1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py 2; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -mcpu=corei7 -mattr=-f16c | FileCheck %s -check-prefix=LIBCALL 3; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -mcpu=corei7 -mattr=+f16c | FileCheck %s -check-prefix=F16C 4; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -mcpu=corei7 -mattr=-f16c,+soft-float | FileCheck %s -check-prefix=SOFTFLOAT 5; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -mcpu=corei7 -mattr=+f16c,+soft-float | FileCheck %s -check-prefix=SOFTFLOAT 6 7; This is a test for float to half float conversions on x86-64. 8; 9; If flag -soft-float is set, or if there is no F16C support, then: 10; 1) half float to float conversions are 11; translated into calls to __gnu_h2f_ieee defined 12; by the compiler runtime library; 13; 2) float to half float conversions are translated into calls 14; to __gnu_f2h_ieee which expected to be defined by the 15; compiler runtime library. 16; 17; Otherwise (we have F16C support): 18; 1) half float to float conversion are translated using 19; vcvtph2ps instructions; 20; 2) float to half float conversions are translated using 21; vcvtps2ph instructions 22 23 24define void @test1(float %src, ptr %dest) nounwind { 25; LIBCALL-LABEL: test1: 26; LIBCALL: # %bb.0: 27; LIBCALL-NEXT: pushq %rbx 28; LIBCALL-NEXT: movq %rdi, %rbx 29; LIBCALL-NEXT: callq __truncsfhf2@PLT 30; LIBCALL-NEXT: pextrw $0, %xmm0, (%rbx) 31; LIBCALL-NEXT: popq %rbx 32; LIBCALL-NEXT: retq 33; 34; F16C-LABEL: test1: 35; F16C: # %bb.0: 36; F16C-NEXT: vcvtps2ph $4, %xmm0, %xmm0 37; F16C-NEXT: vpextrw $0, %xmm0, (%rdi) 38; F16C-NEXT: retq 39; 40; SOFTFLOAT-LABEL: test1: 41; SOFTFLOAT: # %bb.0: 42; SOFTFLOAT-NEXT: pushq %rbx 43; SOFTFLOAT-NEXT: movq %rsi, %rbx 44; SOFTFLOAT-NEXT: callq __gnu_f2h_ieee@PLT 45; SOFTFLOAT-NEXT: movw %ax, (%rbx) 46; SOFTFLOAT-NEXT: popq %rbx 47; SOFTFLOAT-NEXT: retq 48 %1 = tail call i16 @llvm.convert.to.fp16.f32(float %src) 49 store i16 %1, ptr %dest, align 2 50 ret void 51} 52 53define float @test2(ptr nocapture %src) nounwind { 54; LIBCALL-LABEL: test2: 55; LIBCALL: # %bb.0: 56; LIBCALL-NEXT: pinsrw $0, (%rdi), %xmm0 57; LIBCALL-NEXT: jmp __extendhfsf2@PLT # TAILCALL 58; 59; F16C-LABEL: test2: 60; F16C: # %bb.0: 61; F16C-NEXT: vpinsrw $0, (%rdi), %xmm0, %xmm0 62; F16C-NEXT: vcvtph2ps %xmm0, %xmm0 63; F16C-NEXT: retq 64; 65; SOFTFLOAT-LABEL: test2: 66; SOFTFLOAT: # %bb.0: 67; SOFTFLOAT-NEXT: pushq %rax 68; SOFTFLOAT-NEXT: movzwl (%rdi), %edi 69; SOFTFLOAT-NEXT: callq __gnu_h2f_ieee@PLT 70; SOFTFLOAT-NEXT: popq %rcx 71; SOFTFLOAT-NEXT: retq 72 %1 = load i16, ptr %src, align 2 73 %2 = tail call float @llvm.convert.from.fp16.f32(i16 %1) 74 ret float %2 75} 76 77define float @test3(float %src) nounwind uwtable readnone { 78; LIBCALL-LABEL: test3: 79; LIBCALL: # %bb.0: 80; LIBCALL-NEXT: pushq %rax 81; LIBCALL-NEXT: .cfi_def_cfa_offset 16 