1; RUN: llc < %s -mtriple=x86_64-apple-darwin | FileCheck %s 2 3define <4 x float> @foo(<4 x float> %val, <4 x float> %test) nounwind { 4; CHECK-LABEL: LCPI0_0: 5; CHECK-NEXT: .long 1065353216 ## 0x3f800000 6; CHECK-NEXT: .long 1065353216 ## 0x3f800000 7; CHECK-NEXT: .long 1065353216 ## 0x3f800000 8; CHECK-NEXT: .long 1065353216 ## 0x3f800000 9; CHECK-LABEL: foo: 10; CHECK: ## %bb.0: 11; CHECK-NEXT: cmpeqps %xmm1, %xmm0 12; CHECK-NEXT: andps {{.*}}(%rip), %xmm0 13; CHECK-NEXT: retq 14 %cmp = fcmp oeq <4 x float> %val, %test 15 %ext = zext <4 x i1> %cmp to <4 x i32> 16 %result = sitofp <4 x i32> %ext to <4 x float> 17 ret <4 x float> %result 18} 19 20; Make sure the operation doesn't try to get folded when the sizes don't match, 21; as that ends up crashing later when trying to form a bitcast operation for 22; the folded nodes. 23define void @foo1(<4 x float> %val, <4 x float> %test, <4 x double>* %p) nounwind { 24; CHECK-LABEL: LCPI1_0: 25; CHECK-NEXT: .long 1 ## 0x1 26; CHECK-NEXT: .long 1 ## 0x1 27; CHECK-NEXT: .long 1 ## 0x1 28; CHECK-NEXT: .long 1 ## 0x1 29; CHECK-LABEL: foo1: 30; CHECK: ## %bb.0: 31; CHECK-NEXT: cmpeqps %xmm1, %xmm0 32; CHECK-NEXT: andps {{.*}}(%rip), %xmm0 33; CHECK-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3] 34; CHECK-NEXT: cvtdq2pd %xmm1, %xmm1 35; CHECK-NEXT: cvtdq2pd %xmm0, %xmm0 36; CHECK-NEXT: movaps %xmm0, (%rdi) 37; CHECK-NEXT: movaps %xmm1, 16(%rdi) 38; CHECK-NEXT: retq 39 %cmp = fcmp oeq <4 x float> %val, %test 40 %ext = zext <4 x i1> %cmp to <4 x i32> 41 %result = sitofp <4 x i32> %ext to <4 x double> 42 store <4 x double> %result, <4 x double>* %p 43 ret void 44} 45 46; Also test the general purpose constant folding of int->fp. 47define void @foo2(<4 x float>* noalias %result) nounwind { 48; CHECK-LABEL: LCPI2_0: 49; CHECK-NEXT: .long 0x40800000 ## float 4 50; CHECK-NEXT: .long 0x40a00000 ## float 5 51; CHECK-NEXT: .long 0x40c00000 ## float 6 52; CHECK-NEXT: .long 0x40e00000 ## float 7 53; CHECK-LABEL: foo2: 54; CHECK: ## %bb.0: 55; CHECK-NEXT: movaps {{.*#+}} xmm0 = [4.0E+0,5.0E+0,6.0E+0,7.0E+0] 56; CHECK-NEXT: movaps %xmm0, (%rdi) 57; CHECK-NEXT: retq 58 %val = uitofp <4 x i32> <i32 4, i32 5, i32 6, i32 7> to <4 x float> 59 store <4 x float> %val, <4 x float>* %result 60 ret void 61} 62 63; Fold explicit AND operations when the constant isn't a splat of a single 64; scalar value like what the zext creates. 65define <4 x float> @foo3(<4 x float> %val, <4 x float> %test) nounwind { 66; CHECK-LABEL: LCPI3_0: 67; CHECK-NEXT: .long 1065353216 ## 0x3f800000 68; CHECK-NEXT: .long 0 ## 0x0 69; CHECK-NEXT: .long 1065353216 ## 0x3f800000 70; CHECK-NEXT: .long 0 ## 0x0 71; CHECK-LABEL: foo3: 72; CHECK: ## %bb.0: 73; CHECK-NEXT: cmpeqps %xmm1, %xmm0 74; CHECK-NEXT: andps {{.*}}(%rip), %xmm0 75; CHECK-NEXT: retq 76 %cmp = fcmp oeq <4 x float> %val, %test 77 %ext = zext <4 x i1> %cmp to <4 x i32> 78 %and = and <4 x i32> %ext, <i32 255, i32 256, i32 257, i32 258> 79 %result = sitofp <4 x i32> %and to <4 x float> 80 ret <4 x float> %result 81} 82 83; Test the general purpose constant folding of uint->fp. 84define void @foo4(<4 x float>* noalias %result) nounwind { 85; CHECK-LABEL: LCPI4_0: 86; CHECK-NEXT: .long 0x3f800000 ## float 1 87; CHECK-NEXT: .long 0x42fe0000 ## float 127 88; CHECK-NEXT: .long 0x43000000 ## float 128 89; CHECK-NEXT: .long 0x437f0000 ## float 255 90; CHECK-LABEL: foo4: 91; CHECK: ## %bb.0: 92; CHECK-NEXT: movaps {{.