151 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=i686-unknown -mattr=+sse2 | FileCheck %s --check-prefix=X86-SSE23; RUN: llc < %s -mtriple=i686-unknown -mattr=+sse4.2 | FileCheck %s --check-prefix=X86-SSE424; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+sse2 | FileCheck %s --check-prefix=X64-SSE25; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+sse4.2 | FileCheck %s --check-prefix=X64-SSE426 7; unsigned to float v7i16 to v7f328 9define void @convert_v7i16_v7f32(ptr %dst.addr, <7 x i16> %src) nounwind {10; X86-SSE2-LABEL: convert_v7i16_v7f32:11; X86-SSE2: # %bb.0: # %entry12; X86-SSE2-NEXT: movl {{[0-9]+}}(%esp), %eax13; X86-SSE2-NEXT: pxor %xmm1, %xmm114; X86-SSE2-NEXT: movdqa %xmm0, %xmm215; X86-SSE2-NEXT: punpckhwd {{.*#+}} xmm2 = xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]16; X86-SSE2-NEXT: cvtdq2ps %xmm2, %xmm217; X86-SSE2-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]18; X86-SSE2-NEXT: cvtdq2ps %xmm0, %xmm019; X86-SSE2-NEXT: movups %xmm0, (%eax)20; X86-SSE2-NEXT: movss %xmm2, 16(%eax)21; X86-SSE2-NEXT: movaps %xmm2, %xmm022; X86-SSE2-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm2[1]23; X86-SSE2-NEXT: movss %xmm0, 24(%eax)24; X86-SSE2-NEXT: shufps {{.*#+}} xmm2 = xmm2[1,1,1,1]25; X86-SSE2-NEXT: movss %xmm2, 20(%eax)26; X86-SSE2-NEXT: retl27;28; X86-SSE42-LABEL: convert_v7i16_v7f32:29; X86-SSE42: # %bb.0: # %entry30; X86-SSE42-NEXT: movl {{[0-9]+}}(%esp), %eax31; X86-SSE42-NEXT: pxor %xmm1, %xmm132; X86-SSE42-NEXT: pmovzxwd {{.*#+}} xmm2 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero33; X86-SSE42-NEXT: punpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]34; X86-SSE42-NEXT: cvtdq2ps %xmm0, %xmm035; X86-SSE42-NEXT: cvtdq2ps %xmm2, %xmm136; X86-SSE42-NEXT: extractps $2, %xmm0, 24(%eax)37; X86-SSE42-NEXT: extractps $1, %xmm0, 20(%eax)38; X86-SSE42-NEXT: movups %xmm1, (%eax)39; X86-SSE42-NEXT: movss %xmm0, 16(%eax)40; X86-SSE42-NEXT: retl41;42; X64-SSE2-LABEL: convert_v7i16_v7f32:43; X64-SSE2: # %bb.0: # %entry44; X64-SSE2-NEXT: pxor %xmm1, %xmm145; X64-SSE2-NEXT: movdqa %xmm0, %xmm246; X64-SSE2-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]47; X64-SSE2-NEXT: cvtdq2ps %xmm2, %xmm248; X64-SSE2-NEXT: punpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]49; X64-SSE2-NEXT: cvtdq2ps %xmm0, %xmm050; X64-SSE2-NEXT: movlps %xmm0, 16(%rdi)51; X64-SSE2-NEXT: movups %xmm2, (%rdi)52; X64-SSE2-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1]53; X64-SSE2-NEXT: movss %xmm0, 24(%rdi)54; X64-SSE2-NEXT: retq55;56; X64-SSE42-LABEL: convert_v7i16_v7f32:57; X64-SSE42: # %bb.0: # %entry58; X64-SSE42-NEXT: pxor %xmm1, %xmm159; X64-SSE42-NEXT: pmovzxwd {{.*#+}} xmm2 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero60; X64-SSE42-NEXT: punpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]61; X64-SSE42-NEXT: cvtdq2ps %xmm0, %xmm062; X64-SSE42-NEXT: cvtdq2ps %xmm2, %xmm163; X64-SSE42-NEXT: extractps $2, %xmm0, 24(%rdi)64; X64-SSE42-NEXT: movlps %xmm0, 16(%rdi)65; X64-SSE42-NEXT: movups %xmm1, (%rdi)66; X64-SSE42-NEXT: retq67entry:68 %val = uitofp <7 x i16> %src to <7 x float>69 store <7 x float> %val, ptr %dst.addr, align 470 ret void71}72 73; unsigned to float v3i8 to v3f3274 75define void @convert_v3i8_to_v3f32(ptr %dst.addr, ptr %src.addr) nounwind {76; X86-SSE2-LABEL: convert_v3i8_to_v3f32:77; X86-SSE2: # %bb.0: # %entry78; X86-SSE2-NEXT: movl {{[0-9]+}}(%esp), %eax79; X86-SSE2-NEXT: movl {{[0-9]+}}(%esp), %ecx80; X86-SSE2-NEXT: movzwl (%ecx), %edx81; X86-SSE2-NEXT: movd %edx, %xmm082; X86-SSE2-NEXT: movdqa {{.