1432 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=x86_64-- | FileCheck %s3 4;5; 32-bit float to signed integer6;7 8declare <4 x i1> @llvm.fptosi.sat.v4i1.v4f32(<4 x float>)9declare <4 x i8> @llvm.fptosi.sat.v4i8.v4f32(<4 x float>)10declare <4 x i16> @llvm.fptosi.sat.v4i16.v4f32(<4 x float>)11declare <4 x i32> @llvm.fptosi.sat.v4i32.v4f32(<4 x float>)12declare <4 x i64> @llvm.fptosi.sat.v4i64.v4f32(<4 x float>)13declare <4 x i128> @llvm.fptosi.sat.v4i128.v4f32(<4 x float>)14 15define <4 x i1> @test_signed_v4i1_v4f32(<4 x float> %f) nounwind {16; CHECK-LABEL: test_signed_v4i1_v4f32:17; CHECK: # %bb.0:18; CHECK-NEXT: movaps %xmm0, %xmm119; CHECK-NEXT: shufps {{.*#+}} xmm1 = xmm1[3,3],xmm0[3,3]20; CHECK-NEXT: movss {{.*#+}} xmm2 = [-1.0E+0,0.0E+0,0.0E+0,0.0E+0]21; CHECK-NEXT: xorl %eax, %eax22; CHECK-NEXT: ucomiss %xmm1, %xmm123; CHECK-NEXT: maxss %xmm2, %xmm124; CHECK-NEXT: xorps %xmm3, %xmm325; CHECK-NEXT: minss %xmm3, %xmm126; CHECK-NEXT: cvttss2si %xmm1, %ecx27; CHECK-NEXT: cmovpl %eax, %ecx28; CHECK-NEXT: movd %ecx, %xmm129; CHECK-NEXT: movaps %xmm0, %xmm430; CHECK-NEXT: unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm0[1]31; CHECK-NEXT: ucomiss %xmm4, %xmm432; CHECK-NEXT: maxss %xmm2, %xmm433; CHECK-NEXT: minss %xmm3, %xmm434; CHECK-NEXT: cvttss2si %xmm4, %ecx35; CHECK-NEXT: cmovpl %eax, %ecx36; CHECK-NEXT: movd %ecx, %xmm437; CHECK-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm1[0],xmm4[1],xmm1[1]38; CHECK-NEXT: movaps %xmm0, %xmm139; CHECK-NEXT: maxss %xmm2, %xmm140; CHECK-NEXT: minss %xmm3, %xmm141; CHECK-NEXT: cvttss2si %xmm1, %ecx42; CHECK-NEXT: ucomiss %xmm0, %xmm043; CHECK-NEXT: cmovpl %eax, %ecx44; CHECK-NEXT: movd %ecx, %xmm145; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1,1,1]46; CHECK-NEXT: ucomiss %xmm0, %xmm047; CHECK-NEXT: maxss %xmm2, %xmm048; CHECK-NEXT: minss %xmm3, %xmm049; CHECK-NEXT: cvttss2si %xmm0, %ecx50; CHECK-NEXT: cmovpl %eax, %ecx51; CHECK-NEXT: movd %ecx, %xmm052; CHECK-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]53; CHECK-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm4[0]54; CHECK-NEXT: movdqa %xmm1, %xmm055; CHECK-NEXT: retq56 %x = call <4 x i1> @llvm.fptosi.sat.v4i1.v4f32(<4 x float> %f)57 ret <4 x i1> %x58}59 60define <4 x i1> @test_freeze_signed_v4i1_v4f32(<4 x float> %f) nounwind {61; CHECK-LABEL: test_freeze_signed_v4i1_v4f32:62; CHECK: # %bb.0:63; CHECK-NEXT: movaps %xmm0, %xmm164; CHECK-NEXT: shufps {{.*#+}} xmm1 = xmm1[3,3],xmm0[3,3]65; CHECK-NEXT: movss {{.*#+}} xmm2 = [-1.0E+0,0.0E+0,0.0E+0,0.0E+0]66; CHECK-NEXT: xorl %eax, %eax67; CHECK-NEXT: ucomiss %xmm1, %xmm168; CHECK-NEXT: maxss %xmm2, %xmm169; CHECK-NEXT: xorps %xmm3, %xmm370; CHECK-NEXT: minss %xmm3, %xmm171; CHECK-NEXT: cvttss2si %xmm1, %ecx72; CHECK-NEXT: cmovpl %eax, %ecx73; CHECK-NEXT: movd %ecx, %xmm174; CHECK-NEXT: movaps %xmm0, %xmm475; CHECK-NEXT: unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm0[1]76; CHECK-NEXT: ucomiss %xmm4, %xmm477; CHECK-NEXT: maxss %xmm2, %xmm478; CHECK-NEXT: minss %xmm3, %xmm479; CHECK-NEXT: cvttss2si %xmm4, %ecx80; CHECK-NEXT: cmovpl %eax, %ecx81; CHECK-NEXT: movd %ecx, %xmm482; CHECK-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm1[0],xmm4[1],xmm1[1]83; CHECK-NEXT: movaps %xmm0, %xmm184; CHECK-NEXT: maxss %xmm2, %xmm185; CHECK-NEXT: minss %xmm3, %xmm186; CHECK-NEXT: cvttss2si %xmm1, %ecx87; CHECK-NEXT: ucomiss %xmm0, %xmm088; CHECK-NEXT: cmovpl %eax, %ecx89; CHECK-NEXT: movd %ecx, %xmm190; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1,1,1]91; CHECK-NEXT: ucomiss %xmm0, %xmm092; CHECK-NEXT: maxss %xmm2, %xmm093; CHECK-NEXT: minss %xmm3, %xmm094; CHECK-NEXT: cvttss2si %xmm0, %ecx95; CHECK-NEXT: cmovpl %eax, %ecx96; CHECK-NEXT: movd %ecx, %xmm097; CHECK-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]98; CHECK-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm4[0]99; CHECK-NEXT: movdqa %xmm1, %xmm0100; CHECK-NEXT: retq101 %x = call <4 x i1> @llvm.fptosi.sat.v4i1.v4f32(<4 x float> %f)102 %y = freeze <4 x i1> %x103 ret <4 x i1> %y104}105 106define <4 x i8> @test_signed_v4i8_v4f32(<4 x float> %f) nounwind {107; CHECK-LABEL: test_signed_v4i8_v4f32:108; CHECK: # %bb.0:109; CHECK-NEXT: movss {{.*#+}} xmm1 = [-1.28E+2,0.0E+0,0.0E+0,0.0E+0]110; CHECK-NEXT: movaps %xmm1, %xmm3111; CHECK-NEXT: maxss %xmm0, %xmm3112; CHECK-NEXT: movss {{.*#+}} xmm2 = [1.27E+2,0.0E+0,0.0E+0,0.0E+0]113; CHECK-NEXT: movaps %xmm2, %xmm4114; CHECK-NEXT: minss %xmm3, %xmm4115; CHECK-NEXT: cvttss2si %xmm4, %eax116; CHECK-NEXT: movzbl %al, %eax117; CHECK-NEXT: movaps %xmm0, %xmm3118; CHECK-NEXT: shufps {{.*#+}} xmm3 = xmm3[1,1],xmm0[1,1]119; CHECK-NEXT: movaps %xmm1, %xmm4120; CHECK-NEXT: maxss %xmm3, %xmm4121; CHECK-NEXT: movaps %xmm2, %xmm3122; CHECK-NEXT: minss %xmm4, %xmm3123; CHECK-NEXT: cvttss2si %xmm3, %ecx124; CHECK-NEXT: movzbl %cl, %ecx125; CHECK-NEXT: shll $8, %ecx126; CHECK-NEXT: orl %eax, %ecx127; CHECK-NEXT: movaps %xmm0, %xmm3128; CHECK-NEXT: unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm0[1]129; CHECK-NEXT: movaps %xmm1, %xmm4130; CHECK-NEXT: maxss %xmm3, %xmm4131; CHECK-NEXT: movaps %xmm2, %xmm3132; CHECK-NEXT: minss %xmm4, %xmm3133; CHECK-NEXT: cvttss2si %xmm3, %eax134; CHECK-NEXT: movzbl %al, %eax135; CHECK-NEXT: shll $16, %eax136; CHECK-NEXT: orl %ecx, %eax137; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3,3,3]138; CHECK-NEXT: maxss %xmm0, %xmm1139; CHECK-NEXT: minss %xmm1, %xmm2140; CHECK-NEXT: cvttss2si %xmm2, %ecx141; CHECK-NEXT: shll $24, %ecx142; CHECK-NEXT: orl %eax, %ecx143; CHECK-NEXT: movd %ecx, %xmm0144; CHECK-NEXT: retq145 %x = call <4 x i8> @llvm.fptosi.sat.v4i8.v4f32(<4 x float> %f)146 ret <4 x i8> %x147}148 149define <4 x i16> @test_signed_v4i16_v4f32(<4 x float> %f) nounwind {150; CHECK-LABEL: test_signed_v4i16_v4f32:151; CHECK: # %bb.0:152; CHECK-NEXT: movaps %xmm0, %xmm1153; CHECK-NEXT: shufps {{.*#+}} xmm1 = xmm1[1,1],xmm0[1,1]154; CHECK-NEXT: movss {{.*#+}} xmm2 = [-3.2768E+4,0.0E+0,0.0E+0,0.0E+0]155; CHECK-NEXT: movaps %xmm2, %xmm3156; CHECK-NEXT: maxss %xmm1, %xmm3157; CHECK-NEXT: movss {{.*#+}} xmm4 = [3.2767E+4,0.0E+0,0.0E+0,0.0E+0]158; CHECK-NEXT: movaps %xmm4, %xmm1159; CHECK-NEXT: minss %xmm3, %xmm1160; CHECK-NEXT: cvttss2si %xmm1, %eax161; CHECK-NEXT: movaps %xmm2, %xmm1162; CHECK-NEXT: maxss %xmm0, %xmm1163; CHECK-NEXT: movaps %xmm4, %xmm3164; CHECK-NEXT: minss %xmm1, %xmm3165; CHECK-NEXT: cvttss2si %xmm3, %ecx166; CHECK-NEXT: movd %ecx, %xmm1167; CHECK-NEXT: pinsrw $1, %eax, %xmm1168; CHECK-NEXT: movaps %xmm0, %xmm3169; CHECK-NEXT: unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm0[1]170; CHECK-NEXT: movaps %xmm2, %xmm5171; CHECK-NEXT: maxss %xmm3, %xmm5172; CHECK-NEXT: movaps %xmm4, %xmm3173; CHECK-NEXT: minss %xmm5, %xmm3174; CHECK-NEXT: cvttss2si %xmm3, %eax175; CHECK-NEXT: pinsrw $2, %eax, %xmm1176; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3,3,3]177; CHECK-NEXT: maxss %xmm0, %xmm2178; CHECK-NEXT: minss %xmm2, %xmm4179; CHECK-NEXT: cvttss2si %xmm4, %eax180; CHECK-NEXT: pinsrw $3, %eax, %xmm1181; CHECK-NEXT: movdqa %xmm1, %xmm0182; CHECK-NEXT: retq183 %x = call <4 x i16> @llvm.fptosi.sat.v4i16.v4f32(<4 x float> %f)184 ret <4 x i16> %x185}186 187define <4 x i32> @test_signed_v4i32_v4f32(<4 x float> %f) nounwind {188; CHECK-LABEL: test_signed_v4i32_v4f32:189; CHECK: # %bb.0:190; CHECK-NEXT: movaps %xmm0, %xmm1191; CHECK-NEXT: shufps {{.*#+}} xmm1 = xmm1[3,3],xmm0[3,3]192; CHECK-NEXT: cvttss2si %xmm1, %edx193; CHECK-NEXT: movss {{.