90 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc -mtriple=x86_64-linux -mcpu=skylake -mattr=+avx2 < %s | FileCheck %s3 4define <2 x float> @gather_v2f32_scale_512(ptr %result, <2 x i64> %idx, <2 x i1> %mask) {5; CHECK-LABEL: gather_v2f32_scale_512:6; CHECK: # %bb.0:7; CHECK-NEXT: vpsllq $9, %xmm0, %xmm28; CHECK-NEXT: vpshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]9; CHECK-NEXT: vpslld $31, %xmm0, %xmm110; CHECK-NEXT: vpxor %xmm0, %xmm0, %xmm011; CHECK-NEXT: vgatherqps %xmm1, (%rdi,%xmm2), %xmm012; CHECK-NEXT: retq13 %gep = getelementptr inbounds [512 x i8], ptr %result, <2 x i64> %idx14 %res = call <2 x float> @llvm.masked.gather.v2f32.v2p0(<2 x ptr> %gep, i32 0, <2 x i1> %mask, <2 x float> zeroinitializer)15 ret <2 x float> %res16}17 18define <2 x float> @gather_v2f32_scale_16(ptr %result, <2 x i64> %idx, <2 x i1> %mask) {19; CHECK-LABEL: gather_v2f32_scale_16:20; CHECK: # %bb.0:21; CHECK-NEXT: vpsllq $4, %xmm0, %xmm222; CHECK-NEXT: vpshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]23; CHECK-NEXT: vpslld $31, %xmm0, %xmm124; CHECK-NEXT: vpxor %xmm0, %xmm0, %xmm025; CHECK-NEXT: vgatherqps %xmm1, (%rdi,%xmm2), %xmm026; CHECK-NEXT: retq27 %gep = getelementptr inbounds [16 x i8], ptr %result, <2 x i64> %idx28 %res = call <2 x float> @llvm.masked.gather.v2f32.v2p0(<2 x ptr> %gep, i32 0, <2 x i1> %mask, <2 x float> zeroinitializer)29 ret <2 x float> %res30}31 32define <2 x float> @gather_v2f32_scale_8(ptr %result, <2 x i64> %idx, <2 x i1> %mask) {33; CHECK-LABEL: gather_v2f32_scale_8:34; CHECK: # %bb.0:35; CHECK-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]36; CHECK-NEXT: vpslld $31, %xmm1, %xmm237; CHECK-NEXT: vpxor %xmm1, %xmm1, %xmm138; CHECK-NEXT: vgatherqps %xmm2, (%rdi,%xmm0,8), %xmm139; CHECK-NEXT: vmovaps %xmm1, %xmm040; CHECK-NEXT: retq41 %gep = getelementptr inbounds [8 x i8], ptr %result, <2 x i64> %idx42 %res = call <2 x float> @llvm.masked.gather.v2f32.v2p0(<2 x ptr> %gep, i32 0, <2 x i1> %mask, <2 x float> zeroinitializer)43 ret <2 x float> %res44}45 46define <2 x float> @gather_v2f32_scale_4(ptr %result, <2 x i64> %idx, <2 x i1> %mask) {47; CHECK-LABEL: gather_v2f32_scale_4:48; CHECK: # %bb.0:49; CHECK-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]50; CHECK-NEXT: vpslld $31, %xmm1, %xmm251; CHECK-NEXT: vpxor %xmm1, %xmm1, %xmm152; CHECK-NEXT: vgatherqps %xmm2, (%rdi,%xmm0,4), %xmm153; CHECK-NEXT: vmovaps %xmm1, %xmm054; CHECK-NEXT: retq55 %gep = getelementptr inbounds [4 x i8], ptr %result, <2 x i64> %idx56 %res = call <2 x float> @llvm.masked.gather.v2f32.v2p0(<2 x ptr> %gep, i32 0, <2 x i1> %mask, <2 x float> zeroinitializer)57 ret <2 x float> %res58}59 60define <2 x float> @gather_v2f32_scale_3(ptr %result, <2 x i64> %idx, <2 x i1> %mask) {61; CHECK-LABEL: gather_v2f32_scale_3:62; CHECK: # %bb.0:63; CHECK-NEXT: vpaddq %xmm0, %xmm0, %xmm264; CHECK-NEXT: vpaddq %xmm0, %xmm2, %xmm265; CHECK-NEXT: vpshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]66; CHECK-NEXT: vpslld $31, %xmm0, %xmm167; CHECK-NEXT: vpxor %xmm0, %xmm0, %xmm068; CHECK-NEXT: vgatherqps %xmm1, (%rdi,%xmm2), %xmm069; CHECK-NEXT: retq70 %gep = getelementptr inbounds [3 x i8], ptr %result, <2 x i64> %idx71 %res = call <2 x float> @llvm.masked.gather.v2f32.v2p0(<2 x ptr> %gep, i32 0, <2 x i1> %mask, <2 x float> zeroinitializer)72 ret <2 x float> %res73}74 75define <2 x float> @gather_v2f32_scale_1(ptr %result, <2 x i64> %idx, <2 x i1> %mask) {76; CHECK-LABEL: gather_v2f32_scale_1:77; CHECK: # %bb.0:78; CHECK-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]79; CHECK-NEXT: vpslld $31, %xmm1, %xmm280; CHECK-NEXT: vpxor %xmm1, %xmm1, %xmm181; CHECK-NEXT: vgatherqps %xmm2, (%rdi,%xmm0), %xmm182; CHECK-NEXT: vmovaps %xmm1, %xmm083; CHECK-NEXT: retq84 %gep = getelementptr inbounds [1 x i8], ptr %result, <2 x i64> %idx85 %res = call <2 x float> @llvm.masked.gather.v2f32.v2p0(<2 x ptr> %gep, i32 0, <2 x i1> %mask, <2 x float> zeroinitializer)86 ret <2 x float> %res87}88 89declare <2 x float> @llvm.masked.gather.v2f32.v2p0(<2 x ptr>, i32 immarg, <2 x i1>, <2 x float>)90