880 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=x86_64-linux | FileCheck %s --check-prefix=X643; RUN: llc < %s -mtriple=x86_64-linux -mattr=+avx2 | FileCheck %s --check-prefix=X64-AVX24; RUN: llc < %s -mtriple=i686 -mattr=cmov | FileCheck %s --check-prefix=X865 6declare <2 x i64> @llvm.sshl.sat.v2i64(<2 x i64>, <2 x i64>)7declare <4 x i32> @llvm.sshl.sat.v4i32(<4 x i32>, <4 x i32>)8declare <8 x i16> @llvm.sshl.sat.v8i16(<8 x i16>, <8 x i16>)9declare <16 x i8> @llvm.sshl.sat.v16i8(<16 x i8>, <16 x i8>)10 11define <2 x i64> @vec_v2i64(<2 x i64> %x, <2 x i64> %y) nounwind {12; X64-LABEL: vec_v2i64:13; X64: # %bb.0:14; X64-NEXT: movdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]15; X64-NEXT: movdqa %xmm2, %xmm316; X64-NEXT: psrlq %xmm1, %xmm317; X64-NEXT: pshufd {{.*#+}} xmm4 = xmm1[2,3,2,3]18; X64-NEXT: movdqa %xmm2, %xmm519; X64-NEXT: psrlq %xmm4, %xmm520; X64-NEXT: movsd {{.*#+}} xmm5 = xmm3[0],xmm5[1]21; X64-NEXT: movdqa %xmm0, %xmm622; X64-NEXT: psllq %xmm1, %xmm623; X64-NEXT: movdqa %xmm0, %xmm324; X64-NEXT: psllq %xmm4, %xmm325; X64-NEXT: movdqa %xmm3, %xmm726; X64-NEXT: movsd {{.*#+}} xmm3 = xmm6[0],xmm3[1]27; X64-NEXT: psrlq %xmm1, %xmm628; X64-NEXT: psrlq %xmm4, %xmm729; X64-NEXT: movsd {{.*#+}} xmm7 = xmm6[0],xmm7[1]30; X64-NEXT: xorpd %xmm5, %xmm731; X64-NEXT: psubq %xmm5, %xmm732; X64-NEXT: pcmpeqd %xmm0, %xmm733; X64-NEXT: pshufd {{.*#+}} xmm1 = xmm7[1,0,3,2]34; X64-NEXT: pand %xmm7, %xmm135; X64-NEXT: andpd %xmm1, %xmm336; X64-NEXT: pshufd {{.*#+}} xmm4 = xmm0[1,1,3,3]37; X64-NEXT: pand %xmm2, %xmm038; X64-NEXT: pxor %xmm5, %xmm539; X64-NEXT: pcmpgtd %xmm4, %xmm540; X64-NEXT: por %xmm2, %xmm541; X64-NEXT: pcmpeqd %xmm2, %xmm242; X64-NEXT: pxor %xmm5, %xmm243; X64-NEXT: por %xmm0, %xmm244; X64-NEXT: pandn %xmm2, %xmm145; X64-NEXT: por %xmm3, %xmm146; X64-NEXT: movdqa %xmm1, %xmm047; X64-NEXT: retq48;49; X64-AVX2-LABEL: vec_v2i64:50; X64-AVX2: # %bb.0:51; X64-AVX2-NEXT: vmovddup {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]52; X64-AVX2-NEXT: # xmm2 = mem[0,0]53; X64-AVX2-NEXT: vmovddup {{.*#+}} xmm3 = [9223372036854775807,9223372036854775807]54; X64-AVX2-NEXT: # xmm3 = mem[0,0]55; X64-AVX2-NEXT: vblendvpd %xmm0, %xmm2, %xmm3, %xmm356; X64-AVX2-NEXT: vpsrlvq %xmm1, %xmm2, %xmm257; X64-AVX2-NEXT: vpsllvq %xmm1, %xmm0, %xmm458; X64-AVX2-NEXT: vpsrlvq %xmm1, %xmm4, %xmm159; X64-AVX2-NEXT: vpxor %xmm2, %xmm1, %xmm160; X64-AVX2-NEXT: vpsubq %xmm2, %xmm1, %xmm161; X64-AVX2-NEXT: vpcmpeqq %xmm1, %xmm0, %xmm062; X64-AVX2-NEXT: vblendvpd %xmm0, %xmm4, %xmm3, %xmm063; X64-AVX2-NEXT: retq64;65; X86-LABEL: vec_v2i64:66; X86: # %bb.0:67; X86-NEXT: pushl %ebp68; X86-NEXT: pushl %ebx69; X86-NEXT: pushl %edi70; X86-NEXT: pushl %esi71; X86-NEXT: subl $20, %esp72; X86-NEXT: movzbl {{[0-9]+}}(%esp), %ecx73; X86-NEXT: movl {{[0-9]+}}(%esp), %edx74; X86-NEXT: movl {{[0-9]+}}(%esp), %edi75; X86-NEXT: movl %edx, %eax76; X86-NEXT: shll %cl, %eax77; X86-NEXT: shldl %cl, %edx, %edi78; X86-NEXT: xorl %edx, %edx79; X86-NEXT: testb $32, %cl80; X86-NEXT: cmovnel %eax, %edi81; X86-NEXT: cmovnel %edx, %eax82; X86-NEXT: movl %eax, (%esp) # 4-byte Spill83; X86-NEXT: movl %edi, %ebx84; X86-NEXT: sarl %cl, %ebx85; X86-NEXT: movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill86; X86-NEXT: movl %edi, %eax87; X86-NEXT: movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill88; X86-NEXT: sarl $31, %eax89; X86-NEXT: testb $32, %cl90; X86-NEXT: cmovel %ebx, %eax91; X86-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill92; X86-NEXT: movb {{[0-9]+}}(%esp), %ch93; X86-NEXT: movl {{[0-9]+}}(%esp), %esi94; X86-NEXT: movl %esi, %eax95; X86-NEXT: movb %ch, %cl96; X86-NEXT: shll %cl, %eax97; X86-NEXT: movl {{[0-9]+}}(%esp), %ebp98; X86-NEXT: movl %ebp, %ebx99; X86-NEXT: shldl %cl, %esi, %ebx100; X86-NEXT: testb $32, %ch101; X86-NEXT: cmovnel %eax, %ebx102; X86-NEXT: cmovnel %edx, %eax103; X86-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill104; X86-NEXT: movl %ebx, %esi105; X86-NEXT: sarl %cl, %esi106; X86-NEXT: movl %ebx, %edx107; X86-NEXT: sarl $31, %edx108; X86-NEXT: testb $32, %ch109; X86-NEXT: cmovel %esi, %edx110; X86-NEXT: movl (%esp), %eax # 4-byte