1; RUN: llc < %s -mtriple=x86_64-apple-macosx | FileCheck --check-prefix=CHECK --check-prefix=SSE --check-prefix=CST %s 2; RUN: llc < %s -mtriple=x86_64-apple-macosx -mattr=+sse4.1 | FileCheck --check-prefix=CHECK --check-prefix=SSE41 --check-prefix=CST %s 3; RUN: llc < %s -mtriple=x86_64-apple-macosx -mattr=+avx | FileCheck --check-prefix=CHECK --check-prefix=AVX --check-prefix=CST %s 4; RUN: llc < %s -mtriple=x86_64-apple-macosx -mattr=+avx2 | FileCheck --check-prefix=CHECK --check-prefix=AVX2 %s 5 6; Check that the constant used in the vectors are the right ones. 7; SSE: [[MASKCSTADDR:LCPI0_[0-9]+]]: 8; SSE-NEXT: .long 65535 ## 0xffff 9; SSE-NEXT: .long 65535 ## 0xffff 10; SSE-NEXT: .long 65535 ## 0xffff 11; SSE-NEXT: .long 65535 ## 0xffff 12 13; CST: [[LOWCSTADDR:LCPI0_[0-9]+]]: 14; CST-NEXT: .long 1258291200 ## 0x4b000000 15; CST-NEXT: .long 1258291200 ## 0x4b000000 16; CST-NEXT: .long 1258291200 ## 0x4b000000 17; CST-NEXT: .long 1258291200 ## 0x4b000000 18 19; CST: [[HIGHCSTADDR:LCPI0_[0-9]+]]: 20; CST-NEXT: .long 1392508928 ## 0x53000000 21; CST-NEXT: .long 1392508928 ## 0x53000000 22; CST-NEXT: .long 1392508928 ## 0x53000000 23; CST-NEXT: .long 1392508928 ## 0x53000000 24 25; CST: [[MAGICCSTADDR:LCPI0_[0-9]+]]: 26; CST-NEXT: .long 0x53000080 ## float 5.49764202E+11 27; CST-NEXT: .long 0x53000080 ## float 5.49764202E+11 28; CST-NEXT: .long 0x53000080 ## float 5.49764202E+11 29; CST-NEXT: .long 0x53000080 ## float 5.49764202E+11 30 31; AVX2: [[LOWCSTADDR:LCPI0_[0-9]+]]: 32; AVX2-NEXT: .long 1258291200 ## 0x4b000000 33 34; AVX2: [[HIGHCSTADDR:LCPI0_[0-9]+]]: 35; AVX2-NEXT: .long 1392508928 ## 0x53000000 36 37; AVX2: [[MAGICCSTADDR:LCPI0_[0-9]+]]: 38; AVX2-NEXT: .long 0x53000080 ## float 5.49764202E+11 39 40define <4 x float> @test1(<4 x i32> %A) nounwind { 41; CHECK-LABEL: test1: 42; 43; SSE: movdqa [[MASKCSTADDR]](%rip), [[MASK:%xmm[0-9]+]] 44; SSE-NEXT: pand %xmm0, [[MASK]] 45; After this instruction, MASK will have the value of the low parts 46; of the vector. 47; SSE-NEXT: por [[LOWCSTADDR]](%rip), [[MASK]] 48; SSE-NEXT: psrld $16, %xmm0 49; SSE-NEXT: por [[HIGHCSTADDR]](%rip), %xmm0 50; SSE-NEXT: subps [[MAGICCSTADDR]](%rip), %xmm0 51; SSE-NEXT: addps [[MASK]], %xmm0 52; SSE-NEXT: retq 53; 54; Currently we commute the arguments of the first blend, but this could be 55; improved to match the lowering of the second blend. 