llvm/test/CodeGen/X86/cvtv2f32.ll
Simon Pilgrim 7ee24bd8b6 [X86][SSE] Let 64-bit targets use the fast 2i32-2f32 UINT_TO_FP conversion as well as 32-bit
The 2i32-2i64 legalization means that we can use the slightly quicker double bits + fptrunc approach for the same results



git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@277271 91177308-0d34-0410-b5e6-96231b3b80d8
2016-07-30 14:06:59 +00:00

67 lines
2.7 KiB
LLVM

; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc < %s -mtriple=i686-linux-pc -mcpu=corei7 | FileCheck %s --check-prefix=X32
; RUN: llc < %s -mtriple=x86_64-linux-pc -mcpu=corei7 | FileCheck %s --check-prefix=X64
; uitofp <2 x i32> codegen from buildvector or legalization is different but gives the same results
; across the full 0 - 0xFFFFFFFF u32 range.
define <2 x float> @uitofp_2i32_buildvector(i32 %x, i32 %y, <2 x float> %v) {
; X32-LABEL: uitofp_2i32_buildvector:
; X32: # BB#0:
; X32-NEXT: movq {{.*#+}} xmm1 = mem[0],zero
; X32-NEXT: movdqa {{.*#+}} xmm2 = [1258291200,1258291200,1258291200,1258291200]
; X32-NEXT: pblendw {{.*#+}} xmm2 = xmm1[0],xmm2[1],xmm1[2],xmm2[3],xmm1[4],xmm2[5],xmm1[6],xmm2[7]
; X32-NEXT: psrld $16, %xmm1
; X32-NEXT: pblendw {{.*#+}} xmm1 = xmm1[0],mem[1],xmm1[2],mem[3],xmm1[4],mem[5],xmm1[6],mem[7]
; X32-NEXT: addps {{\.LCPI.*}}, %xmm1
; X32-NEXT: addps %xmm2, %xmm1
; X32-NEXT: mulps %xmm1, %xmm0
; X32-NEXT: retl
;
; X64-LABEL: uitofp_2i32_buildvector:
; X64: # BB#0:
; X64-NEXT: movd %edi, %xmm1
; X64-NEXT: pinsrd $1, %esi, %xmm1
; X64-NEXT: movdqa {{.*#+}} xmm2 = [1258291200,1258291200,1258291200,1258291200]
; X64-NEXT: pblendw {{.*#+}} xmm2 = xmm1[0],xmm2[1],xmm1[2],xmm2[3],xmm1[4],xmm2[5],xmm1[6],xmm2[7]
; X64-NEXT: psrld $16, %xmm1
; X64-NEXT: pblendw {{.*#+}} xmm1 = xmm1[0],mem[1],xmm1[2],mem[3],xmm1[4],mem[5],xmm1[6],mem[7]
; X64-NEXT: addps {{.*}}(%rip), %xmm1
; X64-NEXT: addps %xmm2, %xmm1
; X64-NEXT: mulps %xmm1, %xmm0
; X64-NEXT: retq
%t1 = uitofp i32 %x to float
%t2 = insertelement <2 x float> undef, float %t1, i32 0
%t3 = uitofp i32 %y to float
%t4 = insertelement <2 x float> %t2, float %t3, i32 1
%t5 = fmul <2 x float> %v, %t4
ret <2 x float> %t5
}
define <2 x float> @uitofp_2i32_legalized(<2 x i32> %in, <2 x float> %v) {
; X32-LABEL: uitofp_2i32_legalized:
; X32: # BB#0:
; X32-NEXT: pxor %xmm2, %xmm2
; X32-NEXT: pblendw {{.*#+}} xmm2 = xmm0[0,1],xmm2[2,3],xmm0[4,5],xmm2[6,7]
; X32-NEXT: movdqa {{.*#+}} xmm0 = [4.503600e+15,4.503600e+15]
; X32-NEXT: por %xmm0, %xmm2
; X32-NEXT: subpd %xmm0, %xmm2
; X32-NEXT: cvtpd2ps %xmm2, %xmm0
; X32-NEXT: mulps %xmm1, %xmm0
; X32-NEXT: retl
;
; X64-LABEL: uitofp_2i32_legalized:
; X64: # BB#0:
; X64-NEXT: pxor %xmm2, %xmm2
; X64-NEXT: pblendw {{.*#+}} xmm2 = xmm0[0,1],xmm2[2,3],xmm0[4,5],xmm2[6,7]
; X64-NEXT: movdqa {{.*#+}} xmm0 = [4.503600e+15,4.503600e+15]
; X64-NEXT: por %xmm0, %xmm2
; X64-NEXT: subpd %xmm0, %xmm2
; X64-NEXT: cvtpd2ps %xmm2, %xmm0
; X64-NEXT: mulps %xmm1, %xmm0
; X64-NEXT: retq
%t1 = uitofp <2 x i32> %in to <2 x float>
%t2 = fmul <2 x float> %v, %t1
ret <2 x float> %t2
}