*#+}} xmm2 = [2.14748352E+9,0.0E+0,0.0E+0,0.0E+0]194; CHECK-NEXT: ucomiss %xmm2, %xmm1195; CHECK-NEXT: movl $2147483647, %eax # imm = 0x7FFFFFFF196; CHECK-NEXT: cmoval %eax, %edx197; CHECK-NEXT: xorl %ecx, %ecx198; CHECK-NEXT: ucomiss %xmm1, %xmm1199; CHECK-NEXT: cmovpl %ecx, %edx200; CHECK-NEXT: movd %edx, %xmm1201; CHECK-NEXT: movaps %xmm0, %xmm3202; CHECK-NEXT: unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm0[1]203; CHECK-NEXT: cvttss2si %xmm3, %edx204; CHECK-NEXT: ucomiss %xmm2, %xmm3205; CHECK-NEXT: cmoval %eax, %edx206; CHECK-NEXT: ucomiss %xmm3, %xmm3207; CHECK-NEXT: cmovpl %ecx, %edx208; CHECK-NEXT: movd %edx, %xmm3209; CHECK-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1]210; CHECK-NEXT: cvttss2si %xmm0, %edx211; CHECK-NEXT: ucomiss %xmm2, %xmm0212; CHECK-NEXT: cmoval %eax, %edx213; CHECK-NEXT: ucomiss %xmm0, %xmm0214; CHECK-NEXT: cmovpl %ecx, %edx215; CHECK-NEXT: movd %edx, %xmm1216; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1,1,1]217; CHECK-NEXT: cvttss2si %xmm0, %edx218; CHECK-NEXT: ucomiss %xmm2, %xmm0219; CHECK-NEXT: cmoval %eax, %edx220; CHECK-NEXT: ucomiss %xmm0, %xmm0221; CHECK-NEXT: cmovpl %ecx, %edx222; CHECK-NEXT: movd %edx, %xmm0223; CHECK-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]224; CHECK-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm3[0]225; CHECK-NEXT: movdqa %xmm1, %xmm0226; CHECK-NEXT: retq227 %x = call <4 x i32> @llvm.fptosi.sat.v4i32.v4f32(<4 x float> %f)228 ret <4 x i32> %x229}230 231define <4 x i64> @test_signed_v4i64_v4f32(<4 x float> %f) nounwind {232; CHECK-LABEL: test_signed_v4i64_v4f32:233; CHECK: # %bb.0:234; CHECK-NEXT: cvttss2si %xmm0, %rdx235; CHECK-NEXT: movss {{.*#+}} xmm1 = [9.22337149E+18,0.0E+0,0.0E+0,0.0E+0]236; CHECK-NEXT: ucomiss %xmm1, %xmm0237; CHECK-NEXT: movabsq $9223372036854775807, %rax # imm = 0x7FFFFFFFFFFFFFFF238; CHECK-NEXT: cmovaq %rax, %rdx239; CHECK-NEXT: xorl %ecx, %ecx240; CHECK-NEXT: ucomiss %xmm0, %xmm0241; CHECK-NEXT: cmovpq %rcx, %rdx242; CHECK-NEXT: movq %rdx, %xmm2243; CHECK-NEXT: movaps %xmm0, %xmm3244; CHECK-NEXT: shufps {{.*#+}} xmm3 = xmm3[1,1],xmm0[1,1]245; CHECK-NEXT: cvttss2si %xmm3, %rdx246; CHECK-NEXT: ucomiss %xmm1, %xmm3247; CHECK-NEXT: cmovaq %rax, %rdx248; CHECK-NEXT: ucomiss %xmm3, %xmm3249; CHECK-NEXT: cmovpq %rcx, %rdx250; CHECK-NEXT: movq %rdx, %xmm3251; CHECK-NEXT: punpcklqdq {{.*#+}} xmm2 = xmm2[0],xmm3[0]252; CHECK-NEXT: movaps %xmm0, %xmm3253; CHECK-NEXT: shufps {{.*#+}} xmm3 = xmm3[3,3],xmm0[3,3]254; CHECK-NEXT: cvttss2si %xmm3, %rdx255; CHECK-NEXT: ucomiss %xmm1, %xmm3256; CHECK-NEXT: cmovaq %rax, %rdx257; CHECK-NEXT: ucomiss %xmm3, %xmm3258; CHECK-NEXT: cmovpq %rcx, %rdx259; CHECK-NEXT: movq %rdx, %xmm3260; CHECK-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1]261; CHECK-NEXT: cvttss2si %xmm0, %rdx262; CHECK-NEXT: ucomiss %xmm1, %xmm0263; CHECK-NEXT: cmovaq %rax, %rdx264; CHECK-NEXT: ucomiss %xmm0, %xmm0265; CHECK-NEXT: cmovpq %rcx, %rdx266; CHECK-NEXT: movq %rdx, %xmm1267; CHECK-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm3[0]268; CHECK-NEXT: movdqa %xmm2, %xmm0269; CHECK-NEXT: retq270 %x = call <4 x i64> @llvm.fptosi.sat.v4i64.v4f32(<4 x float> %f)271 ret <4 x i64> %x272}273 274define <4 x i128> @test_signed_v4i128_v4f32(<4 x float> %f) nounwind {275; CHECK-LABEL: test_signed_v4i128_v4f32:276; CHECK: # %bb.0:277; CHECK-NEXT: pushq %rbp278; CHECK-NEXT: pushq %r15279; CHECK-NEXT: pushq %r14280; CHECK-NEXT: pushq %r13281; CHECK-NEXT: pushq %r12282; CHECK-NEXT: pushq %rbx283; CHECK-NEXT: subq $56, %rsp284; CHECK-NEXT: movaps %xmm0, (%rsp) # 16-byte Spill285; CHECK-NEXT: movq %rdi, %rbx286; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1,1,1]287; CHECK-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill288; CHECK-NEXT: callq __fixsfti@PLT289; CHECK-NEXT: movq %rdx, %r15290; CHECK-NEXT: xorl %r14d, %r14d291; CHECK-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload292; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0293; CHECK-NEXT: cmovbq %r14, %rax294; CHECK-NEXT: movabsq $-9223372036854775808, %rcx # imm = 0x8000000000000000295; CHECK-NEXT: cmovbq %rcx, %r15296; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0297; CHECK-NEXT: movabsq $9223372036854775807, %rbp # imm = 0x7FFFFFFFFFFFFFFF298; CHECK-NEXT: cmovaq %rbp, %r15299; CHECK-NEXT: movq $-1, %rcx300; CHECK-NEXT: cmovaq %rcx, %rax301; CHECK-NEXT: ucomiss %xmm0, %xmm0302; CHECK-NEXT: cmovpq %r14, %rax303; CHECK-NEXT: movq %rax, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill304; CHECK-NEXT: cmovpq %r14, %r15305; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload306; CHECK-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1]307; CHECK-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill308; CHECK-NEXT: callq __fixsfti@PLT309; CHECK-NEXT: movq %rax, %r12310; CHECK-NEXT: movq %rdx, %r13311; CHECK-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload312; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0313; CHECK-NEXT: cmovbq %r14, %r12314; CHECK-NEXT: movabsq $-9223372036854775808, %rax # imm = 0x8000000000000000315; CHECK-NEXT: cmovbq %rax, %r13316; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0317; CHECK-NEXT: cmovaq %rbp, %r13318; CHECK-NEXT: movq $-1, %rax319; CHECK-NEXT: cmovaq %rax, %r12320; CHECK-NEXT: ucomiss %xmm0, %xmm0321; CHECK-NEXT: cmovpq %r14, %r12322; CHECK-NEXT: cmovpq %r14, %r13323; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload324; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3,3,3]325; CHECK-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill326; CHECK-NEXT: callq __fixsfti@PLT327; CHECK-NEXT: movq %rax, %rbp328; CHECK-NEXT: movq %rdx, %r14329; CHECK-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload330; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0331; CHECK-NEXT: movl $0, %eax332; CHECK-NEXT: cmovbq %rax, %rbp333; CHECK-NEXT: movabsq $-9223372036854775808, %rcx # imm = 0x8000000000000000334; CHECK-NEXT: cmovbq %rcx, %r14335; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0336; CHECK-NEXT: movabsq $9223372036854775807, %rcx # imm = 0x7FFFFFFFFFFFFFFF337; CHECK-NEXT: cmovaq %rcx, %r14338; CHECK-NEXT: movq $-1, %rcx339; CHECK-NEXT: cmovaq %rcx, %rbp340; CHECK-NEXT: ucomiss %xmm0, %xmm0341; CHECK-NEXT: cmovpq %rax, %rbp342; CHECK-NEXT: cmovpq %rax, %r14343; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload344; CHECK-NEXT: callq __fixsfti@PLT345; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload346; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0347; CHECK-NEXT: movl $0, %esi348; CHECK-NEXT: cmovbq %rsi, %rax349; CHECK-NEXT: movabsq $-9223372036854775808, %rcx # imm = 0x8000000000000000350; CHECK-NEXT: cmovbq %rcx, %rdx351; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0352; CHECK-NEXT: movabsq $9223372036854775807, %rcx # imm = 0x7FFFFFFFFFFFFFFF353; CHECK-NEXT: cmovaq %rcx, %rdx354; CHECK-NEXT: movq $-1, %rcx355; CHECK-NEXT: cmovaq %rcx, %rax356; CHECK-NEXT: ucomiss %xmm0, %xmm0357; CHECK-NEXT: cmovpq %rsi, %rax358; CHECK-NEXT: movl $0, %ecx359; CHECK-NEXT: cmovpq %rcx, %rdx360; CHECK-NEXT: movq %rdx, 8(%rbx)361; CHECK-NEXT: movq %rax, (%rbx)362; CHECK-NEXT: movq %r14, 56(%rbx)363; CHECK-NEXT: movq %rbp, 48(%rbx)364; CHECK-NEXT: movq %r13, 40(%rbx)365; CHECK-NEXT: movq %r12, 32(%rbx)366; CHECK-NEXT: movq %r15, 24(%rbx)367; CHECK-NEXT: movq {{[-0-9]+}}(%r{{[sb]}}p), %rax # 8-byte Reload368; CHECK-NEXT: movq %rax, 16(%rbx)369; CHECK-NEXT: movq %rbx, %rax370; CHECK-NEXT: addq $56, %rsp371; CHECK-NEXT: popq %rbx372; CHECK-NEXT: popq %r12373; CHECK-NEXT: popq %r13374; CHECK-NEXT: popq %r14375; CHECK-NEXT: popq %r15376; CHECK-NEXT: popq %rbp377; CHECK-NEXT: retq378 %x = call <4 x i128> @llvm.fptosi.sat.v4i128.v4f32(<4 x float> %f)379 ret <4 x i128> %x380}381 382;383; 64-bit float to signed integer384;385 386declare <2 x i1> @llvm.fptosi.sat.v2i1.v2f64(<2 x double>)387declare <2 x i8> @llvm.fptosi.sat.v2i8.v2f64(<2 x double>)388declare <2 x i16> @llvm.fptosi.sat.v2i16.v2f64(<2 x double>)389declare <2 x i32> @llvm.fptosi.sat.v2i32.v2f64(<2 x double>)390declare <2 x i64> @llvm.fptosi.sat.v2i64.v2f64(<2 x double>)391declare <2 x i128> @llvm.fptosi.sat.v2i128.v2f64(<2 x double>)392 393define <2 x i1> @test_signed_v2i1_v2f64(<2 x double> %f) nounwind {394; CHECK-LABEL: test_signed_v2i1_v2f64:395; CHECK: # %bb.0:396; CHECK-NEXT: movsd {{.