Reload111; X86-NEXT: movb {{[0-9]+}}(%esp), %cl112; X86-NEXT: shrdl %cl, %edi, %eax113; X86-NEXT: testb $32, %cl114; X86-NEXT: cmovnel {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload115; X86-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload116; X86-NEXT: movb %ch, %cl117; X86-NEXT: shrdl %cl, %ebx, %edi118; X86-NEXT: testb $32, %ch119; X86-NEXT: cmovnel %esi, %edi120; X86-NEXT: xorl {{[0-9]+}}(%esp), %eax121; X86-NEXT: movl {{[0-9]+}}(%esp), %esi122; X86-NEXT: xorl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Folded Spill123; X86-NEXT: sarl $31, %esi124; X86-NEXT: movl %esi, %ecx125; X86-NEXT: xorl $2147483647, %ecx # imm = 0x7FFFFFFF126; X86-NEXT: orl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload127; X86-NEXT: notl %esi128; X86-NEXT: cmovel (%esp), %esi # 4-byte Folded Reload129; X86-NEXT: movl %esi, (%esp) # 4-byte Spill130; X86-NEXT: cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload131; X86-NEXT: xorl {{[0-9]+}}(%esp), %edi132; X86-NEXT: xorl %ebp, %edx133; X86-NEXT: sarl $31, %ebp134; X86-NEXT: movl %ebp, %esi135; X86-NEXT: xorl $2147483647, %esi # imm = 0x7FFFFFFF136; X86-NEXT: orl %edx, %edi137; X86-NEXT: notl %ebp138; X86-NEXT: cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload139; X86-NEXT: cmovel %ebx, %esi140; X86-NEXT: movl {{[0-9]+}}(%esp), %eax141; X86-NEXT: movl %esi, 12(%eax)142; X86-NEXT: movl %ebp, 8(%eax)143; X86-NEXT: movl %ecx, 4(%eax)144; X86-NEXT: movl (%esp), %ecx # 4-byte Reload145; X86-NEXT: movl %ecx, (%eax)146; X86-NEXT: addl $20, %esp147; X86-NEXT: popl %esi148; X86-NEXT: popl %edi149; X86-NEXT: popl %ebx150; X86-NEXT: popl %ebp151; X86-NEXT: retl $4152 %tmp = call <2 x i64> @llvm.sshl.sat.v2i64(<2 x i64> %x, <2 x i64> %y)153 ret <2 x i64> %tmp154}155 156define <4 x i32> @vec_v4i32(<4 x i32> %x, <4 x i32> %y) nounwind {157; X64-LABEL: vec_v4i32:158; X64: # %bb.0:159; X64-NEXT: pshufd {{.*#+}} xmm3 = xmm1[2,3,2,3]160; X64-NEXT: pshuflw {{.*#+}} xmm4 = xmm1[2,3,3,3,4,5,6,7]161; X64-NEXT: pshuflw {{.*#+}} xmm5 = xmm1[0,1,1,1,4,5,6,7]162; X64-NEXT: pslld $23, %xmm1163; X64-NEXT: paddd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1164; X64-NEXT: cvttps2dq %xmm1, %xmm6165; X64-NEXT: movdqa %xmm0, %xmm1166; X64-NEXT: pmuludq %xmm6, %xmm1167; X64-NEXT: pshufd {{.*#+}} xmm2 = xmm1[0,2,2,3]168; X64-NEXT: pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3]169; X64-NEXT: pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]170; X64-NEXT: pmuludq %xmm7, %xmm6171; X64-NEXT: pshufd {{.*#+}} xmm6 = xmm6[0,2,2,3]172; X64-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1]173; X64-NEXT: pshuflw {{.*#+}} xmm6 = xmm3[2,3,3,3,4,5,6,7]174; X64-NEXT: movdqa %xmm2, %xmm7175; X64-NEXT: psrad %xmm6, %xmm7176; X64-NEXT: pshuflw {{.*#+}} xmm3 = xmm3[0,1,1,1,4,5,6,7]177; X64-NEXT: movdqa %xmm1, %xmm6178; X64-NEXT: psrad %xmm3, %xmm6179; X64-NEXT: punpckhqdq {{.*#+}} xmm6 = xmm6[1],xmm7[1]180; X64-NEXT: movdqa %xmm2, %xmm3181; X64-NEXT: psrad %xmm4, %xmm3182; X64-NEXT: psrad %xmm5, %xmm1183; X64-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm3[0]184; X64-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,3],xmm6[0,3]185; X64-NEXT: pcmpeqd %xmm0, %xmm1186; X64-NEXT: pand %xmm1, %xmm2187; X64-NEXT: pxor %xmm3, %xmm3188; X64-NEXT: pcmpgtd %xmm0, %xmm3189; X64-NEXT: pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3190; X64-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0191; X64-NEXT: por %xmm3, %xmm0192; X64-NEXT: pandn %xmm0, %xmm1193; X64-NEXT: por %xmm2, %xmm1194; X64-NEXT: movdqa %xmm1, %xmm0195; X64-NEXT: retq196;197; X64-AVX2-LABEL: vec_v4i32:198; X64-AVX2: # %bb.0:199; X64-AVX2-NEXT: vbroadcastss {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]200; X64-AVX2-NEXT: vbroadcastss {{.