56; SSE41: movdqa [[LOWCSTADDR]](%rip), [[LOWVEC:%xmm[0-9]+]] 57; SSE41-NEXT: pblendw $85, %xmm0, [[LOWVEC]] 58; SSE41-NEXT: psrld $16, %xmm0 59; SSE41-NEXT: pblendw $170, [[HIGHCSTADDR]](%rip), %xmm0 60; SSE41-NEXT: subps [[MAGICCSTADDR]](%rip), %xmm0 61; SSE41-NEXT: addps [[LOWVEC]], %xmm0 62; SSE41-NEXT: retq 63; 64; AVX: vpblendw $170, [[LOWCSTADDR]](%rip), %xmm0, [[LOWVEC:%xmm[0-9]+]] 65; AVX-NEXT: vpsrld $16, %xmm0, [[SHIFTVEC:%xmm[0-9]+]] 66; AVX-NEXT: vpblendw $170, [[HIGHCSTADDR]](%rip), [[SHIFTVEC]], [[HIGHVEC:%xmm[0-9]+]] 67; AVX-NEXT: vsubps [[MAGICCSTADDR]](%rip), [[HIGHVEC]], [[TMP:%xmm[0-9]+]] 68; AVX-NEXT: vaddps [[TMP]], [[LOWVEC]], %xmm0 69; AVX-NEXT: retq 70; 71; The lowering for AVX2 is a bit messy, because we select broadcast 72; instructions, instead of folding the constant loads. 73; AVX2: vpbroadcastd [[LOWCSTADDR]](%rip), [[LOWCST:%xmm[0-9]+]] 74; AVX2-NEXT: vpblendw $170, [[LOWCST]], %xmm0, [[LOWVEC:%xmm[0-9]+]] 75; AVX2-NEXT: vpsrld $16, %xmm0, [[SHIFTVEC:%xmm[0-9]+]] 76; AVX2-NEXT: vpbroadcastd [[HIGHCSTADDR]](%rip), [[HIGHCST:%xmm[0-9]+]] 77; AVX2-NEXT: vpblendw $170, [[HIGHCST]], [[SHIFTVEC]], [[HIGHVEC:%xmm[0-9]+]] 78; AVX2-NEXT: vbroadcastss [[MAGICCSTADDR]](%rip), [[MAGICCST:%xmm[0-9]+]] 79; AVX2-NEXT: vsubps [[MAGICCST]], [[HIGHVEC]], [[TMP:%xmm[0-9]+]] 80; AVX2-NEXT: vaddps [[TMP]], [[LOWVEC]], %xmm0 81; AVX2-NEXT: retq 82 %C = uitofp <4 x i32> %A to <4 x float> 83 ret <4 x float> %C 84} 85 86; Match the AVX2 constants used in the next function 87; AVX2: [[LOWCSTADDR:LCPI1_[0-9]+]]: 88; AVX2-NEXT: .long 1258291200 ## 0x4b000000 89 90; AVX2: [[HIGHCSTADDR:LCPI1_[0-9]+]]: 91; AVX2-NEXT: .long 1392508928 ## 0x53000000 92 93; AVX2: [[MAGICCSTADDR:LCPI1_[0-9]+]]: 94; AVX2-NEXT: .long 0x53000080 ## float 5.49764202E+11 95 96define <8 x float> @test2(<8 x i32> %A) nounwind { 97; CHECK-LABEL: test2: 98; Legalization will break the thing is 2 x <4 x i32> on anthing prior AVX. 99; The constant used for in the vector instruction are shared between the 100; two sequences of instructions. 101; 102; SSE: movdqa {{.*#+}} [[MASK:xmm[0-9]+]] = [65535,65535,65535,65535] 103; SSE-NEXT: movdqa %xmm0, [[VECLOW:%xmm[0-9]+]] 104; SSE-NEXT: pand %[[MASK]], [[VECLOW]] 105; SSE-NEXT: movdqa {{.*#+}} [[LOWCST:xmm[0-9]+]] = [1258291200,1258291200,1258291200,1258291200] 106; SSE-NEXT: por %[[LOWCST]], [[VECLOW]] 107; SSE-NEXT: psrld $16, %xmm0 108; SSE-NEXT: movdqa {{.*#+}} [[HIGHCST:xmm[0-9]+]] = [1392508928,1392508928,1392508928,1392508928] 109; SSE-NEXT: por %[[HIGHCST]], %xmm0 110; SSE-NEXT: movaps {{.