*#+}} xmm2 = [-1.0E+0,0.0E+0]397; CHECK-NEXT: movapd %xmm0, %xmm1398; CHECK-NEXT: maxsd %xmm2, %xmm1399; CHECK-NEXT: xorpd %xmm3, %xmm3400; CHECK-NEXT: minsd %xmm3, %xmm1401; CHECK-NEXT: cvttsd2si %xmm1, %rax402; CHECK-NEXT: xorl %ecx, %ecx403; CHECK-NEXT: ucomisd %xmm0, %xmm0404; CHECK-NEXT: cmovpq %rcx, %rax405; CHECK-NEXT: movq %rax, %xmm1406; CHECK-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1,1]407; CHECK-NEXT: ucomisd %xmm0, %xmm0408; CHECK-NEXT: maxsd %xmm2, %xmm0409; CHECK-NEXT: minsd %xmm3, %xmm0410; CHECK-NEXT: cvttsd2si %xmm0, %rax411; CHECK-NEXT: cmovpq %rcx, %rax412; CHECK-NEXT: movq %rax, %xmm0413; CHECK-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0]414; CHECK-NEXT: movdqa %xmm1, %xmm0415; CHECK-NEXT: retq416 %x = call <2 x i1> @llvm.fptosi.sat.v2i1.v2f64(<2 x double> %f)417 ret <2 x i1> %x418}419 420define <2 x i8> @test_signed_v2i8_v2f64(<2 x double> %f) nounwind {421; CHECK-LABEL: test_signed_v2i8_v2f64:422; CHECK: # %bb.0:423; CHECK-NEXT: movsd {{.*#+}} xmm1 = [-1.28E+2,0.0E+0]424; CHECK-NEXT: movapd %xmm1, %xmm2425; CHECK-NEXT: maxsd %xmm0, %xmm2426; CHECK-NEXT: movsd {{.*#+}} xmm3 = [1.27E+2,0.0E+0]427; CHECK-NEXT: movapd %xmm3, %xmm4428; CHECK-NEXT: minsd %xmm2, %xmm4429; CHECK-NEXT: cvttsd2si %xmm4, %eax430; CHECK-NEXT: movzbl %al, %eax431; CHECK-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1,1]432; CHECK-NEXT: maxsd %xmm0, %xmm1433; CHECK-NEXT: minsd %xmm1, %xmm3434; CHECK-NEXT: cvttsd2si %xmm3, %ecx435; CHECK-NEXT: shll $8, %ecx436; CHECK-NEXT: orl %eax, %ecx437; CHECK-NEXT: movd %ecx, %xmm0438; CHECK-NEXT: retq439 %x = call <2 x i8> @llvm.fptosi.sat.v2i8.v2f64(<2 x double> %f)440 ret <2 x i8> %x441}442 443define <2 x i16> @test_signed_v2i16_v2f64(<2 x double> %f) nounwind {444; CHECK-LABEL: test_signed_v2i16_v2f64:445; CHECK: # %bb.0:446; CHECK-NEXT: movsd {{.*#+}} xmm1 = [-3.2768E+4,0.0E+0]447; CHECK-NEXT: movapd %xmm1, %xmm2448; CHECK-NEXT: maxsd %xmm0, %xmm1449; CHECK-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1,1]450; CHECK-NEXT: maxsd %xmm0, %xmm2451; CHECK-NEXT: movsd {{.*#+}} xmm0 = [3.2767E+4,0.0E+0]452; CHECK-NEXT: movapd %xmm0, %xmm3453; CHECK-NEXT: minsd %xmm2, %xmm3454; CHECK-NEXT: cvttsd2si %xmm3, %eax455; CHECK-NEXT: minsd %xmm1, %xmm0456; CHECK-NEXT: cvttsd2si %xmm0, %ecx457; CHECK-NEXT: movd %ecx, %xmm0458; CHECK-NEXT: pinsrw $1, %eax, %xmm0459; CHECK-NEXT: retq460 %x = call <2 x i16> @llvm.fptosi.sat.v2i16.v2f64(<2 x double> %f)461 ret <2 x i16> %x462}463 464define <2 x i32> @test_signed_v2i32_v2f64(<2 x double> %f) nounwind {465; CHECK-LABEL: test_signed_v2i32_v2f64:466; CHECK: # %bb.0:467; CHECK-NEXT: movsd {{.*#+}} xmm2 = [-2.147483648E+9,0.0E+0]468; CHECK-NEXT: movapd %xmm0, %xmm1469; CHECK-NEXT: maxsd %xmm2, %xmm1470; CHECK-NEXT: movsd {{.*#+}} xmm3 = [2.147483647E+9,0.0E+0]471; CHECK-NEXT: minsd %xmm3, %xmm1472; CHECK-NEXT: cvttsd2si %xmm1, %eax473; CHECK-NEXT: xorl %ecx, %ecx474; CHECK-NEXT: ucomisd %xmm0, %xmm0475; CHECK-NEXT: cmovpl %ecx, %eax476; CHECK-NEXT: movd %eax, %xmm1477; CHECK-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1,1]478; CHECK-NEXT: ucomisd %xmm0, %xmm0479; CHECK-NEXT: maxsd %xmm2, %xmm0480; CHECK-NEXT: minsd %xmm3, %xmm0481; CHECK-NEXT: cvttsd2si %xmm0, %eax482; CHECK-NEXT: cmovpl %ecx, %eax483; CHECK-NEXT: movd %eax, %xmm0484; CHECK-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]485; CHECK-NEXT: movdqa %xmm1, %xmm0486; CHECK-NEXT: retq487 %x = call <2 x i32> @llvm.fptosi.sat.v2i32.v2f64(<2 x double> %f)488 ret <2 x i32> %x489}490 491define <2 x i64> @test_signed_v2i64_v2f64(<2 x double> %f) nounwind {492; CHECK-LABEL: test_signed_v2i64_v2f64:493; CHECK: # %bb.0:494; CHECK-NEXT: cvttsd2si %xmm0, %rax495; CHECK-NEXT: movsd {{.*#+}} xmm2 = [9.2233720368547748E+18,0.0E+0]496; CHECK-NEXT: ucomisd %xmm2, %xmm0497; CHECK-NEXT: movabsq $9223372036854775807, %rcx # imm = 0x7FFFFFFFFFFFFFFF498; CHECK-NEXT: cmovaq %rcx, %rax499; CHECK-NEXT: xorl %edx, %edx500; CHECK-NEXT: ucomisd %xmm0, %xmm0501; CHECK-NEXT: cmovpq %rdx, %rax502; CHECK-NEXT: movq %rax, %xmm1503; CHECK-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1,1]504; CHECK-NEXT: cvttsd2si %xmm0, %rax505; CHECK-NEXT: ucomisd %xmm2, %xmm0506; CHECK-NEXT: cmovaq %rcx, %rax507; CHECK-NEXT: ucomisd %xmm0, %xmm0508; CHECK-NEXT: cmovpq %rdx, %rax509; CHECK-NEXT: movq %rax, %xmm0510; CHECK-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0]511; CHECK-NEXT: movdqa %xmm1, %xmm0512; CHECK-NEXT: retq513 %x = call <2 x i64> @llvm.fptosi.sat.v2i64.v2f64(<2 x double> %f)514 ret <2 x i64> %x515}516 517define <2 x i128> @test_signed_v2i128_v2f64(<2 x double> %f) nounwind {518; CHECK-LABEL: test_signed_v2i128_v2f64:519; CHECK: # %bb.0:520; CHECK-NEXT: pushq %rbp521; CHECK-NEXT: pushq %r15522; CHECK-NEXT: pushq %r14523; CHECK-NEXT: pushq %r13524; CHECK-NEXT: pushq %r12525; CHECK-NEXT: pushq %rbx526; CHECK-NEXT: subq $40, %rsp527; CHECK-NEXT: movaps %xmm0, (%rsp) # 16-byte Spill528; CHECK-NEXT: movq %rdi, %rbx529; CHECK-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1]530; CHECK-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill531; CHECK-NEXT: callq __fixdfti@PLT532; CHECK-NEXT: movq %rax, %r14533; CHECK-NEXT: movq %rdx, %r15534; CHECK-NEXT: xorl %r12d, %r12d535; CHECK-NEXT: movapd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload536; CHECK-NEXT: ucomisd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0537; CHECK-NEXT: cmovbq %r12, %r14538; CHECK-NEXT: movabsq $-9223372036854775808, %rax # imm = 0x8000000000000000539; CHECK-NEXT: cmovbq %rax, %r15540; CHECK-NEXT: ucomisd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0541; CHECK-NEXT: movabsq $9223372036854775807, %rbp # imm = 0x7FFFFFFFFFFFFFFF542; CHECK-NEXT: cmovaq %rbp, %r15543; CHECK-NEXT: movq $-1, %r13544; CHECK-NEXT: cmovaq %r13, %r14545; CHECK-NEXT: ucomisd %xmm0, %xmm0546; CHECK-NEXT: cmovpq %r12, %r14547; CHECK-NEXT: cmovpq %r12, %r15548; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload549; CHECK-NEXT: callq __fixdfti@PLT550; CHECK-NEXT: movapd (%rsp), %xmm0 # 16-byte Reload551; CHECK-NEXT: ucomisd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0552; CHECK-NEXT: cmovbq %r12, %rax553; CHECK-NEXT: movabsq $-9223372036854775808, %rcx # imm = 0x8000000000000000554; CHECK-NEXT: cmovbq %rcx, %rdx555; CHECK-NEXT: ucomisd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0556; CHECK-NEXT: cmovaq %rbp, %rdx557; CHECK-NEXT: cmovaq %r13, %rax558; CHECK-NEXT: ucomisd %xmm0, %xmm0559; CHECK-NEXT: cmovpq %r12, %rax560; CHECK-NEXT: cmovpq %r12, %rdx561; CHECK-NEXT: movq %rdx, 8(%rbx)562; CHECK-NEXT: movq %rax, (%rbx)563; CHECK-NEXT: movq %r15, 24(%rbx)564; CHECK-NEXT: movq %r14, 16(%rbx)565; CHECK-NEXT: movq %rbx, %rax566; CHECK-NEXT: addq $40, %rsp567; CHECK-NEXT: popq %rbx568; CHECK-NEXT: popq %r12569; CHECK-NEXT: popq %r13570; CHECK-NEXT: popq %r14571; CHECK-NEXT: popq %r15572; CHECK-NEXT: popq %rbp573; CHECK-NEXT: retq574 %x = call <2 x i128> @llvm.fptosi.sat.v2i128.v2f64(<2 x double> %f)575 ret <2 x i128> %x576}577 578;579; 16-bit float to signed integer580;581 582declare <8 x i1> @llvm.fptosi.sat.v8i1.v8f16(<8 x half>)583declare <8 x i8> @llvm.fptosi.sat.v8i8.v8f16(<8 x half>)584declare <8 x i16> @llvm.fptosi.sat.v8i16.v8f16(<8 x