*#+}} xmm3 = [2147483647,2147483647,2147483647,2147483647]201; X64-AVX2-NEXT: vblendvps %xmm0, %xmm2, %xmm3, %xmm2202; X64-AVX2-NEXT: vpsllvd %xmm1, %xmm0, %xmm3203; X64-AVX2-NEXT: vpsravd %xmm1, %xmm3, %xmm1204; X64-AVX2-NEXT: vpcmpeqd %xmm1, %xmm0, %xmm0205; X64-AVX2-NEXT: vblendvps %xmm0, %xmm3, %xmm2, %xmm0206; X64-AVX2-NEXT: retq207;208; X86-LABEL: vec_v4i32:209; X86: # %bb.0:210; X86-NEXT: pushl %ebp211; X86-NEXT: pushl %ebx212; X86-NEXT: pushl %edi213; X86-NEXT: pushl %esi214; X86-NEXT: movl {{[0-9]+}}(%esp), %esi215; X86-NEXT: movb {{[0-9]+}}(%esp), %ch216; X86-NEXT: movb {{[0-9]+}}(%esp), %cl217; X86-NEXT: movl {{[0-9]+}}(%esp), %edi218; X86-NEXT: movl %edi, %edx219; X86-NEXT: shll %cl, %edx220; X86-NEXT: movl %edx, %ebp221; X86-NEXT: sarl %cl, %ebp222; X86-NEXT: xorl %ebx, %ebx223; X86-NEXT: testl %edi, %edi224; X86-NEXT: sets %bl225; X86-NEXT: addl $2147483647, %ebx # imm = 0x7FFFFFFF226; X86-NEXT: cmpl %ebp, %edi227; X86-NEXT: movl {{[0-9]+}}(%esp), %edi228; X86-NEXT: cmovel %edx, %ebx229; X86-NEXT: movl %edi, %ebp230; X86-NEXT: movb %ch, %cl231; X86-NEXT: shll %cl, %ebp232; X86-NEXT: movl %ebp, %eax233; X86-NEXT: sarl %cl, %eax234; X86-NEXT: xorl %edx, %edx235; X86-NEXT: testl %edi, %edi236; X86-NEXT: sets %dl237; X86-NEXT: addl $2147483647, %edx # imm = 0x7FFFFFFF238; X86-NEXT: cmpl %eax, %edi239; X86-NEXT: cmovel %ebp, %edx240; X86-NEXT: movl %esi, %edi241; X86-NEXT: movzbl {{[0-9]+}}(%esp), %ecx242; X86-NEXT: shll %cl, %edi243; X86-NEXT: movl %edi, %ebp244; X86-NEXT: sarl %cl, %ebp245; X86-NEXT: xorl %eax, %eax246; X86-NEXT: testl %esi, %esi247; X86-NEXT: sets %al248; X86-NEXT: addl $2147483647, %eax # imm = 0x7FFFFFFF249; X86-NEXT: cmpl %ebp, %esi250; X86-NEXT: movl {{[0-9]+}}(%esp), %esi251; X86-NEXT: cmovel %edi, %eax252; X86-NEXT: movl %esi, %edi253; X86-NEXT: movzbl {{[0-9]+}}(%esp), %ecx254; X86-NEXT: shll %cl, %edi255; X86-NEXT: movl %edi, %ebp256; X86-NEXT: sarl %cl, %ebp257; X86-NEXT: xorl %ecx, %ecx258; X86-NEXT: testl %esi, %esi259; X86-NEXT: sets %cl260; X86-NEXT: addl $2147483647, %ecx # imm = 0x7FFFFFFF261; X86-NEXT: cmpl %ebp, %esi262; X86-NEXT: cmovel %edi, %ecx263; X86-NEXT: movl {{[0-9]+}}(%esp), %esi264; X86-NEXT: movl %ecx, 12(%esi)265; X86-NEXT: movl %eax, 8(%esi)266; X86-NEXT: movl %edx, 4(%esi)267; X86-NEXT: movl %ebx, (%esi)268; X86-NEXT: movl %esi, %eax269; X86-NEXT: popl %esi270; X86-NEXT: popl %edi271; X86-NEXT: popl %ebx272; X86-NEXT: popl %ebp273; X86-NEXT: retl $4274 %tmp = call <4 x i32> @llvm.sshl.sat.v4i32(<4 x i32> %x, <4 x i32> %y)275 ret <4 x i32> %tmp276}277 278define <8 x i16> @vec_v8i16(<8 x i16> %x, <8 x i16> %y) nounwind {279; X64-LABEL: vec_v8i16:280; X64: # %bb.0:281; X64-NEXT: movdqa %xmm1, %xmm2282; X64-NEXT: punpckhwd {{.*#+}} xmm2 = xmm2[4,4,5,5,6,6,7,7]283; X64-NEXT: pslld $23, %xmm2284; X64-NEXT: movdqa {{.*#+}} xmm3 = [1065353216,1065353216,1065353216,1065353216]285; X64-NEXT: paddd %xmm3, %xmm2286; X64-NEXT: cvttps2dq %xmm2, %xmm2287; X64-NEXT: pslld $16, %xmm2288; X64-NEXT: psrad $16, %xmm2289; X64-NEXT: movdqa %xmm1, %xmm4290; X64-NEXT: punpcklwd {{.*#+}} xmm4 = xmm4[0,0,1,1,2,2,3,3]291; X64-NEXT: pslld $23, %xmm4292; X64-NEXT: paddd %xmm3, %xmm4293; X64-NEXT: cvttps2dq %xmm4, %xmm3294; X64-NEXT: pslld $16, %xmm3295; X64-NEXT: psrad $16, %xmm3296; X64-NEXT: packssdw %xmm2, %xmm3297; X64-NEXT: pmullw %xmm0, %xmm3298; X64-NEXT: psllw $12, %xmm1299; X64-NEXT: movdqa %xmm1, %xmm2300; X64-NEXT: psraw $15, %xmm2301; X64-NEXT: movdqa %xmm3, %xmm4302; X64-NEXT: psraw $8, %xmm4303; X64-NEXT: pand %xmm2, %xmm4304; X64-NEXT: pandn %xmm3, %xmm2305; X64-NEXT: por %xmm4, %xmm2306; X64-NEXT: paddw %xmm1, %xmm1307; X64-NEXT: movdqa %xmm1, %xmm4308; X64-NEXT: psraw $15, %xmm4309; X64-NEXT: movdqa %xmm4, %xmm5310; X64-NEXT: pandn %xmm2, %xmm5311; X64-NEXT: psraw $4, %xmm2312; X64-NEXT: pand %xmm4, %xmm2313; X64-NEXT: por %xmm5, %xmm2314; X64-NEXT: paddw %xmm1, %xmm1315; X64-NEXT: movdqa %xmm1, %xmm4316; X64-NEXT: psraw $15, %xmm4317; X64-NEXT: movdqa %xmm4, %xmm5318; X64-NEXT: pandn %xmm2, %xmm5319; X64-NEXT: psraw $2, %xmm2320; X64-NEXT: pand %xmm4, %xmm2321; X64-NEXT: por %xmm5, %xmm2322; X64-NEXT: paddw %xmm1, %xmm1323; X64-NEXT: psraw $15, %xmm1324; X64-NEXT: movdqa %xmm1, %xmm4325; X64-NEXT: pandn %xmm2, %xmm4326; X64-NEXT: psraw $1, %xmm2327; X64-NEXT: pand %xmm1, %xmm2328; X64-NEXT: por %xmm4, %xmm2329; X64-NEXT: pcmpeqw %xmm0, %xmm2330; X64-NEXT: pand %xmm2, %xmm3331; X64-NEXT: pxor %xmm1, %xmm1332; X64-NEXT: pcmpgtw %xmm0, %xmm1333; X64-NEXT: pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1334; X64-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0335; X64-NEXT: por %xmm1, %xmm0336; X64-NEXT: pandn %xmm0, %xmm2337; X64-NEXT: por %xmm3, %xmm2338; X64-NEXT: movdqa %xmm2, %xmm0339; X64-NEXT: retq340;341; X64-AVX2-LABEL: vec_v8i16:342; X64-AVX2: # %bb.0:343; X64-AVX2-NEXT: vpmovzxwd {{.