16C-NEXT: vcvtps2ph $4, %xmm0, %xmm048; F16C-NEXT: retq49;50; F16C-O0-LABEL: ir_fadd_v1f16:51; F16C-O0: # %bb.0:52; F16C-O0-NEXT: vcvtph2ps %xmm1, %xmm153; F16C-O0-NEXT: vcvtph2ps %xmm0, %xmm054; F16C-O0-NEXT: vaddss %xmm1, %xmm0, %xmm055; F16C-O0-NEXT: vcvtps2ph $4, %xmm0, %xmm056; F16C-O0-NEXT: retq57 %retval = fadd <1 x half> %arg0, %arg158 ret <1 x half> %retval59}60 61define <2 x half> @ir_fadd_v2f16(<2 x half> %arg0, <2 x half> %arg1) nounwind {62; X86-LABEL: ir_fadd_v2f16:63; X86: # %bb.0:64; X86-NEXT: subl $84, %esp65; X86-NEXT: movdqu %xmm0, {{[-0-9]+}}(%e{{[sb]}}p) # 16-byte Spill66; X86-NEXT: psrld $16, %xmm067; X86-NEXT: movdqu %xmm0, {{[-0-9]+}}(%e{{[sb]}}p) # 16-byte Spill68; X86-NEXT: movdqa %xmm1, %xmm069; X86-NEXT: psrld $16, %xmm070; X86-NEXT: movdqu %xmm0, {{[-0-9]+}}(%e{{[sb]}}p) # 16-byte Spill71; X86-NEXT: pextrw $0, %xmm1, %eax72; X86-NEXT: movw %ax, (%esp)73; X86-NEXT: calll __extendhfsf274; X86-NEXT: fstpt {{[-0-9]+}}(%e{{[sb]}}p) # 10-byte Folded Spill75; X86-NEXT: movdqu {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 # 16-byte Reload76; X86-NEXT: pextrw $0, %xmm0, %eax77; X86-NEXT: movw %ax, (%esp)78; X86-NEXT: calll __extendhfsf279; X86-NEXT: fstpt {{[-0-9]+}}(%e{{[sb]}}p) # 10-byte Folded Spill80; X86-NEXT: movdqu {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 # 16-byte Reload81; X86-NEXT: pextrw $0, %xmm0, %eax82; X86-NEXT: movw %ax, (%esp)83; X86-NEXT: calll __extendhfsf284; X86-NEXT: movdqu {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 # 16-byte Reload85; X86-NEXT: pextrw $0, %xmm0, %eax86; X86-NEXT: movw %ax, (%esp)87; X86-NEXT: fstps {{[0-9]+}}(%esp)88; X86-NEXT: fldt {{[-0-9]+}}(%e{{[sb]}}p) # 10-byte Folded Reload89; X86-NEXT: fstps {{[0-9]+}}(%esp)90; X86-NEXT: calll __extendhfsf291; X86-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero92; X86-NEXT: addss {{[0-9]+}}(%esp), %xmm093; X86-NEXT: movss %xmm0, (%esp)94; X86-NEXT: fstps {{[0-9]+}}(%esp)95; X86-NEXT: fldt {{[-0-9]+}}(%e{{[sb]}}p) # 10-byte Folded Reload96; X86-NEXT: fstps {{[0-9]+}}(%esp)97; X86-NEXT: calll __truncsfhf298; X86-NEXT: movups %xmm0, {{[-0-9]+}}(%e{{[sb]}}p) # 16-byte Spill99; X86-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero100; X86-NEXT: addss {{[0-9]+}}(%esp), %xmm0101; X86-NEXT: movss %xmm0, (%esp)102; X86-NEXT: calll __truncsfhf2103; X86-NEXT: movdqu {{[-0-9]+}}(%e{{[sb]}}p), %xmm1 # 16-byte Reload104; X86-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]105; X86-NEXT: addl $84, %esp106; X86-NEXT: retl107;108; X64-LABEL: ir_fadd_v2f16:109; X64: # %bb.0:110; X64-NEXT: subq $72, %rsp111; X64-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill112; X64-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill113; X64-NEXT: psrld $16, %xmm0114; X64-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill115; X64-NEXT: movdqa %xmm1, %xmm0116; X64-NEXT: psrld $16, %xmm0117; X64-NEXT: callq __extendhfsf2@PLT118; X64-NEXT: movd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Folded Spill119; X64-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload120; X64-NEXT: callq __extendhfsf2@PLT121; X64-NEXT: addss {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 4-byte Folded Reload122; X64-NEXT: callq __truncsfhf2@PLT123; X64-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill124; X64-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload125; X64-NEXT: callq __extendhfsf2@PLT126; X64-NEXT: movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill127; X64-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload128; X64-NEXT: callq __extendhfsf2@PLT129; X64-NEXT: addss {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 4-byte Folded Reload130; X64-NEXT: callq __truncsfhf2@PLT131; X64-NEXT: punpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload132; X64-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]133; X64-NEXT: addq $72, %rsp134; X64-NEXT: retq135;136; F16C-LABEL: ir_fadd_v2f16:137; F16C: # %bb.0:138; F16C-NEXT: vcvtph2ps %xmm1, %ymm1139; F16C-NEXT: vcvtph2ps %xmm0, %ymm0140; F16C-NEXT: vaddps %ymm1, %ymm0, %ymm0141; F16C-NEXT: vcvtps2ph $4, %ymm0, %xmm0142; F16C-NEXT: vzeroupper143; F16C-NEXT: retq144;145; F16C-O0-LABEL: ir_fadd_v2f16:146; F16C-O0: # %bb.0:147; F16C-O0-NEXT: vcvtph2ps %xmm1, %ymm1148; F16C-O0-NEXT: vcvtph2ps %xmm0, %ymm0149; F16C-O0-NEXT: vaddps %ymm1, %ymm0, %ymm0150; F16C-O0-NEXT: vcvtps2ph $4, %ymm0, %xmm0151; F16C-O0-NEXT: vzeroupper152; F16C-O0-NEXT: retq153 %retval = fadd <2 x half> %arg0, %arg1154 ret <2 x half> %retval155}156