414 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse2 -O3 | FileCheck %s --check-prefixes=SSE23; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+f16c -O3 | FileCheck %s --check-prefixes=AVX,F16C4; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f -O3 | FileCheck %s --check-prefixes=AVX,AVX5125; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+avx512fp16 -mattr=+avx512vl -O3 | FileCheck %s --check-prefixes=X866; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512fp16 -mattr=+avx512vl -O3 | FileCheck %s --check-prefixes=X647 8declare i1 @llvm.experimental.constrained.fptosi.i1.f16(half, metadata)9declare i8 @llvm.experimental.constrained.fptosi.i8.f16(half, metadata)10declare i16 @llvm.experimental.constrained.fptosi.i16.f16(half, metadata)11declare i32 @llvm.experimental.constrained.fptosi.i32.f16(half, metadata)12declare i64 @llvm.experimental.constrained.fptosi.i64.f16(half, metadata)13declare i1 @llvm.experimental.constrained.fptoui.i1.f16(half, metadata)14declare i8 @llvm.experimental.constrained.fptoui.i8.f16(half, metadata)15declare i16 @llvm.experimental.constrained.fptoui.i16.f16(half, metadata)16declare i32 @llvm.experimental.constrained.fptoui.i32.f16(half, metadata)17declare i64 @llvm.experimental.constrained.fptoui.i64.f16(half, metadata)18 19define i1 @fptosi_f16toi1(half %x) #0 {20; SSE2-LABEL: fptosi_f16toi1:21; SSE2: # %bb.0:22; SSE2-NEXT: pushq %rax23; SSE2-NEXT: callq __extendhfsf2@PLT24; SSE2-NEXT: cvttss2si %xmm0, %eax25; SSE2-NEXT: # kill: def $al killed $al killed $eax26; SSE2-NEXT: popq %rcx27; SSE2-NEXT: retq28;29; AVX-LABEL: fptosi_f16toi1:30; AVX: # %bb.0:31; AVX-NEXT: vpextrw $0, %xmm0, %eax32; AVX-NEXT: movzwl %ax, %eax33; AVX-NEXT: vmovd %eax, %xmm034; AVX-NEXT: vcvtph2ps %xmm0, %xmm035; AVX-NEXT: vcvttss2si %xmm0, %eax36; AVX-NEXT: # kill: def $al killed $al killed $eax37; AVX-NEXT: retq38;39; X86-LABEL: fptosi_f16toi1:40; X86: # %bb.0:41; X86-NEXT: vcvttsh2si {{[0-9]+}}(%esp), %eax42; X86-NEXT: # kill: def $al killed $al killed $eax43; X86-NEXT: retl44;45; X64-LABEL: fptosi_f16toi1:46; X64: # %bb.0:47; X64-NEXT: vcvttsh2si %xmm0, %eax48; X64-NEXT: # kill: def $al killed $al killed $eax49; X64-NEXT: retq50 %result = call i1 @llvm.experimental.constrained.fptosi.i1.f16(half %x,51 metadata !"fpexcept.strict") #052 ret i1 %result53}54 55define i8 @fptosi_f16toi8(half %x) #0 {56; SSE2-LABEL: fptosi_f16toi8:57; SSE2: # %bb.0:58; SSE2-NEXT: pushq %rax59; SSE2-NEXT: callq __extendhfsf2@PLT60; SSE2-NEXT: cvttss2si %xmm0, %eax61; SSE2-NEXT: # kill: def $al killed $al killed $eax62; SSE2-NEXT: popq %rcx63; SSE2-NEXT: retq64;65; AVX-LABEL: fptosi_f16toi8:66; AVX: # %bb.0:67; AVX-NEXT: vpextrw $0, %xmm0, %eax68; AVX-NEXT: movzwl %ax, %eax69; AVX-NEXT: vmovd %eax, %xmm070; AVX-NEXT: vcvtph2ps %xmm0, %xmm071; AVX-NEXT: vcvttss2si %xmm0, %eax72; AVX-NEXT: # kill: def $al killed $al killed $eax73; AVX-NEXT: retq74;75; X86-LABEL: fptosi_f16toi8:76; X86: # %bb.0:77; X86-NEXT: vcvttsh2si {{[0-9]+}}(%esp), %eax78; X86-NEXT: # kill: def $al killed $al killed $eax79; X86-NEXT: retl80;81; X64-LABEL: fptosi_f16toi8:82; X64: # %bb.0:83; X64-NEXT: vcvttsh2si %xmm0, %eax84; X64-NEXT: # kill: def $al killed $al killed $eax85; X64-NEXT: retq86 %result = call i8 @llvm.experimental.constrained.fptosi.i8.f16(half %x,87 metadata !"fpexcept.strict") #088 ret i8 %result89}90 91define i16 @fptosi_f16toi16(half %x) #0 {92; SSE2-LABEL: fptosi_f16toi16:93; SSE2: # %bb.0:94; SSE2-NEXT: pushq %rax95; SSE2-NEXT: callq __extendhfsf2@PLT96; SSE2-NEXT: cvttss2si %xmm0, %eax97; SSE2-NEXT: # kill: def $ax killed $ax killed $eax98; SSE2-NEXT: popq %rcx99; SSE2-NEXT: retq100;101; AVX-LABEL: fptosi_f16toi16:102; AVX: # %bb.0:103; AVX-NEXT: vpextrw $0, %xmm0, %eax104; AVX-NEXT: movzwl %ax, %eax105; AVX-NEXT: vmovd %eax, %xmm0106; AVX-NEXT: vcvtph2ps %xmm0, %xmm0107; AVX-NEXT: vcvttss2si %xmm0, %eax108; AVX-NEXT: # kill: def $ax killed $ax killed $eax109; AVX-NEXT: retq110;111; X86-LABEL: fptosi_f16toi16:112; X86: # %bb.0:113; X86-NEXT: vcvttsh2si {{[0-9]+}}(%esp), %eax114; X86-NEXT: # kill: def $ax killed $ax killed $eax115; X86-NEXT: retl116;117; X64-LABEL: fptosi_f16toi16:118; X64: # %bb.0:119; X64-NEXT: vcvttsh2si %xmm0, %eax120; X64-NEXT: # kill: def $ax killed $ax killed $eax121; X64-NEXT: retq122 %result = call i16 @llvm.experimental.constrained.fptosi.i16.f16(half %x,123 metadata !"fpexcept.strict") #0124 ret i16 %result125}126 127define i32 @fptosi_f16toi32(half %x) #0 {128; SSE2-LABEL: fptosi_f16toi32:129; SSE2: # %bb.0:130; SSE2-NEXT: pushq %rax131; SSE2-NEXT: callq __extendhfsf2@PLT132; SSE2-NEXT: cvttss2si %xmm0, %eax133; SSE2-NEXT: popq %rcx134; SSE2-NEXT: retq135;136; AVX-LABEL: fptosi_f16toi32:137; AVX: # %bb.0:138; AVX-NEXT: vpextrw $0, %xmm0, %eax139; AVX-NEXT: movzwl %ax, %eax140; AVX-NEXT: vmovd %eax, %xmm0141; AVX-NEXT: vcvtph2ps %xmm0, %xmm0142; AVX-NEXT: vcvttss2si %xmm0, %eax143; AVX-NEXT: retq144;145; X86-LABEL: fptosi_f16toi32:146; X86: # %bb.0:147; X86-NEXT: vcvttsh2si {{[0-9]+}}(%esp), %eax148; X86-NEXT: retl149;150; X64-LABEL: fptosi_f16toi32:151; X64: # %bb.0:152; X64-NEXT: vcvttsh2si %xmm0, %eax153; X64-NEXT: retq154 %result = call