82; LIBCALL-NEXT: callq __truncsfhf2@PLT 83; LIBCALL-NEXT: popq %rax 84; LIBCALL-NEXT: .cfi_def_cfa_offset 8 85; LIBCALL-NEXT: jmp __extendhfsf2@PLT # TAILCALL 86; 87; F16C-LABEL: test3: 88; F16C: # %bb.0: 89; F16C-NEXT: vcvtps2ph $4, %xmm0, %xmm0 90; F16C-NEXT: vcvtph2ps %xmm0, %xmm0 91; F16C-NEXT: retq 92; 93; SOFTFLOAT-LABEL: test3: 94; SOFTFLOAT: # %bb.0: 95; SOFTFLOAT-NEXT: pushq %rax 96; SOFTFLOAT-NEXT: .cfi_def_cfa_offset 16 97; SOFTFLOAT-NEXT: callq __gnu_f2h_ieee@PLT 98; SOFTFLOAT-NEXT: movzwl %ax, %edi 99; SOFTFLOAT-NEXT: callq __gnu_h2f_ieee@PLT 100; SOFTFLOAT-NEXT: popq %rcx 101; SOFTFLOAT-NEXT: .cfi_def_cfa_offset 8 102; SOFTFLOAT-NEXT: retq 103 %1 = tail call i16 @llvm.convert.to.fp16.f32(float %src) 104 %2 = tail call float @llvm.convert.from.fp16.f32(i16 %1) 105 ret float %2 106} 107 108define double @test4(ptr nocapture %src) nounwind { 109; LIBCALL-LABEL: test4: 110; LIBCALL: # %bb.0: 111; LIBCALL-NEXT: pushq %rax 112; LIBCALL-NEXT: pinsrw $0, (%rdi), %xmm0 113; LIBCALL-NEXT: callq __extendhfsf2@PLT 114; LIBCALL-NEXT: cvtss2sd %xmm0, %xmm0 115; LIBCALL-NEXT: popq %rax 116; LIBCALL-NEXT: retq 117; 118; F16C-LABEL: test4: 119; F16C: # %bb.0: 120; F16C-NEXT: vpinsrw $0, (%rdi), %xmm0, %xmm0 121; F16C-NEXT: vcvtph2ps %xmm0, %xmm0 122; F16C-NEXT: vcvtss2sd %xmm0, %xmm0, %xmm0 123; F16C-NEXT: retq 124; 125; SOFTFLOAT-LABEL: test4: 126; SOFTFLOAT: # %bb.0: 127; SOFTFLOAT-NEXT: pushq %rax 128; SOFTFLOAT-NEXT: movzwl (%rdi), %edi 129; SOFTFLOAT-NEXT: callq __gnu_h2f_ieee@PLT 130; SOFTFLOAT-NEXT: movl %eax, %edi 131; SOFTFLOAT-NEXT: callq __extendsfdf2@PLT 132; SOFTFLOAT-NEXT: popq %rcx 133; SOFTFLOAT-NEXT: retq 134 %1 = load i16, ptr %src, align 2 135 %2 = tail call double @llvm.convert.from.fp16.f64(i16 %1) 136 ret double %2 137} 138 139define i16 @test5(double %src) nounwind { 140; LIBCALL-LABEL: test5: 141; LIBCALL: # %bb.0: 142; LIBCALL-NEXT: pushq %rax 143; LIBCALL-NEXT: callq __truncdfhf2@PLT 144; LIBCALL-NEXT: pextrw $0, %xmm0, %eax 145; LIBCALL-NEXT: # kill: def $ax killed $ax killed $eax 146; LIBCALL-NEXT: popq %rcx 147; LIBCALL-NEXT: retq 148; 149; F16C-LABEL: test5: 150; F16C: # %bb.0: 151; F16C-NEXT: pushq %rax 152; F16C-NEXT: callq __truncdfhf2@PLT 153; F16C-NEXT: vpextrw $0, %xmm0, %eax 154; F16C-NEXT: # kill: def $ax killed $ax killed $eax 155; F16C-NEXT: popq %rcx 156; F16C-NEXT: retq 157; 158; SOFTFLOAT-LABEL: test5: 159; SOFTFLOAT: # %bb.0: 160; SOFTFLOAT-NEXT: pushq %rax 161; SOFTFLOAT-NEXT: callq __truncdfhf2@PLT 162; SOFTFLOAT-NEXT: popq %rcx 163; SOFTFLOAT-NEXT: retq 164 %val = tail call i16 @llvm.convert.to.fp16.f64(double %src) 165 ret i16 %val 166} 167 168declare float @llvm.convert.from.fp16.f32(i16) nounwind readnone 169declare i16 @llvm.convert.to.fp16.f32(float) nounwind readnone 170declare double @llvm.convert.from.fp16.f64(i16) nounwind readnone 171declare i16 @llvm.convert.to.fp16.f64(double) nounwind readnone 172