*#+}} ymm1 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero,xmm1[4],zero,xmm1[5],zero,xmm1[6],zero,xmm1[7],zero344; X64-AVX2-NEXT: vpmovzxwd {{.*#+}} ymm2 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero345; X64-AVX2-NEXT: vpsllvd %ymm1, %ymm2, %ymm2346; X64-AVX2-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[0,1,4,5,8,9,12,13,u,u,u,u,u,u,u,u,16,17,20,21,24,25,28,29,u,u,u,u,u,u,u,u]347; X64-AVX2-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,2,2,3]348; X64-AVX2-NEXT: vpmovsxwd %xmm2, %ymm3349; X64-AVX2-NEXT: vpsravd %ymm1, %ymm3, %ymm1350; X64-AVX2-NEXT: vextracti128 $1, %ymm1, %xmm3351; X64-AVX2-NEXT: vpackssdw %xmm3, %xmm1, %xmm1352; X64-AVX2-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm1353; X64-AVX2-NEXT: vpxor %xmm3, %xmm3, %xmm3354; X64-AVX2-NEXT: vpcmpgtw %xmm0, %xmm3, %xmm0355; X64-AVX2-NEXT: vpbroadcastw {{.*#+}} xmm3 = [32767,32767,32767,32767,32767,32767,32767,32767]356; X64-AVX2-NEXT: vpblendvb %xmm0, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3, %xmm0357; X64-AVX2-NEXT: vpblendvb %xmm1, %xmm2, %xmm0, %xmm0358; X64-AVX2-NEXT: vzeroupper359; X64-AVX2-NEXT: retq360;361; X86-LABEL: vec_v8i16:362; X86: # %bb.0:363; X86-NEXT: pushl %ebp364; X86-NEXT: pushl %ebx365; X86-NEXT: pushl %edi366; X86-NEXT: pushl %esi367; X86-NEXT: subl $16, %esp368; X86-NEXT: movl {{[0-9]+}}(%esp), %eax369; X86-NEXT: movl {{[0-9]+}}(%esp), %edx370; X86-NEXT: movl {{[0-9]+}}(%esp), %esi371; X86-NEXT: movl {{[0-9]+}}(%esp), %edi372; X86-NEXT: movl %edi, %ebx373; X86-NEXT: movzbl {{[0-9]+}}(%esp), %ecx374; X86-NEXT: shll %cl, %edi375; X86-NEXT: movswl %di, %ebp376; X86-NEXT: sarl %cl, %ebp377; X86-NEXT: xorl %ecx, %ecx378; X86-NEXT: testw %bx, %bx379; X86-NEXT: sets %cl380; X86-NEXT: addl $32767, %ecx # imm = 0x7FFF381; X86-NEXT: cmpw %bp, %bx382; X86-NEXT: movl %esi, %ebx383; X86-NEXT: cmovel %edi, %ecx384; X86-NEXT: movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill385; X86-NEXT: movzbl {{[0-9]+}}(%esp), %ecx386; X86-NEXT: shll %cl, %esi387; X86-NEXT: movswl %si, %edi388; X86-NEXT: sarl %cl, %edi389; X86-NEXT: xorl %ecx, %ecx390; X86-NEXT: testw %bx, %bx391; X86-NEXT: sets %cl392; X86-NEXT: addl $32767, %ecx # imm = 0x7FFF393; X86-NEXT: movl %ecx, %ebp394; X86-NEXT: cmpw %di, %bx395; X86-NEXT: movl %edx, %edi396; X86-NEXT: movzbl {{[0-9]+}}(%esp), %ecx397; X86-NEXT: cmovel %esi, %ebp398; X86-NEXT: shll %cl, %edx399; X86-NEXT: movswl %dx, %esi400; X86-NEXT: sarl %cl, %esi401; X86-NEXT: xorl %ebx, %ebx402; X86-NEXT: testw %di, %di403; X86-NEXT: sets %bl404; X86-NEXT: addl $32767, %ebx # imm = 0x7FFF405; X86-NEXT: cmpw %si, %di406; X86-NEXT: movl %eax, %esi407; X86-NEXT: movzbl {{[0-9]+}}(%esp), %ecx408; X86-NEXT: cmovel %edx, %ebx409; X86-NEXT: movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill410; X86-NEXT: shll %cl, %eax411; X86-NEXT: movswl %ax, %edx412; X86-NEXT: sarl %cl, %edx413; X86-NEXT: xorl %ecx, %ecx414; X86-NEXT: testw %si, %si415; X86-NEXT: sets %cl416; X86-NEXT: addl $32767, %ecx # imm = 0x7FFF417; X86-NEXT: cmpw %dx, %si418; X86-NEXT: movl {{[0-9]+}}(%esp), %edx419; X86-NEXT: movl %edx, %esi420; X86-NEXT: cmovel %eax, %ecx421; X86-NEXT: movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill422; X86-NEXT: movzbl {{[0-9]+}}(%esp), %ecx423; X86-NEXT: shll %cl, %edx424; X86-NEXT: movswl %dx, %eax425; X86-NEXT: sarl %cl, %eax426; X86-NEXT: xorl %ecx, %ecx427; X86-NEXT: testw %si, %si428; X86-NEXT: sets %cl429; X86-NEXT: addl $32767, %ecx # imm = 0x7FFF430; X86-NEXT: cmpw %ax, %si431; X86-NEXT: movl {{[0-9]+}}(%esp), %eax432; X86-NEXT: cmovel %edx, %ecx433; X86-NEXT: movl %ecx, (%esp) # 4-byte Spill434; X86-NEXT: movl %eax, %edx435; X86-NEXT: movzbl {{[0-9]+}}(%esp), %ecx436; X86-NEXT: shll %cl, %eax437; X86-NEXT: movswl %ax, %esi438; X86-NEXT: sarl %cl, %esi439; X86-NEXT: xorl %ebx, %ebx440; X86-NEXT: testw %dx, %dx441; X86-NEXT: sets %bl442; X86-NEXT: addl $32767, %ebx # imm = 0x7FFF443; X86-NEXT: cmpw %si, %dx444; X86-NEXT: cmovel %eax, %ebx445; X86-NEXT: movl {{[0-9]+}}(%esp), %eax446; X86-NEXT: movl %eax, %esi447; X86-NEXT: movzbl {{[0-9]+}}(%esp), %ecx448; X86-NEXT: shll %cl, %eax449; X86-NEXT: movswl %ax, %edi450; X86-NEXT: sarl %cl, %edi451; X86-NEXT: xorl %edx, %edx452; X86-NEXT: testw %si, %si453; X86-NEXT: sets %dl454; X86-NEXT: addl $32767, %edx # imm = 0x7FFF455; X86-NEXT: cmpw %di, %si456; X86-NEXT: cmovel %eax, %edx457; X86-NEXT: movl {{[0-9]+}}(%esp), %eax458; X86-NEXT: movl %eax, %esi459; X86-NEXT: movzbl {{[0-9]+}}(%esp), %ecx460; X86-NEXT: shll %cl, %eax461; X86-NEXT: movswl %ax, %edi462; X86-NEXT: sarl %cl, %edi463; X86-NEXT: xorl %ecx, %ecx464; X86-NEXT: testw %si, %si465; X86-NEXT: sets %cl466; X86-NEXT: addl $32767, %ecx # imm = 0x7FFF467; X86-NEXT: cmpw %di, %si468; X86-NEXT: cmovel %eax, %ecx469; X86-NEXT: movl {{[0-9]+}}(%esp), %eax470; X86-NEXT: movw %cx, 14(%eax)471; X86-NEXT: movw %dx, 12(%eax)472; X86-NEXT: movw %bx, 10(%eax)473; X86-NEXT: movl (%esp), %ecx # 4-byte Reload474; X86-NEXT: movw %cx, 8(%eax)475; X86-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload476; X86-NEXT: movw %cx, 6(%eax)477; X86-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload478; X86-NEXT: movw %cx, 4(%eax)479; X86-NEXT: movw %bp, 2(%eax)480; X86-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload481; X86-NEXT: movw %cx, (%eax)482; X86-NEXT: addl $16, %esp483; X86-NEXT: popl %esi484; X86-NEXT: popl %edi485; X86-NEXT: popl %ebx486; X86-NEXT: popl %ebp487; X86-NEXT: retl $4488 %tmp = call <8 x i16> @llvm.sshl.sat.v8i16(<8 x i16> %x, <8 x i16> %y)489 ret <8 x i16> %tmp490}491 492define <16 x i8> @vec_v16i8(<16 x i8> %x, <16 x i8> %y) nounwind {493; X64-LABEL: vec_v16i8:494; X64: # %bb.0:495; X64-NEXT: psllw $5, %xmm1496; X64-NEXT: pxor %xmm3, %xmm3497; X64-NEXT: pxor %xmm4, %xmm4498; X64-NEXT: pcmpgtb %xmm1, %xmm4499; X64-NEXT: movdqa %xmm0, %xmm2500; X64-NEXT: psllw $4, %xmm2501; X64-NEXT: pand %xmm4, %xmm2502; X64-NEXT: pandn %xmm0, %xmm4503; X64-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2504; X64-NEXT: por %xmm4, %xmm2505; X64-NEXT: punpckhbw {{.*#+}} xmm4 = xmm4[8],xmm1[8],xmm4[9],xmm1[9],xmm4[10],xmm1[10],xmm4[11],xmm1[11],xmm4[12],xmm1[12],xmm4[13],xmm1[13],xmm4[14],xmm1[14],xmm4[15],xmm1[15]506; X64-NEXT: punpcklbw {{.*#+}} xmm5 = xmm5[0],xmm1[0],xmm5[1],xmm1[1],xmm5[2],xmm1[2],xmm5[3],xmm1[3],xmm5[4],xmm1[4],xmm5[5],xmm1[5],xmm5[6],xmm1[6],xmm5[7],xmm1[7]507; X64-NEXT: paddb %xmm1, %xmm1508; X64-NEXT: pxor %xmm6, %xmm6509; X64-NEXT: pcmpgtb %xmm1, %xmm6510; X64-NEXT: movdqa %xmm6, %xmm7511; X64-NEXT: pandn %xmm2, %xmm7512; X64-NEXT: psllw $2, %xmm2513; X64-NEXT: pand %xmm6, %xmm2514; X64-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2515; X64-NEXT: por %xmm7, %xmm2516; X64-NEXT: paddb %xmm1, %xmm1517; X64-NEXT: pxor %xmm6, %xmm6518; X64-NEXT: pcmpgtb %xmm1, %xmm6519; X64-NEXT: movdqa %xmm6, %xmm1520; X64-NEXT: pandn %xmm2, %xmm1521; X64-NEXT: paddb %xmm2, %xmm2522; X64-NEXT: pand %xmm6, %xmm2523; X64-NEXT: por %xmm1, %xmm2524; X64-NEXT: punpckhbw {{.*#+}} xmm6 = xmm6[8],xmm2[8],xmm6[9],xmm2[9],xmm6[10],xmm2[10],xmm6[11],xmm2[11],xmm6[12],xmm2[12],xmm6[13],xmm2[13],xmm6[14],xmm2[14],xmm6[15],xmm2[15]525; X64-NEXT: pxor %xmm1, %xmm1526; X64-NEXT: pcmpgtw %xmm4, %xmm1527; X64-NEXT: movdqa %xmm1, %xmm7528; X64-NEXT: pandn %xmm6, %xmm7529; X64-NEXT: psraw $4, %xmm6530; X64-NEXT: pand %xmm1, %xmm6531; X64-NEXT: por %xmm7, %xmm6532; X64-NEXT: paddw %xmm4, %xmm4533; X64-NEXT: pxor %xmm1, %xmm1534; X64-NEXT: pcmpgtw %xmm4, %xmm1535; X64-NEXT: movdqa %xmm1, %xmm7536; X64-NEXT: pandn %xmm6, %xmm7537; X64-NEXT: psraw $2, %xmm6538; X64-NEXT: pand %xmm1, %xmm6539; X64-NEXT: por %xmm7, %xmm6540; X64-NEXT: paddw %xmm4, %xmm4541; X64-NEXT: pxor %xmm1, %xmm1542; X64-NEXT: pcmpgtw %xmm4, %xmm1543; X64-NEXT: movdqa %xmm1, %xmm4544; X64-NEXT: pandn %xmm6, %xmm4545; X64-NEXT: psraw $1, %xmm6546; X64-NEXT: pand %xmm1, %xmm6547; X64-NEXT: por %xmm4, %xmm6548; X64-NEXT: psrlw $8, %xmm6549; X64-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3],xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]550; X64-NEXT: pxor %xmm4, %xmm4551; X64-NEXT: pcmpgtw %xmm5, %xmm4552; X64-NEXT: movdqa %xmm4, %xmm7553; X64-NEXT: pandn %xmm1, %xmm7554; X64-NEXT: psraw $4, %xmm1555; X64-NEXT: pand %xmm4, %xmm1556; X64-NEXT: por %xmm7, %xmm1557; X64-NEXT: paddw %xmm5, %xmm5558; X64-NEXT: pxor %xmm4, %xmm4559; X64-NEXT: pcmpgtw %xmm5, %xmm4560; X64-NEXT: movdqa %xmm4, %xmm7561; X64-NEXT: pandn %xmm1, %xmm7562; X64-NEXT: psraw $2, %xmm1563; X64-NEXT: pand %xmm4, %xmm1564; X64-NEXT: por %xmm7, %xmm1565; X64-NEXT: paddw %xmm5, %xmm5566; X64-NEXT: pxor %xmm4, %xmm4567; X64-NEXT: pcmpgtw %xmm5, %xmm4568; X64-NEXT: movdqa %xmm4, %xmm5569; X64-NEXT: pandn %xmm1, %xmm5570; X64-NEXT: psraw $1, %xmm1571; X64-NEXT: pand %xmm4, %xmm1572; X64-NEXT: por %xmm5, %xmm1573; X64-NEXT: psrlw $8, %xmm1574; X64-NEXT: packuswb %xmm6, %xmm1575; X64-NEXT: pcmpeqb %xmm0, %xmm1576; X64-NEXT: pand %xmm1, %xmm2577; X64-NEXT: pcmpgtb %xmm0, %xmm3578; X64-NEXT: pandn {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3579; X64-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0580; X64-NEXT: por %xmm3, %xmm0581; X64-NEXT: pandn %xmm0, %xmm1582; X64-NEXT: por %xmm2, %xmm1583; X64-NEXT: movdqa %xmm1, %xmm0584; X64-NEXT: retq585;586; X64-AVX2-LABEL: vec_v16i8:587; X64-AVX2: # %bb.0:588; X64-AVX2-NEXT: vpsllw $5, %xmm1, %xmm1589; X64-AVX2-NEXT: vpsllw $4, %xmm0, %xmm2590; X64-AVX2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2, %xmm2591; X64-AVX2-NEXT: vpblendvb %xmm1, %xmm2, %xmm0, %xmm2592; X64-AVX2-NEXT: vpsllw $2, %xmm2, %xmm3593; X64-AVX2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3, %xmm3594; X64-AVX2-NEXT: vpaddb %xmm1, %xmm1, %xmm4595; X64-AVX2-NEXT: vpblendvb %xmm4, %xmm3, %xmm2, %xmm2596; X64-AVX2-NEXT: vpaddb %xmm2, %xmm2, %xmm3597; X64-AVX2-NEXT: vpaddb %xmm4, %xmm4, %xmm4598; X64-AVX2-NEXT: vpblendvb %xmm4, %xmm3, %xmm2, %xmm2599; X64-AVX2-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm2[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]600; X64-AVX2-NEXT: vpsraw $4, %xmm3, %xmm4601; X64-AVX2-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm1[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]602; X64-AVX2-NEXT: vpblendvb %xmm5, %xmm4, %xmm3, %xmm3603; X64-AVX2-NEXT: vpsraw $2, %xmm3, %xmm4604; X64-AVX2-NEXT: vpaddw %xmm5, %xmm5, %xmm6605; X64-AVX2-NEXT: vpblendvb %xmm6, %xmm4, %xmm3, %xmm3606; X64-AVX2-NEXT: vpsraw $1, %xmm3, %xmm4607; X64-AVX2-NEXT: vpsllw $2, %xmm5, %xmm5608; X64-AVX2-NEXT: vpblendvb %xmm5, %xmm4, %xmm3, %xmm3609; X64-AVX2-NEXT: vpsrlw $8, %xmm3, %xmm3610; X64-AVX2-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm2[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]611; X64-AVX2-NEXT: vpsraw $4, %xmm4, %xmm5612; X64-AVX2-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm1[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]613; X64-AVX2-NEXT: vpblendvb %xmm1, %xmm5, %xmm4, %xmm4614; X64-AVX2-NEXT: vpsraw $2, %xmm4, %xmm5615; X64-AVX2-NEXT: vpaddw %xmm1, %xmm1, %xmm6616; X64-AVX2-NEXT: vpblendvb %xmm6, %xmm5, %xmm4, %xmm4617; X64-AVX2-NEXT: vpsraw $1, %xmm4, %xmm5618; X64-AVX2-NEXT: vpsllw $2, %xmm1, %xmm1619; X64-AVX2-NEXT: vpblendvb %xmm1, %xmm5, %xmm4, %xmm1620; X64-AVX2-NEXT: vpsrlw $8, %xmm1, %xmm1621; X64-AVX2-NEXT: vpackuswb %xmm3, %xmm1, %xmm1622; X64-AVX2-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm1623; X64-AVX2-NEXT: vpbroadcastb {{.*#+}} xmm3 = [127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127]624; X64-AVX2-NEXT: vpblendvb %xmm0, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3, %xmm0625; X64-AVX2-NEXT: vpblendvb %xmm1, %xmm2, %xmm0, %xmm0626; X64-AVX2-NEXT: retq627;628; X86-LABEL: vec_v16i8:629; X86: # %bb.0:630; X86-NEXT: pushl %ebp631; X86-NEXT: pushl %ebx632; X86-NEXT: pushl %edi633; X86-NEXT: pushl %esi634; X86-NEXT: subl $44, %esp635; X86-NEXT: movb {{[0-9]+}}(%esp), %dh636; X86-NEXT: movb {{[0-9]+}}(%esp), %dl637; X86-NEXT: movzbl {{[0-9]+}}(%esp), %ebx638; X86-NEXT: movzbl {{[0-9]+}}(%esp), %ecx639; X86-NEXT: movb {{[0-9]+}}(%esp), %ch640; X86-NEXT: movb %ch, %bh641; X86-NEXT: shlb %cl, %bh642; X86-NEXT: movzbl %bh, %esi643; X86-NEXT: sarb %cl, %bh644; X86-NEXT: xorl %eax, %eax645; X86-NEXT: testb %ch, %ch646; X86-NEXT: sets %al647; X86-NEXT: addl $127, %eax648; X86-NEXT: cmpb %bh, %ch649; X86-NEXT: cmovel %esi, %eax650; X86-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill651; X86-NEXT: movl %ebx, %eax652; X86-NEXT: movl %edx, %ecx653; X86-NEXT: shlb %cl, %al654; X86-NEXT: movzbl %al, %esi655; X86-NEXT: sarb %cl, %al656; X86-NEXT: xorl %ecx, %ecx657; X86-NEXT: testb %bl, %bl658; X86-NEXT: sets %cl659; X86-NEXT: addl $127, %ecx660; X86-NEXT: cmpb %al, %bl661; X86-NEXT: cmovel %esi, %ecx662; X86-NEXT: movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill663; X86-NEXT: movb %dh, %al664; X86-NEXT: movzbl {{[0-9]+}}(%esp), %ecx665; X86-NEXT: shlb %cl, %al666; X86-NEXT: movzbl %al, %esi667; X86-NEXT: sarb %cl, %al668; X86-NEXT: xorl %ecx, %ecx669; X86-NEXT: testb %dh, %dh670; X86-NEXT: sets %cl671; X86-NEXT: addl $127, %ecx672; X86-NEXT: cmpb %al, %dh673; X86-NEXT: cmovel %esi, %ecx674; X86-NEXT: movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill675; X86-NEXT: movb {{[0-9]+}}(%esp), %ah676; X86-NEXT: movb %ah, %al677; X86-NEXT: movzbl {{[0-9]+}}(%esp), %ecx678; X86-NEXT: shlb %cl, %al679; X86-NEXT: movzbl %al, %esi680; X86-NEXT: sarb %cl, %al681; X86-NEXT: xorl %edx, %edx682; X86-NEXT: testb %ah, %ah683; X86-NEXT: sets %dl684; X86-NEXT: addl $127, %edx685; X86-NEXT: cmpb %al, %ah686; X86-NEXT: movzbl {{[0-9]+}}(%esp), %ecx687; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax688; X86-NEXT: cmovel %esi, %edx689; X86-NEXT: movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill690; X86-NEXT: movl %eax, %edx691; X86-NEXT: shlb %cl, %dl692; X86-NEXT: movzbl %dl, %esi693; X86-NEXT: sarb %cl, %dl694; X86-NEXT: xorl %ecx, %ecx695; X86-NEXT: testb %al, %al696; X86-NEXT: sets %cl697; X86-NEXT: addl $127, %ecx698; X86-NEXT: cmpb %dl, %al699; X86-NEXT: cmovel %esi, %ecx700; X86-NEXT: movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill701; X86-NEXT: movzbl {{[0-9]+}}(%esp), %ecx702; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax703; X86-NEXT: movl %eax, %edx704; X86-NEXT: shlb %cl, %dl705; X86-NEXT: movzbl %dl, %esi706; X86-NEXT: sarb %cl, %dl707; X86-NEXT: xorl %ecx, %ecx708; X86-NEXT: testb %al, %al709; X86-NEXT: sets %cl710; X86-NEXT: addl $127, %ecx711; X86-NEXT: cmpb %dl, %al712; X86-NEXT: cmovel %esi, %ecx713; X86-NEXT: movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill714; X86-NEXT: movzbl {{[0-9]+}}(%esp), %ecx715; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax716; X86-NEXT: movl %eax, %edx717; X86-NEXT: shlb %cl, %dl718; X86-NEXT: movzbl %dl, %esi719; X86-NEXT: sarb %cl, %dl720; X86-NEXT: xorl %ecx, %ecx721; X86-NEXT: testb %al, %al722; X86-NEXT: sets %cl723; X86-NEXT: addl $127, %ecx724; X86-NEXT: cmpb %dl, %al725; X86-NEXT: cmovel %esi, %ecx726; X86-NEXT: movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill727; X86-NEXT: movzbl {{[0-9]+}}(%esp), %ecx728; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax729; X86-NEXT: movl %eax, %edx730; X86-NEXT: shlb %cl, %dl731; X86-NEXT: movzbl %dl, %esi732; X86-NEXT: sarb %cl, %dl733; X86-NEXT: xorl %ecx, %ecx734; X86-NEXT: testb %al, %al735; X86-NEXT: sets %cl736; X86-NEXT: addl $127, %ecx737; X86-NEXT: cmpb %dl, %al738; X86-NEXT: cmovel %esi, %ecx739; X86-NEXT: movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill740; X86-NEXT: movzbl {{[0-9]+}}(%esp), %ecx741; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax742; X86-NEXT: movl %eax, %edx743; X86-NEXT: shlb %cl, %dl744; X86-NEXT: movzbl %dl, %esi745; X86-NEXT: sarb %cl, %dl746; X86-NEXT: xorl %ecx, %ecx747; X86-NEXT: testb %al, %al748; X86-NEXT: sets %cl749; X86-NEXT: addl $127, %ecx750; X86-NEXT: cmpb %dl, %al751; X86-NEXT: cmovel %esi, %ecx752; X86-NEXT: movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill753; X86-NEXT: movzbl {{[0-9]+}}(%esp), %ecx754; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax755; X86-NEXT: movl %eax, %edx756; X86-NEXT: shlb %cl, %dl757; X86-NEXT: movzbl %dl, %esi758; X86-NEXT: sarb %cl, %dl759; X86-NEXT: xorl %ecx, %ecx760; X86-NEXT: testb %al, %al761; X86-NEXT: sets %cl762; X86-NEXT: addl $127, %ecx763; X86-NEXT: cmpb %dl, %al764; X86-NEXT: cmovel %esi, %ecx765; X86-NEXT: movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill766; X86-NEXT: movzbl {{[0-9]+}}(%esp), %ecx767; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax768; X86-NEXT: movl %eax, %edx769; X86-NEXT: shlb %cl, %dl770; X86-NEXT: movzbl %dl, %esi771; X86-NEXT: sarb %cl, %dl772; X86-NEXT: xorl %ecx, %ecx773; X86-NEXT: testb %al, %al774; X86-NEXT: sets %cl775; X86-NEXT: addl $127, %ecx776; X86-NEXT: cmpb %dl, %al777; X86-NEXT: cmovel %esi, %ecx778; X86-NEXT: movl %ecx, (%esp) # 4-byte Spill779; X86-NEXT: movzbl {{[0-9]+}}(%esp), %ecx780; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax781; X86-NEXT: movl %eax, %edx782; X86-NEXT: shlb %cl, %dl783; X86-NEXT: movzbl %dl, %esi784; X86-NEXT: sarb %cl, %dl785; X86-NEXT: xorl %ecx, %ecx786; X86-NEXT: testb %al, %al787; X86-NEXT: sets %cl788; X86-NEXT: addl $127, %ecx789; X86-NEXT: cmpb %dl, %al790; X86-NEXT: cmovel %esi, %ecx791; X86-NEXT: movl %ecx, %ebp792; X86-NEXT: movzbl {{[0-9]+}}(%esp), %ecx793; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax794; X86-NEXT: movl %eax, %edx795; X86-NEXT: shlb %cl, %dl796; X86-NEXT: movzbl %dl, %esi797; X86-NEXT: sarb %cl, %dl798; X86-NEXT: xorl %ecx, %ecx799; X86-NEXT: testb %al, %al800; X86-NEXT: sets %cl801; X86-NEXT: addl $127, %ecx802; X86-NEXT: cmpb %dl, %al803; X86-NEXT: cmovel %esi, %ecx804; X86-NEXT: movl %ecx, %edi805; X86-NEXT: movzbl {{[0-9]+}}(%esp), %ecx806; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax807; X86-NEXT: movl %eax, %edx808; X86-NEXT: shlb %cl, %dl809; X86-NEXT: movzbl %dl, %esi810; X86-NEXT: sarb %cl, %dl811; X86-NEXT: xorl %ebx, %ebx812; X86-NEXT: testb %al, %al813; X86-NEXT: sets %bl814; X86-NEXT: addl $127, %ebx815; X86-NEXT: cmpb %dl, %al816; X86-NEXT: cmovel %esi, %ebx817; X86-NEXT: movzbl {{[0-9]+}}(%esp), %ecx818; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax819; X86-NEXT: movb %al, %ah820; X86-NEXT: shlb %cl, %ah821; X86-NEXT: movzbl %ah, %esi822; X86-NEXT: sarb %cl, %ah823; X86-NEXT: xorl %edx, %edx824; X86-NEXT: testb %al, %al825; X86-NEXT: sets %dl826; X86-NEXT: addl $127, %edx827; X86-NEXT: cmpb %ah, %al828; X86-NEXT: cmovel %esi, %edx829; X86-NEXT: movzbl {{[0-9]+}}(%esp), %ecx830; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax831; X86-NEXT: movb %al, %ah832; X86-NEXT: shlb %cl, %ah833; X86-NEXT: movzbl %ah, %esi834; X86-NEXT: sarb %cl, %ah835; X86-NEXT: xorl %ecx, %ecx836; X86-NEXT: testb %al, %al837; X86-NEXT: sets %cl838; X86-NEXT: addl $127, %ecx839; X86-NEXT: cmpb %ah, %al840; X86-NEXT: cmovel %esi, %ecx841; X86-NEXT: movl {{[0-9]+}}(%esp), %eax842; X86-NEXT: movb %cl, 15(%eax)843; X86-NEXT: movb %dl, 14(%eax)844; X86-NEXT: movb %bl, 13(%eax)845; X86-NEXT: movl %edi, %ecx846; X86-NEXT: movb %cl, 12(%eax)847; X86-NEXT: movl %ebp, %ecx848; X86-NEXT: movb %cl, 11(%eax)849; X86-NEXT: movl (%esp), %ecx # 4-byte Reload850; X86-NEXT: movb %cl, 10(%eax)851; X86-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload852; X86-NEXT: movb %cl, 9(%eax)853; X86-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload854; X86-NEXT: movb %cl, 8(%eax)855; X86-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload856; X86-NEXT: movb %cl, 7(%eax)857; X86-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload858; X86-NEXT: movb %cl, 6(%eax)859; X86-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload860; X86-NEXT: movb %cl, 5(%eax)861; X86-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload862; X86-NEXT: movb %cl, 4(%eax)863; X86-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload864; X86-NEXT: movb %cl, 3(%eax)865; X86-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload866; X86-NEXT: movb %cl, 2(%eax)867; X86-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload868; X86-NEXT: movb %cl, 1(%eax)869; X86-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload870; X86-NEXT: movb %cl, (%eax)871; X86-NEXT: addl $44, %esp872; X86-NEXT: popl %esi873; X86-NEXT: popl %edi874; X86-NEXT: popl %ebx875; X86-NEXT: popl %ebp876; X86-NEXT: retl $4877 %tmp = call <16 x i8> @llvm.sshl.sat.v16i8(<16 x i8> %x, <16 x i8> %y)878 ret <16 x i8> %tmp879}880