i32 @llvm.experimental.constrained.fptosi.i32.f16(half %x,155 metadata !"fpexcept.strict") #0156 ret i32 %result157}158 159define i64 @fptosi_f16toi64(half %x) #0 {160; SSE2-LABEL: fptosi_f16toi64:161; SSE2: # %bb.0:162; SSE2-NEXT: pushq %rax163; SSE2-NEXT: callq __extendhfsf2@PLT164; SSE2-NEXT: cvttss2si %xmm0, %rax165; SSE2-NEXT: popq %rcx166; SSE2-NEXT: retq167;168; AVX-LABEL: fptosi_f16toi64:169; AVX: # %bb.0:170; AVX-NEXT: vpextrw $0, %xmm0, %eax171; AVX-NEXT: movzwl %ax, %eax172; AVX-NEXT: vmovd %eax, %xmm0173; AVX-NEXT: vcvtph2ps %xmm0, %xmm0174; AVX-NEXT: vcvttss2si %xmm0, %rax175; AVX-NEXT: retq176;177; X86-LABEL: fptosi_f16toi64:178; X86: # %bb.0:179; X86-NEXT: vmovsh {{.*#+}} xmm0 = mem[0],zero,zero,zero,zero,zero,zero,zero180; X86-NEXT: vcvttph2qq %xmm0, %xmm0181; X86-NEXT: vmovd %xmm0, %eax182; X86-NEXT: vpextrd $1, %xmm0, %edx183; X86-NEXT: retl184;185; X64-LABEL: fptosi_f16toi64:186; X64: # %bb.0:187; X64-NEXT: vcvttsh2si %xmm0, %rax188; X64-NEXT: retq189 %result = call i64 @llvm.experimental.constrained.fptosi.i64.f16(half %x,190 metadata !"fpexcept.strict") #0191 ret i64 %result192}193 194define i1 @fptoui_f16toi1(half %x) #0 {195; SSE2-LABEL: fptoui_f16toi1:196; SSE2: # %bb.0:197; SSE2-NEXT: pushq %rax198; SSE2-NEXT: callq __extendhfsf2@PLT199; SSE2-NEXT: cvttss2si %xmm0, %eax200; SSE2-NEXT: # kill: def $al killed $al killed $eax201; SSE2-NEXT: popq %rcx202; SSE2-NEXT: retq203;204; AVX-LABEL: fptoui_f16toi1:205; AVX: # %bb.0:206; AVX-NEXT: vpextrw $0, %xmm0, %eax207; AVX-NEXT: movzwl %ax, %eax208; AVX-NEXT: vmovd %eax, %xmm0209; AVX-NEXT: vcvtph2ps %xmm0, %xmm0210; AVX-NEXT: vcvttss2si %xmm0, %eax211; AVX-NEXT: # kill: def $al killed $al killed $eax212; AVX-NEXT: retq213;214; X86-LABEL: fptoui_f16toi1:215; X86: # %bb.0:216; X86-NEXT: vcvttsh2si {{[0-9]+}}(%esp), %eax217; X86-NEXT: # kill: def $al killed $al killed $eax218; X86-NEXT: retl219;220; X64-LABEL: fptoui_f16toi1:221; X64: # %bb.0:222; X64-NEXT: vcvttsh2si %xmm0, %eax223; X64-NEXT: # kill: def $al killed $al killed $eax224; X64-NEXT: retq225 %result = call i1 @llvm.experimental.constrained.fptoui.i1.f16(half %x,226 metadata !"fpexcept.strict") #0227 ret i1 %result228}229 230define i8 @fptoui_f16toi8(half %x) #0 {231; SSE2-LABEL: fptoui_f16toi8:232; SSE2: # %bb.0:233; SSE2-NEXT: pushq %rax234; SSE2-NEXT: callq __extendhfsf2@PLT235; SSE2-NEXT: cvttss2si %xmm0, %eax236; SSE2-NEXT: # kill: def $al killed $al killed $eax237; SSE2-NEXT: popq %rcx238; SSE2-NEXT: retq239;240; AVX-LABEL: fptoui_f16toi8:241; AVX: # %bb.0:242; AVX-NEXT: vpextrw $0, %xmm0, %eax243; AVX-NEXT: movzwl %ax, %eax244; AVX-NEXT: vmovd %eax, %xmm0245; AVX-NEXT: vcvtph2ps %xmm0, %xmm0246; AVX-NEXT: vcvttss2si %xmm0, %eax247; AVX-NEXT: # kill: def $al killed $al killed $eax248; AVX-NEXT: retq249;250; X86-LABEL: fptoui_f16toi8:251; X86: # %bb.0:252; X86-NEXT: vcvttsh2si {{[0-9]+}}(%esp), %eax253; X86-NEXT: # kill: def $al killed $al killed $eax254; X86-NEXT: retl255;256; X64-LABEL: fptoui_f16toi8:257; X64: # %bb.0:258; X64-NEXT: vcvttsh2si %xmm0, %eax259; X64-NEXT: # kill: def $al killed $al killed $eax260; X64-NEXT: retq261 %result = call i8 @llvm.experimental.constrained.fptoui.i8.f16(half %x,262 metadata !"fpexcept.strict") #0263 ret i8 %result264}265 266define i16 @fptoui_f16toi16(half %x) #0 {267; SSE2-LABEL: fptoui_f16toi16:268; SSE2: # %bb.0:269; SSE2-NEXT: pushq %rax270; SSE2-NEXT: callq __extendhfsf2@PLT271; SSE2-NEXT: cvttss2si %xmm0, %eax272; SSE2-NEXT: # kill: def $ax killed $ax killed $eax273; SSE2-NEXT: popq %rcx274; SSE2-NEXT: retq275;276; AVX-LABEL: fptoui_f16toi16:277; AVX: # %bb.0:278; AVX-NEXT: vpextrw $0, %xmm0, %eax279; AVX-NEXT: movzwl %ax, %eax280; AVX-NEXT: vmovd %eax, %xmm0281; AVX-NEXT: vcvtph2ps %xmm0, %xmm0282; AVX-NEXT: vcvttss2si %xmm0, %eax283; AVX-NEXT: # kill: def $ax killed $ax killed $eax284; AVX-NEXT: retq285;286; X86-LABEL: fptoui_f16toi16:287; X86: # %bb.0:288; X86-NEXT: vcvttsh2si {{[0-9]+}}(%esp), %eax289; X86-NEXT: # kill: def $ax killed $ax killed $eax290; X86-NEXT: retl291;292; X64-LABEL: fptoui_f16toi16:293; X64: # %bb.0:294; X64-NEXT: vcvttsh2si %xmm0, %eax295; X64-NEXT: # kill: def $ax killed $ax killed $eax296; X64-NEXT: retq297 %result = call i16 @llvm.experimental.constrained.fptoui.i16.f16(half %x,298 metadata !"fpexcept.strict") #0299 ret i16 %result300}301 302define i32 @fptoui_f16toi32(half %x) #0 {303; SSE2-LABEL: fptoui_f16toi32:304; SSE2: # %bb.0:305; SSE2-NEXT: pushq %rax306; SSE2-NEXT: callq __extendhfsf2@PLT307; SSE2-NEXT: cvttss2si %xmm0, %rax308; SSE2-NEXT: # kill: def $eax killed $eax killed $rax309; SSE2-NEXT: popq %rcx310; SSE2-NEXT: retq311;312; F16C-LABEL: fptoui_f16toi32:313; F16C: # %bb.0:314; F16C-NEXT: vpextrw $0, %xmm0, %eax315; F16C-NEXT: movzwl %ax, %eax316; F16C-NEXT: vmovd %eax, %xmm0317; F16C-NEXT: vcvtph2ps %xmm0, %xmm0318; F16C-NEXT: vcvttss2si %xmm0, %rax319; F16C-NEXT: # kill: def $eax killed $eax killed $rax320; F16C-NEXT: retq321;322; AVX512-LABEL: fptoui_f16toi32:323; AVX512: # %bb.0:324; AVX512-NEXT: vpextrw $0, %xmm0, %eax325; AVX512-NEXT: movzwl %ax, %eax326; AVX512-NEXT: vmovd %eax, %xmm0327; AVX512-NEXT: vcvtph2ps %xmm0, %xmm0328; AVX512-NEXT: vcvttss2usi %xmm0, %eax329; AVX512-NEXT: retq330;331; X86-LABEL: fptoui_f16toi32:332; X86: # %bb.0:333; X86-NEXT: vcvttsh2usi {{[0-9]+}}(%esp), %eax334; X86-NEXT: retl335;336; X64-LABEL: fptoui_f16toi32:337; X64: # %bb.0:338; X64-NEXT: vcvttsh2usi %xmm0, %eax339; X64-NEXT: retq340 %result = call i32 @llvm.experimental.constrained.fptoui.i32.f16(half %x,341 metadata !"fpexcept.strict") #0342 ret i32 %result343}344 345define i64 @fptoui_f16toi64(half %x) #0 {346; SSE2-LABEL: fptoui_f16toi64:347; SSE2: # %bb.0:348; SSE2-NEXT: pushq %rax349; SSE2-NEXT: callq __extendhfsf2@PLT350; SSE2-NEXT: movss {{.*#+}} xmm2 = [9.22337203E+18,0.0E+0,0.0E+0,0.0E+0]351; SSE2-NEXT: comiss %xmm2, %xmm0352; SSE2-NEXT: xorps %xmm1, %xmm1353; SSE2-NEXT: jb .LBB9_2354; SSE2-NEXT: # %bb.1:355; SSE2-NEXT: movaps %xmm2, %xmm1356; SSE2-NEXT: .LBB9_2:357; SSE2-NEXT: subss %xmm1, %xmm0358; SSE2-NEXT: cvttss2si %xmm0, %rcx359; SSE2-NEXT: setae %al360; SSE2-NEXT: movzbl %al, %eax361; SSE2-NEXT: shlq $63, %rax362; SSE2-NEXT: xorq %rcx, %rax363; SSE2-NEXT: popq %rcx364; SSE2-NEXT: retq365;366; F16C-LABEL: fptoui_f16toi64:367; F16C: # %bb.0:368; F16C-NEXT: vpextrw $0, %xmm0, %eax369; F16C-NEXT: movzwl %ax, %eax370; F16C-NEXT: vmovd %eax, %xmm0371; F16C-NEXT: vcvtph2ps %xmm0, %xmm0372; F16C-NEXT: vmovss {{.*#+}} xmm1 = [9.22337203E+18,0.0E+0,0.0E+0,0.0E+0]373; F16C-NEXT: vcomiss %xmm1, %xmm0374; F16C-NEXT: vxorps %xmm2, %xmm2, %xmm2375; F16C-NEXT: jb .LBB9_2376; F16C-NEXT: # %bb.1:377; F16C-NEXT: vmovaps %xmm1, %xmm2378; F16C-NEXT: .LBB9_2:379; F16C-NEXT: vsubss %xmm2, %xmm0, %xmm0380; F16C-NEXT: vcvttss2si %xmm0, %rcx381; F16C-NEXT: setae %al382; F16C-NEXT: movzbl %al, %eax383; F16C-NEXT: shlq $63, %rax384; F16C-NEXT: xorq %rcx, %rax385; F16C-NEXT: retq386;387; AVX512-LABEL: fptoui_f16toi64:388; AVX512: # %bb.0:389; AVX512-NEXT: vpextrw $0, %xmm0, %eax390; AVX512-NEXT: movzwl %ax, %eax391; AVX512-NEXT: vmovd %eax, %xmm0392; AVX512-NEXT: vcvtph2ps %xmm0, %xmm0393; AVX512-NEXT: vcvttss2usi %xmm0, %rax394; AVX512-NEXT: retq395;396; X86-LABEL: fptoui_f16toi64:397; X86: # %bb.0:398; X86-NEXT: vmovsh {{.*#+}} xmm0 = mem[0],zero,zero,zero,zero,zero,zero,zero399; X86-NEXT: vcvttph2uqq %xmm0, %xmm0400; X86-NEXT: vmovd %xmm0, %eax401; X86-NEXT: vpextrd $1, %xmm0, %edx402; X86-NEXT: retl403;404; X64-LABEL: fptoui_f16toi64:405; X64: # %bb.0:406; X64-NEXT: vcvttsh2usi %xmm0, %rax407; X64-NEXT: retq408 %result = call i64 @llvm.experimental.constrained.fptoui.i64.f16(half %x,409 metadata !"fpexcept.strict") #0410 ret i64 %result411}412 413attributes #0 = { strictfp nounwind }414