1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py 2; RUN: llc < %s -mtriple=i386-- -mattr=sse2 | FileCheck %s --check-prefixes=X86,X86-SSE 3; RUN: llc < %s -mtriple=i386-- -mattr=sse4.1 | FileCheck %s --check-prefixes=X86,X86-SSE 4; RUN: llc < %s -mtriple=i386-- -mattr=avx2 | FileCheck %s --check-prefixes=X86,X86-AVX2 5; RUN: llc < %s -mtriple=i386-- -mattr=avx512f | FileCheck %s --check-prefixes=X86,X86-AVX512F 6; RUN: llc < %s -mtriple=x86_64-- -mattr=sse2 | FileCheck %s --check-prefix=X64-SSE 7; RUN: llc < %s -mtriple=x86_64-- -mattr=sse4.1 | FileCheck %s --check-prefix=X64-SSE 8; RUN: llc < %s -mtriple=x86_64-- -mattr=avx2 | FileCheck %s --check-prefixes=X64-AVX,X64-AVX2 9; RUN: llc < %s -mtriple=x86_64-- -mattr=avx512f | FileCheck %s --check-prefixes=X64-AVX,X64-AVX512F 10 11; This should do a single load into the fp stack for the return, not diddle with xmm registers. 12 13define float @icmp_select_fp_constants(i32 %x) nounwind readnone { 14; X86-LABEL: icmp_select_fp_constants: 15; X86: # %bb.0: 16; X86-NEXT: xorl %eax, %eax 17; X86-NEXT: cmpl $0, {{[0-9]+}}(%esp) 18; X86-NEXT: sete %al 19; X86-NEXT: flds {{\.LCPI.*}}(,%eax,4) 20; X86-NEXT: retl 21; 22; X64-SSE-LABEL: icmp_select_fp_constants: 23; X64-SSE: # %bb.0: 24; X64-SSE-NEXT: xorl %eax, %eax 25; X64-SSE-NEXT: testl %edi, %edi 26; X64-SSE-NEXT: sete %al 27; X64-SSE-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero 28; X64-SSE-NEXT: retq 29; 30; X64-AVX-LABEL: icmp_select_fp_constants: 31; X64-AVX: # %bb.0: 32; X64-AVX-NEXT: xorl %eax, %eax 33; X64-AVX-NEXT: testl %edi, %edi 34; X64-AVX-NEXT: sete %al 35; X64-AVX-NEXT: vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero 36; X64-AVX-NEXT: retq 37 %c = icmp eq i32 %x, 0 38 %r = select i1 %c, float 42.0, float 23.0 39 ret float %r 40} 41 42define float @fcmp_select_fp_constants(float %x) nounwind readnone { 43; X86-SSE-LABEL: fcmp_select_fp_constants: 44; X86-SSE: # %bb.0: 45; X86-SSE-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero 46; X86-SSE-NEXT: cmpneqss {{[0-9]+}}(%esp), %xmm0 47; X86-SSE-NEXT: movd %xmm0, %eax 48; X86-SSE-NEXT: andl $1, %eax 49; X86-SSE-NEXT: flds {{\.LCPI.*}}(,%eax,4) 50; X86-SSE-NEXT: retl 51; 52; X86-AVX2-LABEL: fcmp_select_fp_constants: 53; X86-AVX2: # %bb.0: 54; X86-AVX2-NEXT: vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero 55; X86-AVX2-NEXT: vcmpneqss {{[0-9]+}}(%esp), %xmm0, %xmm0 56; X86-AVX2-NEXT: vmovd %xmm0, %eax 57; X86-AVX2-NEXT: andl $1, %eax 58; X86-AVX2-NEXT: flds {{\.LCPI.*}}(,%eax,4) 59; X86-AVX2-NEXT: retl 60; 61; X86-AVX512F-LABEL: fcmp_select_fp_constants: 62; X86-AVX512F: # %bb.0: 63; X86-AVX512F-NEXT: vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero 64; X86-AVX512F-NEXT: vcmpneqss {{\.LCPI.*}}, %xmm0, %k0 65; X86-AVX512F-NEXT: kmovw %k0, %eax 66; X86-AVX512F-NEXT: flds {{\.LCPI.*}}(,%eax,4) 67; X86-AVX512F-NEXT: retl 68; 69; X64-SSE-LABEL: fcmp_select_fp_constants: 70; X64-SSE: # %bb.0: 71; X64-SSE-NEXT: cmpneqss {{.*}}(%rip), %xmm0 72; X64-SSE-NEXT: movd %xmm0, %eax 73; X64-SSE-NEXT: andl $1, %eax 74; X64-SSE-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero 75; X64-SSE-NEXT: retq 76; 77; X64-AVX2-LABEL: fcmp_select_fp_constants: 78; X64-AVX2: # %bb.0: 79; X64-AVX2-NEXT: vcmpneqss {{.*}}(%rip), %xmm0, %xmm0 80; X64-AVX2-NEXT: vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero 81; X64-AVX2-NEXT: vmovss {{.*#+}} xmm2 = mem[0],zero,zero,zero 82; X64-AVX2-NEXT: vblendvps %xmm0, %xmm1, %xmm2, %xmm0 83; X64-AVX2-NEXT: retq 84; 85; X64-AVX512F-LABEL: fcmp_select_fp_constants: 86; X64-AVX512F: # %bb.0: 87; X64-AVX512F-NEXT: vcmpneqss {{.*}}(%rip), %xmm0, %k1 88; X64-AVX512F-NEXT: vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero 89; X64-AVX512F-NEXT: vmovss {{.*}}(%rip), %xmm0 {%k1} 90; X64-AVX512F-NEXT: retq 91 %c = fcmp une float %x, -4.0 92 %r = select i1 %c, float 42.0, float 23.0 93 ret float %r 94} 95 96