half>)585declare <8 x i32> @llvm.fptosi.sat.v8i32.v8f16(<8 x half>)586declare <8 x i64> @llvm.fptosi.sat.v8i64.v8f16(<8 x half>)587declare <8 x i128> @llvm.fptosi.sat.v8i128.v8f16(<8 x half>)588 589define <8 x i1> @test_signed_v8i1_v8f16(<8 x half> %f) nounwind {590; CHECK-LABEL: test_signed_v8i1_v8f16:591; CHECK: # %bb.0:592; CHECK-NEXT: pushq %rbp593; CHECK-NEXT: pushq %rbx594; CHECK-NEXT: subq $72, %rsp595; CHECK-NEXT: movdqa %xmm0, (%rsp) # 16-byte Spill596; CHECK-NEXT: psrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero597; CHECK-NEXT: callq __extendhfsf2@PLT598; CHECK-NEXT: cvttss2si %xmm0, %eax599; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0600; CHECK-NEXT: movl $65535, %ebp # imm = 0xFFFF601; CHECK-NEXT: cmovbl %ebp, %eax602; CHECK-NEXT: xorl %ebx, %ebx603; CHECK-NEXT: xorps %xmm1, %xmm1604; CHECK-NEXT: ucomiss %xmm1, %xmm0605; CHECK-NEXT: cmoval %ebx, %eax606; CHECK-NEXT: ucomiss %xmm0, %xmm0607; CHECK-NEXT: cmovpl %ebx, %eax608; CHECK-NEXT: movd %eax, %xmm0609; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill610; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload611; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3,3,3]612; CHECK-NEXT: callq __extendhfsf2@PLT613; CHECK-NEXT: cvttss2si %xmm0, %eax614; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0615; CHECK-NEXT: cmovbl %ebp, %eax616; CHECK-NEXT: xorps %xmm1, %xmm1617; CHECK-NEXT: ucomiss %xmm1, %xmm0618; CHECK-NEXT: cmoval %ebx, %eax619; CHECK-NEXT: ucomiss %xmm0, %xmm0620; CHECK-NEXT: cmovpl %ebx, %eax621; CHECK-NEXT: movd %eax, %xmm0622; CHECK-NEXT: punpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload623; CHECK-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]624; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill625; CHECK-NEXT: movdqa (%rsp), %xmm0 # 16-byte Reload626; CHECK-NEXT: psrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero627; CHECK-NEXT: callq __extendhfsf2@PLT628; CHECK-NEXT: cvttss2si %xmm0, %eax629; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0630; CHECK-NEXT: cmovbl %ebp, %eax631; CHECK-NEXT: xorps %xmm1, %xmm1632; CHECK-NEXT: ucomiss %xmm1, %xmm0633; CHECK-NEXT: cmoval %ebx, %eax634; CHECK-NEXT: ucomiss %xmm0, %xmm0635; CHECK-NEXT: cmovpl %ebx, %eax636; CHECK-NEXT: movd %eax, %xmm0637; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill638; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload639; CHECK-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1]640; CHECK-NEXT: callq __extendhfsf2@PLT641; CHECK-NEXT: cvttss2si %xmm0, %eax642; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0643; CHECK-NEXT: cmovbl %ebp, %eax644; CHECK-NEXT: xorps %xmm1, %xmm1645; CHECK-NEXT: ucomiss %xmm1, %xmm0646; CHECK-NEXT: cmoval %ebx, %eax647; CHECK-NEXT: ucomiss %xmm0, %xmm0648; CHECK-NEXT: cmovpl %ebx, %eax649; CHECK-NEXT: movd %eax, %xmm0650; CHECK-NEXT: punpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload651; CHECK-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]652; CHECK-NEXT: punpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload653; CHECK-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]654; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill655; CHECK-NEXT: movdqa (%rsp), %xmm0 # 16-byte Reload656; CHECK-NEXT: psrlq $48, %xmm0657; CHECK-NEXT: callq __extendhfsf2@PLT658; CHECK-NEXT: cvttss2si %xmm0, %eax659; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0660; CHECK-NEXT: cmovbl %ebp, %eax661; CHECK-NEXT: xorps %xmm1, %xmm1662; CHECK-NEXT: ucomiss %xmm1, %xmm0663; CHECK-NEXT: cmoval %ebx, %eax664; CHECK-NEXT: ucomiss %xmm0, %xmm0665; CHECK-NEXT: cmovpl %ebx, %eax666; CHECK-NEXT: movd %eax, %xmm0667; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill668; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload669; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1,1,1]670; CHECK-NEXT: callq __extendhfsf2@PLT671; CHECK-NEXT: cvttss2si %xmm0, %eax672; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0673; CHECK-NEXT: cmovbl %ebp, %eax674; CHECK-NEXT: xorps %xmm1, %xmm1675; CHECK-NEXT: ucomiss %xmm1, %xmm0676; CHECK-NEXT: cmoval %ebx, %eax677; CHECK-NEXT: ucomiss %xmm0, %xmm0678; CHECK-NEXT: cmovpl %ebx, %eax679; CHECK-NEXT: movd %eax, %xmm0680; CHECK-NEXT: punpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload681; CHECK-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]682; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill683; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload684; CHECK-NEXT: callq __extendhfsf2@PLT685; CHECK-NEXT: cvttss2si %xmm0, %eax686; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0687; CHECK-NEXT: cmovbl %ebp, %eax688; CHECK-NEXT: xorps %xmm1, %xmm1689; CHECK-NEXT: ucomiss %xmm1, %xmm0690; CHECK-NEXT: cmoval %ebx, %eax691; CHECK-NEXT: ucomiss %xmm0, %xmm0692; CHECK-NEXT: cmovpl %ebx, %eax693; CHECK-NEXT: movd %eax, %xmm0694; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill695; CHECK-NEXT: movdqa (%rsp), %xmm0 # 16-byte Reload696; CHECK-NEXT: psrld $16, %xmm0697; CHECK-NEXT: callq __extendhfsf2@PLT698; CHECK-NEXT: cvttss2si %xmm0, %eax699; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0700; CHECK-NEXT: cmovbl %ebp, %eax701; CHECK-NEXT: xorps %xmm1, %xmm1702; CHECK-NEXT: ucomiss %xmm1, %xmm0703; CHECK-NEXT: cmoval %ebx, %eax704; CHECK-NEXT: ucomiss %xmm0, %xmm0705; CHECK-NEXT: cmovpl %ebx, %eax706; CHECK-NEXT: movd %eax, %xmm1707; CHECK-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload708; CHECK-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]709; CHECK-NEXT: punpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload710; CHECK-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]711; CHECK-NEXT: punpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload712; CHECK-NEXT: # xmm0 = xmm0[0],mem[0]713; CHECK-NEXT: addq $72, %rsp714; CHECK-NEXT: popq %rbx715; CHECK-NEXT: popq %rbp716; CHECK-NEXT: retq717 %x = call <8 x i1> @llvm.fptosi.sat.v8i1.v8f16(<8 x half> %f)718 ret <8 x i1> %x719}720 721define <8 x i8> @test_signed_v8i8_v8f16(<8 x half> %f) nounwind {722; CHECK-LABEL: test_signed_v8i8_v8f16:723; CHECK: # %bb.0:724; CHECK-NEXT: pushq %rbp725; CHECK-NEXT: pushq %r15726; CHECK-NEXT: pushq %r14727; CHECK-NEXT: pushq %r13728; CHECK-NEXT: pushq %r12729; CHECK-NEXT: pushq %rbx730; CHECK-NEXT: subq $40, %rsp731; CHECK-NEXT: movdqa %xmm0, (%rsp) # 16-byte Spill732; CHECK-NEXT: psrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero733; CHECK-NEXT: callq __extendhfsf2@PLT734; CHECK-NEXT: cvttss2si %xmm0, %r12d735; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0736; CHECK-NEXT: movl $128, %ebx737; CHECK-NEXT: cmovbl %ebx, %r12d738; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0739; CHECK-NEXT: movl $127, %ebp740; CHECK-NEXT: cmoval %ebp, %r12d741; CHECK-NEXT: xorl %r14d, %r14d742; CHECK-NEXT: ucomiss %xmm0, %xmm0743; CHECK-NEXT: cmovpl %r14d, %r12d744; CHECK-NEXT: shll $8, %r12d745; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload746; CHECK-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1]747; CHECK-NEXT: callq __extendhfsf2@PLT748; CHECK-NEXT: cvttss2si %xmm0, %eax749; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0750; CHECK-NEXT: cmovbl %ebx, %eax751; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0752; CHECK-NEXT: cmoval %ebp, %eax753; CHECK-NEXT: ucomiss %xmm0, %xmm0754; CHECK-NEXT: cmovpl %r14d, %eax755; CHECK-NEXT: movzbl %al, %r15d756; CHECK-NEXT: orl %r12d, %r15d757; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload758; CHECK-NEXT: callq __extendhfsf2@PLT759; CHECK-NEXT: cvttss2si %xmm0, %eax760; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0761; CHECK-NEXT: cmovbl %ebx, %eax762; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0763; CHECK-NEXT: cmoval %ebp, %eax764; CHECK-NEXT: ucomiss %xmm0, %xmm0765; CHECK-NEXT: cmovpl %r14d, %eax766; CHECK-NEXT: movzbl %al, %r12d767; CHECK-NEXT: movdqa (%rsp), %xmm0 # 16-byte Reload768; CHECK-NEXT: psrld $16, %xmm0769; CHECK-NEXT: callq __extendhfsf2@PLT770; CHECK-NEXT: cvttss2si %xmm0, %eax771; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0772; CHECK-NEXT: cmovbl %ebx, %eax773; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0774; CHECK-NEXT: cmoval %ebp, %eax775; CHECK-NEXT: ucomiss %xmm0, %xmm0776; CHECK-NEXT: cmovpl %r14d, %eax777; CHECK-NEXT: movzbl %al, %r13d778; CHECK-NEXT: shll $8, %r13d779; CHECK-NEXT: orl %r12d, %r13d780; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload781; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1,1,1]782; CHECK-NEXT: callq __extendhfsf2@PLT783; CHECK-NEXT: cvttss2si %xmm0, %eax784; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0785; CHECK-NEXT: cmovbl %ebx, %eax786; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0787; CHECK-NEXT: cmoval %ebp, %eax788; CHECK-NEXT: ucomiss %xmm0, %xmm0789; CHECK-NEXT: cmovpl %r14d, %eax790; CHECK-NEXT: movzbl %al, %r12d791; CHECK-NEXT: shll $16, %r12d792; CHECK-NEXT: orl %r13d, %r12d793; CHECK-NEXT: movdqa (%rsp), %xmm0 # 16-byte Reload794; CHECK-NEXT: psrlq $48, %xmm0795; CHECK-NEXT: callq __extendhfsf2@PLT796; CHECK-NEXT: cvttss2si %xmm0, %eax797; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0798; CHECK-NEXT: cmovbl %ebx, %eax799; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0800; CHECK-NEXT: cmoval %ebp, %eax801; CHECK-NEXT: ucomiss %xmm0, %xmm0802; CHECK-NEXT: cmovpl %r14d, %eax803; CHECK-NEXT: shll $24, %eax804; CHECK-NEXT: orl %r12d, %eax805; CHECK-NEXT: movd %eax, %xmm0806; CHECK-NEXT: pinsrw $2, %r15d, %xmm0807; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill808; CHECK-NEXT: movdqa (%rsp), %xmm0 # 16-byte Reload809; CHECK-NEXT: psrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero810; CHECK-NEXT: callq __extendhfsf2@PLT811; CHECK-NEXT: cvttss2si %xmm0, %r15d812; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0813; CHECK-NEXT: cmovbl %ebx, %r15d814; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0815; CHECK-NEXT: cmoval %ebp, %r15d816; CHECK-NEXT: ucomiss %xmm0, %xmm0817; CHECK-NEXT: cmovpl %r14d, %r15d818; CHECK-NEXT: shll $8, %r15d819; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload820; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3,3,3]821; CHECK-NEXT: callq __extendhfsf2@PLT822; CHECK-NEXT: cvttss2si %xmm0, %eax823; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0824; CHECK-NEXT: cmovbl %ebx, %eax825; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0826; CHECK-NEXT: cmoval %ebp, %eax827; CHECK-NEXT: ucomiss %xmm0, %xmm0828; CHECK-NEXT: cmovpl %r14d, %eax829; CHECK-NEXT: movzbl %al, %eax830; CHECK-NEXT: orl %r15d, %eax831; CHECK-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload832; CHECK-NEXT: pinsrw $3, %eax, %xmm0833; CHECK-NEXT: addq $40, %rsp834; CHECK-NEXT: popq %rbx835; CHECK-NEXT: popq %r12836; CHECK-NEXT: popq %r13837; CHECK-NEXT: popq %r14838; CHECK-NEXT: popq %r15839; CHECK-NEXT: popq %rbp840; CHECK-NEXT: retq841 %x = call <8 x i8> @llvm.fptosi.sat.v8i8.v8f16(<8 x half> %f)842 ret <8 x i8> %x843}844 845define <8 x i16> @test_signed_v8i16_v8f16(<8 x half> %f) nounwind {846; CHECK-LABEL: test_signed_v8i16_v8f16:847; CHECK: # %bb.0:848; CHECK-NEXT: pushq %rbp849; CHECK-NEXT: pushq %r14850; CHECK-NEXT: pushq %rbx851; CHECK-NEXT: subq $64, %rsp852; CHECK-NEXT: movdqa %xmm0, (%rsp) # 16-byte Spill853; CHECK-NEXT: psrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero854; CHECK-NEXT: callq __extendhfsf2@PLT855; CHECK-NEXT: cvttss2si %xmm0, %eax856; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0857; CHECK-NEXT: movl $32768, %ebx # imm = 0x8000858; CHECK-NEXT: cmovbl %ebx, %eax859; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0860; CHECK-NEXT: movl $32767, %ebp # imm = 0x7FFF861; CHECK-NEXT: cmoval %ebp, %eax862; CHECK-NEXT: xorl %r14d, %r14d863; CHECK-NEXT: ucomiss %xmm0, %xmm0864; CHECK-NEXT: cmovpl %r14d, %eax865; CHECK-NEXT: movd %eax, %xmm0866; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill867; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload868; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3,3,3]869; CHECK-NEXT: callq __extendhfsf2@PLT870; CHECK-NEXT: cvttss2si %xmm0, %eax871; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0872; CHECK-NEXT: cmovbl %ebx, %eax873; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0874; CHECK-NEXT: cmoval %ebp, %eax875; CHECK-NEXT: ucomiss %xmm0, %xmm0876; CHECK-NEXT: cmovpl %r14d, %eax877; CHECK-NEXT: movd %eax, %xmm0878; CHECK-NEXT: punpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload879; CHECK-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]880; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill881; CHECK-NEXT: movdqa (%rsp), %xmm0 # 16-byte Reload882; CHECK-NEXT: psrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero883; CHECK-NEXT: callq __extendhfsf2@PLT884; CHECK-NEXT: cvttss2si %xmm0, %eax885; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0886; CHECK-NEXT: cmovbl %ebx, %eax887; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0888; CHECK-NEXT: cmoval %ebp, %eax889; CHECK-NEXT: ucomiss %xmm0, %xmm0890; CHECK-NEXT: cmovpl %r14d, %eax891; CHECK-NEXT: movd %eax, %xmm0892; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill893; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload894; CHECK-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1]895; CHECK-NEXT: callq __extendhfsf2@PLT896; CHECK-NEXT: cvttss2si %xmm0, %eax897; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0898; CHECK-NEXT: cmovbl %ebx, %eax899; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0900; CHECK-NEXT: cmoval %ebp, %eax901; CHECK-NEXT: ucomiss %xmm0, %xmm0902; CHECK-NEXT: cmovpl %r14d, %eax903; CHECK-NEXT: movd %eax, %xmm0904; CHECK-NEXT: punpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload905; CHECK-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]906; CHECK-NEXT: punpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload907; CHECK-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]908; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill909; CHECK-NEXT: movdqa (%rsp), %xmm0 # 16-byte Reload910; CHECK-NEXT: psrlq $48, %xmm0911; CHECK-NEXT: callq __extendhfsf2@PLT912; CHECK-NEXT: cvttss2si %xmm0, %eax913; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0914; CHECK-NEXT: cmovbl %ebx, %eax915; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0916; CHECK-NEXT: cmoval %ebp, %eax917; CHECK-NEXT: ucomiss %xmm0, %xmm0918; CHECK-NEXT: cmovpl %r14d, %eax919; CHECK-NEXT: movd %eax, %xmm0920; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill921; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload922; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1,1,1]923; CHECK-NEXT: callq __extendhfsf2@PLT924; CHECK-NEXT: cvttss2si %xmm0, %eax925; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0926; CHECK-NEXT: cmovbl %ebx, %eax927; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0928; CHECK-NEXT: cmoval %ebp, %eax929; CHECK-NEXT: ucomiss %xmm0, %xmm0930; CHECK-NEXT: cmovpl %r14d, %eax931; CHECK-NEXT: movd %eax, %xmm0932; CHECK-NEXT: punpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload933; CHECK-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]934; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill935; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload936; CHECK-NEXT: callq __extendhfsf2@PLT937; CHECK-NEXT: cvttss2si %xmm0, %eax938; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0939; CHECK-NEXT: cmovbl %ebx, %eax940; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0941; CHECK-NEXT: cmoval %ebp, %eax942; CHECK-NEXT: ucomiss %xmm0, %xmm0943; CHECK-NEXT: cmovpl %r14d, %eax944; CHECK-NEXT: movd %eax, %xmm0945; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill946; CHECK-NEXT: movdqa (%rsp), %xmm0 # 16-byte Reload947; CHECK-NEXT: psrld $16, %xmm0948; CHECK-NEXT: callq __extendhfsf2@PLT949; CHECK-NEXT: cvttss2si %xmm0, %eax950; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0951; CHECK-NEXT: cmovbl %ebx, %eax952; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0953; CHECK-NEXT: cmoval %ebp, %eax954; CHECK-NEXT: ucomiss %xmm0, %xmm0955; CHECK-NEXT: cmovpl %r14d, %eax956; CHECK-NEXT: movd %eax, %xmm1957; CHECK-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload958; CHECK-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]959; CHECK-NEXT: punpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload960; CHECK-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]961; CHECK-NEXT: punpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload962; CHECK-NEXT: # xmm0 = xmm0[0],mem[0]963; CHECK-NEXT: addq $64, %rsp964; CHECK-NEXT: popq %rbx965; CHECK-NEXT: popq %r14966; CHECK-NEXT: popq %rbp967; CHECK-NEXT: retq968 %x = call <8 x i16> @llvm.fptosi.sat.v8i16.v8f16(<8 x half> %f)969 ret <8 x i16> %x970}971 972define <8 x i32> @test_signed_v8i32_v8f16(<8 x half> %f) nounwind {973; CHECK-LABEL: test_signed_v8i32_v8f16:974; CHECK: # %bb.0:975; CHECK-NEXT: pushq %rbp976; CHECK-NEXT: pushq %r14977; CHECK-NEXT: pushq %rbx978; CHECK-NEXT: subq $64, %rsp979; CHECK-NEXT: movdqa %xmm0, (%rsp) # 16-byte Spill980; CHECK-NEXT: psrlq $48, %xmm0981; CHECK-NEXT: callq __extendhfsf2@PLT982; CHECK-NEXT: cvttss2si %xmm0, %eax983; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0984; CHECK-NEXT: movl $-2147483648, %ebx # imm = 0x80000000985; CHECK-NEXT: cmovbl %ebx, %eax986; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0987; CHECK-NEXT: movl $2147483647, %ebp # imm = 0x7FFFFFFF988; CHECK-NEXT: cmoval %ebp, %eax989; CHECK-NEXT: xorl %r14d, %r14d990; CHECK-NEXT: ucomiss %xmm0, %xmm0991; CHECK-NEXT: cmovpl %r14d, %eax992; CHECK-NEXT: movd %eax, %xmm0993; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill994; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload995; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1,1,1]996; CHECK-NEXT: callq __extendhfsf2@PLT997; CHECK-NEXT: cvttss2si %xmm0, %eax998; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0999; CHECK-NEXT: cmovbl %ebx, %eax1000; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01001; CHECK-NEXT: cmoval %ebp, %eax1002; CHECK-NEXT: ucomiss %xmm0, %xmm01003; CHECK-NEXT: cmovpl %r14d, %eax1004; CHECK-NEXT: movd %eax, %xmm01005; CHECK-NEXT: punpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload1006; CHECK-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]1007; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1008; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload1009; CHECK-NEXT: callq __extendhfsf2@PLT1010; CHECK-NEXT: cvttss2si %xmm0, %eax1011; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01012; CHECK-NEXT: cmovbl %ebx, %eax1013; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01014; CHECK-NEXT: cmoval %ebp, %eax1015; CHECK-NEXT: ucomiss %xmm0, %xmm01016; CHECK-NEXT: cmovpl %r14d, %eax1017; CHECK-NEXT: movd %eax, %xmm01018; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1019; CHECK-NEXT: movdqa (%rsp), %xmm0 # 16-byte Reload1020; CHECK-NEXT: psrld $16, %xmm01021; CHECK-NEXT: callq __extendhfsf2@PLT1022; CHECK-NEXT: cvttss2si %xmm0, %eax1023; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01024; CHECK-NEXT: cmovbl %ebx, %eax1025; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01026; CHECK-NEXT: cmoval %ebp, %eax1027; CHECK-NEXT: ucomiss %xmm0, %xmm01028; CHECK-NEXT: cmovpl %r14d, %eax1029; CHECK-NEXT: movd %eax, %xmm01030; CHECK-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload1031; CHECK-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]1032; CHECK-NEXT: punpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload1033; CHECK-NEXT: # xmm1 = xmm1[0],mem[0]1034; CHECK-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1035; CHECK-NEXT: movdqa (%rsp), %xmm0 # 16-byte Reload1036; CHECK-NEXT: psrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero1037; CHECK-NEXT: callq __extendhfsf2@PLT1038; CHECK-NEXT: cvttss2si %xmm0, %eax1039; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01040; CHECK-NEXT: cmovbl %ebx, %eax1041; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01042; CHECK-NEXT: cmoval %ebp, %eax1043; CHECK-NEXT: ucomiss %xmm0, %xmm01044; CHECK-NEXT: cmovpl %r14d, %eax1045; CHECK-NEXT: movd %eax, %xmm01046; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1047; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload1048; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3,3,3]1049; CHECK-NEXT: callq __extendhfsf2@PLT1050; CHECK-NEXT: cvttss2si %xmm0, %eax1051; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01052; CHECK-NEXT: cmovbl %ebx, %eax1053; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01054; CHECK-NEXT: cmoval %ebp, %eax1055; CHECK-NEXT: ucomiss %xmm0, %xmm01056; CHECK-NEXT: cmovpl %r14d, %eax1057; CHECK-NEXT: movd %eax, %xmm01058; CHECK-NEXT: punpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload1059; CHECK-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]1060; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1061; CHECK-NEXT: movdqa (%rsp), %xmm0 # 16-byte Reload1062; CHECK-NEXT: psrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero1063; CHECK-NEXT: callq __extendhfsf2@PLT1064; CHECK-NEXT: cvttss2si %xmm0, %eax1065; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01066; CHECK-NEXT: cmovbl %ebx, %eax1067; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01068; CHECK-NEXT: cmoval %ebp, %eax1069; CHECK-NEXT: ucomiss %xmm0, %xmm01070; CHECK-NEXT: cmovpl %r14d, %eax1071; CHECK-NEXT: movd %eax, %xmm01072; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1073; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload1074; CHECK-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1]1075; CHECK-NEXT: callq __extendhfsf2@PLT1076; CHECK-NEXT: cvttss2si %xmm0, %eax1077; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01078; CHECK-NEXT: cmovbl %ebx, %eax1079; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01080; CHECK-NEXT: cmoval %ebp, %eax1081; CHECK-NEXT: ucomiss %xmm0, %xmm01082; CHECK-NEXT: cmovpl %r14d, %eax1083; CHECK-NEXT: movd %eax, %xmm11084; CHECK-NEXT: punpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload1085; CHECK-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]1086; CHECK-NEXT: punpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload1087; CHECK-NEXT: # xmm1 = xmm1[0],mem[0]1088; CHECK-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1089; CHECK-NEXT: addq $64, %rsp1090; CHECK-NEXT: popq %rbx1091; CHECK-NEXT: popq %r141092; CHECK-NEXT: popq %rbp1093; CHECK-NEXT: retq1094 %x = call <8 x i32> @llvm.fptosi.sat.v8i32.v8f16(<8 x half> %f)1095 ret <8 x i32> %x1096}1097 1098define <8 x i64> @test_signed_v8i64_v8f16(<8 x half> %f) nounwind {1099; CHECK-LABEL: test_signed_v8i64_v8f16:1100; CHECK: # %bb.0:1101; CHECK-NEXT: pushq %r151102; CHECK-NEXT: pushq %r141103; CHECK-NEXT: pushq %rbx1104; CHECK-NEXT: subq $80, %rsp1105; CHECK-NEXT: movaps %xmm0, (%rsp) # 16-byte Spill1106; CHECK-NEXT: callq __extendhfsf2@PLT1107; CHECK-NEXT: cvttss2si %xmm0, %rax1108; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01109; CHECK-NEXT: movabsq $-9223372036854775808, %rbx # imm = 0x80000000000000001110; CHECK-NEXT: cmovbq %rbx, %rax1111; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01112; CHECK-NEXT: movabsq $9223372036854775807, %r14 # imm = 0x7FFFFFFFFFFFFFFF1113; CHECK-NEXT: cmovaq %r14, %rax1114; CHECK-NEXT: xorl %r15d, %r15d1115; CHECK-NEXT: ucomiss %xmm0, %xmm01116; CHECK-NEXT: cmovpq %r15, %rax1117; CHECK-NEXT: movq %rax, %xmm01118; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1119; CHECK-NEXT: movdqa (%rsp), %xmm0 # 16-byte Reload1120; CHECK-NEXT: psrld $16, %xmm01121; CHECK-NEXT: callq __extendhfsf2@PLT1122; CHECK-NEXT: cvttss2si %xmm0, %rax1123; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01124; CHECK-NEXT: cmovbq %rbx, %rax1125; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01126; CHECK-NEXT: cmovaq %r14, %rax1127; CHECK-NEXT: ucomiss %xmm0, %xmm01128; CHECK-NEXT: cmovpq %r15, %rax1129; CHECK-NEXT: movq %rax, %xmm01130; CHECK-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload1131; CHECK-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0]1132; CHECK-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1133; CHECK-NEXT: movdqa (%rsp), %xmm0 # 16-byte Reload1134; CHECK-NEXT: psrlq $48, %xmm01135; CHECK-NEXT: callq __extendhfsf2@PLT1136; CHECK-NEXT: cvttss2si %xmm0, %rax1137; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01138; CHECK-NEXT: cmovbq %rbx, %rax1139; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01140; CHECK-NEXT: cmovaq %r14, %rax1141; CHECK-NEXT: ucomiss %xmm0, %xmm01142; CHECK-NEXT: cmovpq %r15, %rax1143; CHECK-NEXT: movq %rax, %xmm01144; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1145; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload1146; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1,1,1]1147; CHECK-NEXT: callq __extendhfsf2@PLT1148; CHECK-NEXT: cvttss2si %xmm0, %rax1149; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01150; CHECK-NEXT: cmovbq %rbx, %rax1151; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01152; CHECK-NEXT: cmovaq %r14, %rax1153; CHECK-NEXT: ucomiss %xmm0, %xmm01154; CHECK-NEXT: cmovpq %r15, %rax1155; CHECK-NEXT: movq %rax, %xmm01156; CHECK-NEXT: punpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload1157; CHECK-NEXT: # xmm0 = xmm0[0],mem[0]1158; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1159; CHECK-NEXT: movdqa (%rsp), %xmm0 # 16-byte Reload1160; CHECK-NEXT: psrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero1161; CHECK-NEXT: callq __extendhfsf2@PLT1162; CHECK-NEXT: cvttss2si %xmm0, %rax1163; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01164; CHECK-NEXT: cmovbq %rbx, %rax1165; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01166; CHECK-NEXT: cmovaq %r14, %rax1167; CHECK-NEXT: ucomiss %xmm0, %xmm01168; CHECK-NEXT: cmovpq %r15, %rax1169; CHECK-NEXT: movq %rax, %xmm01170; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1171; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload1172; CHECK-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1]1173; CHECK-NEXT: callq __extendhfsf2@PLT1174; CHECK-NEXT: cvttss2si %xmm0, %rax1175; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01176; CHECK-NEXT: cmovbq %rbx, %rax1177; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01178; CHECK-NEXT: cmovaq %r14, %rax1179; CHECK-NEXT: ucomiss %xmm0, %xmm01180; CHECK-NEXT: cmovpq %r15, %rax1181; CHECK-NEXT: movq %rax, %xmm01182; CHECK-NEXT: punpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload1183; CHECK-NEXT: # xmm0 = xmm0[0],mem[0]1184; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1185; CHECK-NEXT: movdqa (%rsp), %xmm0 # 16-byte Reload1186; CHECK-NEXT: psrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero1187; CHECK-NEXT: callq __extendhfsf2@PLT1188; CHECK-NEXT: cvttss2si %xmm0, %rax1189; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01190; CHECK-NEXT: cmovbq %rbx, %rax1191; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01192; CHECK-NEXT: cmovaq %r14, %rax1193; CHECK-NEXT: ucomiss %xmm0, %xmm01194; CHECK-NEXT: cmovpq %r15, %rax1195; CHECK-NEXT: movq %rax, %xmm01196; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1197; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload1198; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3,3,3]1199; CHECK-NEXT: callq __extendhfsf2@PLT1200; CHECK-NEXT: cvttss2si %xmm0, %rax1201; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01202; CHECK-NEXT: cmovbq %rbx, %rax1203; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01204; CHECK-NEXT: cmovaq %r14, %rax1205; CHECK-NEXT: ucomiss %xmm0, %xmm01206; CHECK-NEXT: cmovpq %r15, %rax1207; CHECK-NEXT: movq %rax, %xmm31208; CHECK-NEXT: punpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload1209; CHECK-NEXT: # xmm3 = xmm3[0],mem[0]1210; CHECK-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1211; CHECK-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload1212; CHECK-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload1213; CHECK-NEXT: addq $80, %rsp1214; CHECK-NEXT: popq %rbx1215; CHECK-NEXT: popq %r141216; CHECK-NEXT: popq %r151217; CHECK-NEXT: retq1218 %x = call <8 x i64> @llvm.fptosi.sat.v8i64.v8f16(<8 x half> %f)1219 ret <8 x i64> %x1220}1221 1222define <8 x i128> @test_signed_v8i128_v8f16(<8 x half> %f) nounwind {1223; CHECK-LABEL: test_signed_v8i128_v8f16:1224; CHECK: # %bb.0:1225; CHECK-NEXT: pushq %rbp1226; CHECK-NEXT: pushq %r151227; CHECK-NEXT: pushq %r141228; CHECK-NEXT: pushq %r131229; CHECK-NEXT: pushq %r121230; CHECK-NEXT: pushq %rbx1231; CHECK-NEXT: subq $104, %rsp1232; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1233; CHECK-NEXT: movq %rdi, %rbx1234; CHECK-NEXT: psrld $16, %xmm01235; CHECK-NEXT: callq __extendhfsf2@PLT1236; CHECK-NEXT: movd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Folded Spill1237; CHECK-NEXT: callq __fixsfti@PLT1238; CHECK-NEXT: xorl %r12d, %r12d1239; CHECK-NEXT: movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 4-byte Reload1240; CHECK-NEXT: # xmm0 = mem[0],zero,zero,zero1241; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01242; CHECK-NEXT: cmovbq %r12, %rax1243; CHECK-NEXT: movabsq $-9223372036854775808, %rbp # imm = 0x80000000000000001244; CHECK-NEXT: cmovbq %rbp, %rdx1245; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01246; CHECK-NEXT: movabsq $9223372036854775807, %r15 # imm = 0x7FFFFFFFFFFFFFFF1247; CHECK-NEXT: cmovaq %r15, %rdx1248; CHECK-NEXT: movq $-1, %rcx1249; CHECK-NEXT: cmovaq %rcx, %rax1250; CHECK-NEXT: movq $-1, %r131251; CHECK-NEXT: ucomiss %xmm0, %xmm01252; CHECK-NEXT: cmovpq %r12, %rax1253; CHECK-NEXT: movq %rax, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill1254; CHECK-NEXT: cmovpq %r12, %rdx1255; CHECK-NEXT: movq %rdx, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill1256; CHECK-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1257; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1,1,1]1258; CHECK-NEXT: callq __extendhfsf2@PLT1259; CHECK-NEXT: movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill1260; CHECK-NEXT: callq __fixsfti@PLT1261; CHECK-NEXT: movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 4-byte Reload1262; CHECK-NEXT: # xmm0 = mem[0],zero,zero,zero1263; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01264; CHECK-NEXT: cmovbq %r12, %rax1265; CHECK-NEXT: cmovbq %rbp, %rdx1266; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01267; CHECK-NEXT: cmovaq %r15, %rdx1268; CHECK-NEXT: cmovaq %r13, %rax1269; CHECK-NEXT: movq $-1, %r141270; CHECK-NEXT: ucomiss %xmm0, %xmm01271; CHECK-NEXT: cmovpq %r12, %rax1272; CHECK-NEXT: movq %rax, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill1273; CHECK-NEXT: