95 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=i686-linux-pc -mcpu=corei7 | FileCheck %s --check-prefix=X863; RUN: llc < %s -mtriple=x86_64-linux-pc -mcpu=corei7 | FileCheck %s --check-prefix=X644 5; uitofp <2 x i32> codegen from buildvector or legalization is different but gives the same results6; across the full 0 - 0xFFFFFFFF u32 range.7 8define <2 x float> @uitofp_2i32_cvt_buildvector(i32 %x, i32 %y, <2 x float> %v) {9; X86-LABEL: uitofp_2i32_cvt_buildvector:10; X86: # %bb.0:11; X86-NEXT: movss {{.*#+}} xmm1 = mem[0],zero,zero,zero12; X86-NEXT: movsd {{.*#+}} xmm2 = [4.503599627370496E+15,0.0E+0]13; X86-NEXT: orpd %xmm2, %xmm114; X86-NEXT: subsd %xmm2, %xmm115; X86-NEXT: cvtsd2ss %xmm1, %xmm116; X86-NEXT: movss {{.*#+}} xmm3 = mem[0],zero,zero,zero17; X86-NEXT: orpd %xmm2, %xmm318; X86-NEXT: subsd %xmm2, %xmm319; X86-NEXT: xorps %xmm2, %xmm220; X86-NEXT: cvtsd2ss %xmm3, %xmm221; X86-NEXT: insertps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[2,3]22; X86-NEXT: mulps %xmm1, %xmm023; X86-NEXT: retl24;25; X64-LABEL: uitofp_2i32_cvt_buildvector:26; X64: # %bb.0:27; X64-NEXT: movl %edi, %eax28; X64-NEXT: cvtsi2ss %rax, %xmm129; X64-NEXT: movl %esi, %eax30; X64-NEXT: cvtsi2ss %rax, %xmm231; X64-NEXT: insertps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[2,3]32; X64-NEXT: mulps %xmm1, %xmm033; X64-NEXT: retq34 %t1 = uitofp i32 %x to float35 %t2 = insertelement <2 x float> undef, float %t1, i32 036 %t3 = uitofp i32 %y to float37 %t4 = insertelement <2 x float> %t2, float %t3, i32 138 %t5 = fmul <2 x float> %v, %t439 ret <2 x float> %t540}41 42define <2 x float> @uitofp_2i32_buildvector_cvt(i32 %x, i32 %y, <2 x float> %v) {43; X86-LABEL: uitofp_2i32_buildvector_cvt:44; X86: # %bb.0:45; X86-NEXT: movdqa {{.*#+}} xmm1 = [4.503599627370496E+15,4.503599627370496E+15]46; X86-NEXT: pmovzxdq {{.*#+}} xmm2 = mem[0],zero,mem[1],zero47; X86-NEXT: por %xmm1, %xmm248; X86-NEXT: subpd %xmm1, %xmm249; X86-NEXT: cvtpd2ps %xmm2, %xmm150; X86-NEXT: mulps %xmm1, %xmm051; X86-NEXT: retl52;53; X64-LABEL: uitofp_2i32_buildvector_cvt:54; X64: # %bb.0:55; X64-NEXT: movd %edi, %xmm156; X64-NEXT: pinsrd $1, %esi, %xmm157; X64-NEXT: pmovzxdq {{.*#+}} xmm1 = xmm1[0],zero,xmm1[1],zero58; X64-NEXT: movdqa {{.*#+}} xmm2 = [4.503599627370496E+15,4.503599627370496E+15]59; X64-NEXT: por %xmm2, %xmm160; X64-NEXT: subpd %xmm2, %xmm161; X64-NEXT: cvtpd2ps %xmm1, %xmm162; X64-NEXT: mulps %xmm1, %xmm063; X64-NEXT: retq64 %t1 = insertelement <2 x i32> undef, i32 %x, i32 065 %t2 = insertelement <2 x i32> %t1, i32 %y, i32 166 %t3 = uitofp <2 x i32> %t2 to <2 x float>67 %t4 = fmul <2 x float> %v, %t368 ret <2 x float> %t469}70 71define <2 x float> @uitofp_2i32_legalized(<2 x i32> %in, <2 x float> %v) {72; X86-LABEL: uitofp_2i32_legalized:73; X86: # %bb.0:74; X86-NEXT: pmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero75; X86-NEXT: movdqa {{.*#+}} xmm2 = [4.503599627370496E+15,4.503599627370496E+15]76; X86-NEXT: por %xmm2, %xmm077; X86-NEXT: subpd %xmm2, %xmm078; X86-NEXT: cvtpd2ps %xmm0, %xmm079; X86-NEXT: mulps %xmm1, %xmm080; X86-NEXT: retl81;82; X64-LABEL: uitofp_2i32_legalized:83; X64: # %bb.0:84; X64-NEXT: pmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero85; X64-NEXT: movdqa {{.*#+}} xmm2 = [4.503599627370496E+15,4.503599627370496E+15]86; X64-NEXT: por %xmm2, %xmm087; X64-NEXT: subpd %xmm2, %xmm088; X64-NEXT: cvtpd2ps %xmm0, %xmm089; X64-NEXT: mulps %xmm1, %xmm090; X64-NEXT: retq91 %t1 = uitofp <2 x i32> %in to <2 x float>92 %t2 = fmul <2 x float> %v, %t193 ret <2 x float> %t294}95