*#+}} xmm0 = [1.0E+0,1.27E+2,1.28E+2,2.55E+2] 93; CHECK-NEXT: movaps %xmm0, (%rdi) 94; CHECK-NEXT: retq 95 %val = uitofp <4 x i8> <i8 1, i8 127, i8 -128, i8 -1> to <4 x float> 96 store <4 x float> %val, <4 x float>* %result 97 ret void 98} 99 100; Test when we're masking against a sign extended setcc. 101define <4 x float> @foo5(<4 x i32> %a0, <4 x i32> %a1) { 102; CHECK-LABEL: LCPI5_0: 103; CHECK-NEXT: .long 1065353216 ## 0x3f800000 104; CHECK-NEXT: .long 0 ## 0x0 105; CHECK-NEXT: .long 1065353216 ## 0x3f800000 106; CHECK-NEXT: .long 0 ## 0x0 107; CHECK: ## %bb.0: 108; CHECK-NEXT: pcmpgtd %xmm1, %xmm0 109; CHECK-NEXT: pand {{.*}}(%rip), %xmm0 110; CHECK-NEXT: retq 111 %1 = icmp sgt <4 x i32> %a0, %a1 112 %2 = sext <4 x i1> %1 to <4 x i32> 113 %3 = and <4 x i32> %2, <i32 1, i32 0, i32 1, i32 0> 114 %4 = uitofp <4 x i32> %3 to <4 x float> 115 ret <4 x float> %4 116} 117 118; Test when we're masking against mask arithmetic, not the setcc's directly. 119define <4 x float> @foo6(<4 x i32> %a0, <4 x i32> %a1) { 120; CHECK-LABEL: LCPI6_0: 121; CHECK-NEXT: .long 1065353216 ## 0x3f800000 122; CHECK-NEXT: .long 0 ## 0x0 123; CHECK-NEXT: .long 1065353216 ## 0x3f800000 124; CHECK-NEXT: .long 0 ## 0x0 125; CHECK: ## %bb.0: 126; CHECK-NEXT: movdqa %xmm0, %xmm2 127; CHECK-NEXT: pcmpgtd %xmm1, %xmm2 128; CHECK-NEXT: pxor %xmm1, %xmm1 129; CHECK-NEXT: pcmpgtd %xmm1, %xmm0 130; CHECK-NEXT: pand %xmm2, %xmm0 131; CHECK-NEXT: pand {{.*}}(%rip), %xmm0 132; CHECK-NEXT: retq 133 %1 = icmp sgt <4 x i32> %a0, %a1 134 %2 = icmp sgt <4 x i32> %a0, zeroinitializer 135 %3 = and <4 x i1> %1, %2 136 %4 = sext <4 x i1> %3 to <4 x i32> 137 %5 = and <4 x i32> %4, <i32 1, i32 0, i32 1, i32 0> 138 %6 = uitofp <4 x i32> %5 to <4 x float> 139 ret <4 x float> %6 140} 141 142define <4 x float> @foo7(<4 x i64> %a) { 143; CHECK-LABEL: LCPI7_0: 144; CHECK-NEXT: .byte 0 ## 0x0 145; CHECK-NEXT: .byte 255 ## 0xff 146; CHECK-NEXT: .byte 0 ## 0x0 147; CHECK-NEXT: .byte 0 ## 0x0 148; CHECK-NEXT: .byte 0 ## 0x0 149; CHECK-NEXT: .byte 255 ## 0xff 150; CHECK-NEXT: .byte 0 ## 0x0 151; CHECK-NEXT: .byte 0 ## 0x0 152; CHECK-NEXT: .byte 0 ## 0x0 153; CHECK-NEXT: .byte 255 ## 0xff 154; CHECK-NEXT: .byte 0 ## 0x0 155; CHECK-NEXT: .byte 0 ## 0x0 156; CHECK-NEXT: .byte 0 ## 0x0 157; CHECK-NEXT: .byte 255 ## 0xff 158; CHECK-NEXT: .byte 0 ## 0x0 159; CHECK-NEXT: .byte 0 ## 0x0 160; CHECK-LABEL: foo7: 161; CHECK: ## %bb.0: 162; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[0,2] 163; CHECK-NEXT: andps {{.*}}(%rip), %xmm0 164; CHECK-NEXT: cvtdq2ps %xmm0, %xmm0 165; CHECK-NEXT: retq 166 %b = and <4 x i64> %a, <i64 4278255360, i64 4278255360, i64 4278255360, i64 4278255360> 167 %c = and <4 x i64> %b, <i64 65535, i64 65535, i64 65535, i64 65535> 168 %d = uitofp <4 x i64> %c to <4 x float> 169 ret <4 x float> %d 170} 171