cmovpq %r12, %rdx1274; CHECK-NEXT: movq %rdx, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill1275; CHECK-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1276; CHECK-NEXT: psrlq $48, %xmm01277; CHECK-NEXT: callq __extendhfsf2@PLT1278; CHECK-NEXT: movd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Folded Spill1279; CHECK-NEXT: callq __fixsfti@PLT1280; CHECK-NEXT: movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 4-byte Reload1281; CHECK-NEXT: # xmm0 = mem[0],zero,zero,zero1282; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01283; CHECK-NEXT: cmovbq %r12, %rax1284; CHECK-NEXT: cmovbq %rbp, %rdx1285; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01286; CHECK-NEXT: cmovaq %r15, %rdx1287; CHECK-NEXT: cmovaq %r14, %rax1288; CHECK-NEXT: movq $-1, %r141289; CHECK-NEXT: ucomiss %xmm0, %xmm01290; CHECK-NEXT: cmovpq %r12, %rax1291; CHECK-NEXT: movq %rax, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill1292; CHECK-NEXT: cmovpq %r12, %rdx1293; CHECK-NEXT: movq %rdx, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill1294; CHECK-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1295; CHECK-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1]1296; CHECK-NEXT: callq __extendhfsf2@PLT1297; CHECK-NEXT: movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill1298; CHECK-NEXT: callq __fixsfti@PLT1299; CHECK-NEXT: movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 4-byte Reload1300; CHECK-NEXT: # xmm0 = mem[0],zero,zero,zero1301; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01302; CHECK-NEXT: cmovbq %r12, %rax1303; CHECK-NEXT: cmovbq %rbp, %rdx1304; CHECK-NEXT: movq %rbp, %r131305; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01306; CHECK-NEXT: cmovaq %r15, %rdx1307; CHECK-NEXT: cmovaq %r14, %rax1308; CHECK-NEXT: movq $-1, %r141309; CHECK-NEXT: ucomiss %xmm0, %xmm01310; CHECK-NEXT: cmovpq %r12, %rax1311; CHECK-NEXT: movq %rax, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill1312; CHECK-NEXT: cmovpq %r12, %rdx1313; CHECK-NEXT: movq %rdx, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill1314; CHECK-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1315; CHECK-NEXT: psrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero1316; CHECK-NEXT: callq __extendhfsf2@PLT1317; CHECK-NEXT: movd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Folded Spill1318; CHECK-NEXT: callq __fixsfti@PLT1319; CHECK-NEXT: movq %rdx, %rbp1320; CHECK-NEXT: movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 4-byte Reload1321; CHECK-NEXT: # xmm0 = mem[0],zero,zero,zero1322; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01323; CHECK-NEXT: cmovbq %r12, %rax1324; CHECK-NEXT: cmovbq %r13, %rbp1325; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01326; CHECK-NEXT: cmovaq %r15, %rbp1327; CHECK-NEXT: movq %r15, %r131328; CHECK-NEXT: cmovaq %r14, %rax1329; CHECK-NEXT: ucomiss %xmm0, %xmm01330; CHECK-NEXT: cmovpq %r12, %rax1331; CHECK-NEXT: movq %rax, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill1332; CHECK-NEXT: cmovpq %r12, %rbp1333; CHECK-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1334; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3,3,3]1335; CHECK-NEXT: callq __extendhfsf2@PLT1336; CHECK-NEXT: movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill1337; CHECK-NEXT: callq __fixsfti@PLT1338; CHECK-NEXT: movq %rax, %r141339; CHECK-NEXT: movq %rdx, %r151340; CHECK-NEXT: movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 4-byte Reload1341; CHECK-NEXT: # xmm0 = mem[0],zero,zero,zero1342; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01343; CHECK-NEXT: cmovbq %r12, %r141344; CHECK-NEXT: movabsq $-9223372036854775808, %rax # imm = 0x80000000000000001345; CHECK-NEXT: cmovbq %rax, %r151346; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01347; CHECK-NEXT: cmovaq %r13, %r151348; CHECK-NEXT: movq $-1, %rax1349; CHECK-NEXT: cmovaq %rax, %r141350; CHECK-NEXT: ucomiss %xmm0, %xmm01351; CHECK-NEXT: cmovpq %r12, %r141352; CHECK-NEXT: cmovpq %r12, %r151353; CHECK-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1354; CHECK-NEXT: psrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero1355; CHECK-NEXT: callq __extendhfsf2@PLT1356; CHECK-NEXT: movd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Folded Spill1357; CHECK-NEXT: callq __fixsfti@PLT1358; CHECK-NEXT: movq %rax, %r121359; CHECK-NEXT: movq %rdx, %r131360; CHECK-NEXT: movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 4-byte Reload1361; CHECK-NEXT: # xmm0 = mem[0],zero,zero,zero1362; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01363; CHECK-NEXT: movl $0, %eax1364; CHECK-NEXT: cmovbq %rax, %r121365; CHECK-NEXT: movabsq $-9223372036854775808, %rcx # imm = 0x80000000000000001366; CHECK-NEXT: cmovbq %rcx, %r131367; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01368; CHECK-NEXT: movabsq $9223372036854775807, %rcx # imm = 0x7FFFFFFFFFFFFFFF1369; CHECK-NEXT: cmovaq %rcx, %r131370; CHECK-NEXT: movq $-1, %rcx1371; CHECK-NEXT: cmovaq %rcx, %r121372; CHECK-NEXT: ucomiss %xmm0, %xmm01373; CHECK-NEXT: cmovpq %rax, %r121374; CHECK-NEXT: cmovpq %rax, %r131375; CHECK-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1376; CHECK-NEXT: callq __extendhfsf2@PLT1377; CHECK-NEXT: movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill1378; CHECK-NEXT: callq __fixsfti@PLT1379; CHECK-NEXT: movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 4-byte Reload1380; CHECK-NEXT: # xmm0 = mem[0],zero,zero,zero1381; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01382; CHECK-NEXT: movabsq $-9223372036854775808, %rcx # imm = 0x80000000000000001383; CHECK-NEXT: cmovbq %rcx, %rdx1384; CHECK-NEXT: movl $0, %esi1385; CHECK-NEXT: cmovbq %rsi, %rax1386; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm01387; CHECK-NEXT: movabsq $9223372036854775807, %rcx # imm = 0x7FFFFFFFFFFFFFFF1388; CHECK-NEXT: cmovaq %rcx, %rdx1389; CHECK-NEXT: movq $-1, %rcx1390; CHECK-NEXT: cmovaq %rcx, %rax1391; CHECK-NEXT: ucomiss %xmm0, %xmm01392; CHECK-NEXT: cmovpq %rsi, %rax1393; CHECK-NEXT: movl $0, %ecx1394; CHECK-NEXT: cmovpq %rcx, %rdx1395; CHECK-NEXT: movq %rdx, 8(%rbx)1396; CHECK-NEXT: movq %rax, (%rbx)1397; CHECK-NEXT: movq %r13, 120(%rbx)1398; CHECK-NEXT: movq %r12, 112(%rbx)1399; CHECK-NEXT: movq %r15, 104(%rbx)1400; CHECK-NEXT: movq %r14, 96(%rbx)1401; CHECK-NEXT: movq %rbp, 88(%rbx)1402; CHECK-NEXT: movq {{[-0-9]+}}(%r{{[sb]}}p), %rax # 8-byte Reload1403; CHECK-NEXT: movq %rax, 80(%rbx)1404; CHECK-NEXT: movq {{[-0-9]+}}(%r{{[sb]}}p), %rax # 8-byte Reload1405; CHECK-NEXT: movq %rax, 72(%rbx)1406; CHECK-NEXT: movq {{[-0-9]+}}(%r{{[sb]}}p), %rax # 8-byte Reload1407; CHECK-NEXT: movq %rax, 64(%rbx)1408; CHECK-NEXT: movq {{[-0-9]+}}(%r{{[sb]}}p), %rax # 8-byte Reload1409; CHECK-NEXT: movq %rax, 56(%rbx)1410; CHECK-NEXT: movq {{[-0-9]+}}(%r{{[sb]}}p), %rax # 8-byte Reload1411; CHECK-NEXT: movq %rax, 48(%rbx)1412; CHECK-NEXT: movq {{[-0-9]+}}(%r{{[sb]}}p), %rax # 8-byte Reload1413; CHECK-NEXT: movq %rax, 40(%rbx)1414; CHECK-NEXT: movq {{[-0-9]+}}(%r{{[sb]}}p), %rax # 8-byte Reload1415; CHECK-NEXT: movq %rax, 32(%rbx)1416; CHECK-NEXT: movq {{[-0-9]+}}(%r{{[sb]}}p), %rax # 8-byte Reload1417; CHECK-NEXT: movq %rax, 24(%rbx)1418; CHECK-NEXT: movq {{[-0-9]+}}(%r{{[sb]}}p), %rax # 8-byte Reload1419; CHECK-NEXT: movq %rax, 16(%rbx)1420; CHECK-NEXT: movq %rbx, %rax1421; CHECK-NEXT: addq $104, %rsp1422; CHECK-NEXT: popq %rbx1423; CHECK-NEXT: popq %r121424; CHECK-NEXT: popq %r131425; CHECK-NEXT: popq %r141426; CHECK-NEXT: popq %r151427; CHECK-NEXT: popq %rbp1428; CHECK-NEXT: retq1429 %x = call <8 x i128> @llvm.fptosi.sat.v8i128.v8f16(<8 x half> %f)1430 ret <8 x i128> %x1431}1432