*#+}} [[MAGICCST:xmm[0-9]+]] = [5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11] 111; SSE-NEXT: subps %[[MAGICCST]], %xmm0 112; SSE-NEXT: addps [[VECLOW]], %xmm0 113; MASK is the low vector of the second part after this point. 114; SSE-NEXT: pand %xmm1, %[[MASK]] 115; SSE-NEXT: por %[[LOWCST]], %[[MASK]] 116; SSE-NEXT: psrld $16, %xmm1 117; SSE-NEXT: por %[[HIGHCST]], %xmm1 118; SSE-NEXT: subps %[[MAGICCST]], %xmm1 119; SSE-NEXT: addps %[[MASK]], %xmm1 120; SSE-NEXT: retq 121; 122; SSE41: movdqa {{.*#+}} [[LOWCST:xmm[0-9]+]] = [1258291200,1258291200,1258291200,1258291200] 123; SSE41-NEXT: movdqa %xmm0, [[VECLOW:%xmm[0-9]+]] 124; SSE41-NEXT: pblendw $170, %[[LOWCST]], [[VECLOW]] 125; SSE41-NEXT: psrld $16, %xmm0 126; SSE41-NEXT: movdqa {{.*#+}} [[HIGHCST:xmm[0-9]+]] = [1392508928,1392508928,1392508928,1392508928] 127; SSE41-NEXT: pblendw $170, %[[HIGHCST]], %xmm0 128; SSE41-NEXT: movaps {{.*#+}} [[MAGICCST:xmm[0-9]+]] = [5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11] 129; SSE41-NEXT: subps %[[MAGICCST]], %xmm0 130; SSE41-NEXT: addps [[VECLOW]], %xmm0 131; LOWCST is the low vector of the second part after this point. 132; The operands of the blend are inverted because we reuse xmm1 133; in the next shift. 134; SSE41-NEXT: pblendw $85, %xmm1, %[[LOWCST]] 135; SSE41-NEXT: psrld $16, %xmm1 136; SSE41-NEXT: pblendw $170, %[[HIGHCST]], %xmm1 137; SSE41-NEXT: subps %[[MAGICCST]], %xmm1 138; SSE41-NEXT: addps %[[LOWCST]], %xmm1 139; SSE41-NEXT: retq 140; 141; Test that we are not lowering uinttofp to scalars 142; AVX-NOT: cvtsd2ss 143; AVX: retq 144; 145; AVX2: vpbroadcastd [[LOWCSTADDR]](%rip), [[LOWCST:%ymm[0-9]+]] 146; AVX2-NEXT: vpblendw $170, [[LOWCST]], %ymm0, [[LOWVEC:%ymm[0-9]+]] 147; AVX2-NEXT: vpsrld $16, %ymm0, [[SHIFTVEC:%ymm[0-9]+]] 148; AVX2-NEXT: vpbroadcastd [[HIGHCSTADDR]](%rip), [[HIGHCST:%ymm[0-9]+]] 149; AVX2-NEXT: vpblendw $170, [[HIGHCST]], [[SHIFTVEC]], [[HIGHVEC:%ymm[0-9]+]] 150; AVX2-NEXT: vbroadcastss [[MAGICCSTADDR]](%rip), [[MAGICCST:%ymm[0-9]+]] 151; AVX2-NEXT: vsubps [[MAGICCST]], [[HIGHVEC]], [[TMP:%ymm[0-9]+]] 152; AVX2-NEXT: vaddps [[TMP]], [[LOWVEC]], %ymm0 153; AVX2-NEXT: retq 154 %C = uitofp <8 x i32> %A to <8 x float> 155 ret <8 x float> %C 156} 157 158define <4 x double> @test3(<4 x i32> %arg) { 159; CHECK-LABEL: test3: 160; This test used to crash because we were custom lowering it as if it was 161; a conversion between <4 x i32> and <4 x float>. 162; AVX: vsubpd 163; AVX2: vsubpd 164; CHECK: retq 165 %tmp = uitofp <4 x i32> %arg to <4 x double> 166 ret <4 x double> %tmp 167} 168