*#+}} xmm1 = [255,255,0,255,255,255,255,255,255,255,255,255,255,255,255,255]83; X86-SSE2-NEXT: pand %xmm1, %xmm084; X86-SSE2-NEXT: movzbl 2(%ecx), %ecx85; X86-SSE2-NEXT: movd %ecx, %xmm286; X86-SSE2-NEXT: pslld $16, %xmm287; X86-SSE2-NEXT: pandn %xmm2, %xmm188; X86-SSE2-NEXT: por %xmm0, %xmm189; X86-SSE2-NEXT: pxor %xmm0, %xmm090; X86-SSE2-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]91; X86-SSE2-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]92; X86-SSE2-NEXT: cvtdq2ps %xmm1, %xmm093; X86-SSE2-NEXT: movss %xmm0, (%eax)94; X86-SSE2-NEXT: movaps %xmm0, %xmm195; X86-SSE2-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]96; X86-SSE2-NEXT: movss %xmm1, 8(%eax)97; X86-SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1,1,1]98; X86-SSE2-NEXT: movss %xmm0, 4(%eax)99; X86-SSE2-NEXT: retl100;101; X86-SSE42-LABEL: convert_v3i8_to_v3f32:102; X86-SSE42: # %bb.0: # %entry103; X86-SSE42-NEXT: movl {{[0-9]+}}(%esp), %eax104; X86-SSE42-NEXT: movl {{[0-9]+}}(%esp), %ecx105; X86-SSE42-NEXT: movzwl (%ecx), %edx106; X86-SSE42-NEXT: movd %edx, %xmm0107; X86-SSE42-NEXT: pinsrb $2, 2(%ecx), %xmm0108; X86-SSE42-NEXT: pmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero109; X86-SSE42-NEXT: cvtdq2ps %xmm0, %xmm0110; X86-SSE42-NEXT: extractps $2, %xmm0, 8(%eax)111; X86-SSE42-NEXT: extractps $1, %xmm0, 4(%eax)112; X86-SSE42-NEXT: movss %xmm0, (%eax)113; X86-SSE42-NEXT: retl114;115; X64-SSE2-LABEL: convert_v3i8_to_v3f32:116; X64-SSE2: # %bb.0: # %entry117; X64-SSE2-NEXT: movzwl (%rsi), %eax118; X64-SSE2-NEXT: movd %eax, %xmm0119; X64-SSE2-NEXT: movdqa {{.*#+}} xmm1 = [255,255,0,255,255,255,255,255,255,255,255,255,255,255,255,255]120; X64-SSE2-NEXT: pand %xmm1, %xmm0121; X64-SSE2-NEXT: movzbl 2(%rsi), %eax122; X64-SSE2-NEXT: movd %eax, %xmm2123; X64-SSE2-NEXT: pslld $16, %xmm2124; X64-SSE2-NEXT: pandn %xmm2, %xmm1125; X64-SSE2-NEXT: por %xmm0, %xmm1126; X64-SSE2-NEXT: pxor %xmm0, %xmm0127; X64-SSE2-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]128; X64-SSE2-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]129; X64-SSE2-NEXT: cvtdq2ps %xmm1, %xmm0130; X64-SSE2-NEXT: movlps %xmm0, (%rdi)131; X64-SSE2-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1]132; X64-SSE2-NEXT: movss %xmm0, 8(%rdi)133; X64-SSE2-NEXT: retq134;135; X64-SSE42-LABEL: convert_v3i8_to_v3f32:136; X64-SSE42: # %bb.0: # %entry137; X64-SSE42-NEXT: movzwl (%rsi), %eax138; X64-SSE42-NEXT: movd %eax, %xmm0139; X64-SSE42-NEXT: pinsrb $2, 2(%rsi), %xmm0140; X64-SSE42-NEXT: pmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero141; X64-SSE42-NEXT: cvtdq2ps %xmm0, %xmm0142; X64-SSE42-NEXT: extractps $2, %xmm0, 8(%rdi)143; X64-SSE42-NEXT: movlps %xmm0, (%rdi)144; X64-SSE42-NEXT: retq145entry:146 %load = load <3 x i8>, ptr %src.addr, align 1147 %cvt = uitofp <3 x i8> %load to <3 x float>148 store <3 x float> %cvt, ptr %dst.addr, align 4149 ret void150}151