From 7c2fdf27795a3f6576395f0a8007e84d9f52e48b Mon Sep 17 00:00:00 2001 From: Craig Topper Date: Tue, 10 Sep 2019 05:49:48 +0000 Subject: [PATCH] [X86] Add broadcast load unfold tests for VCMPPS/PD. llvm-svn: 371486 --- llvm/test/CodeGen/X86/avx512-broadcast-unfold.ll | 221 +++++++++++++++++++++++ 1 file changed, 221 insertions(+) diff --git a/llvm/test/CodeGen/X86/avx512-broadcast-unfold.ll b/llvm/test/CodeGen/X86/avx512-broadcast-unfold.ll index 86e9cfa..a798102 100644 --- a/llvm/test/CodeGen/X86/avx512-broadcast-unfold.ll +++ b/llvm/test/CodeGen/X86/avx512-broadcast-unfold.ll @@ -4222,3 +4222,224 @@ bb1: ; preds = %bb1, %bb bb10: ; preds = %bb1 ret void } + +define void @bcast_unfold_cmp_v4f32(float* %arg) { +; CHECK-LABEL: bcast_unfold_cmp_v4f32: +; CHECK: # %bb.0: # %bb +; CHECK-NEXT: movq $-4096, %rax # imm = 0xF000 +; CHECK-NEXT: vbroadcastss {{.*#+}} xmm0 = [3.0E+0,3.0E+0,3.0E+0,3.0E+0] +; CHECK-NEXT: .p2align 4, 0x90 +; CHECK-NEXT: .LBB120_1: # %bb1 +; CHECK-NEXT: # =>This Inner Loop Header: Depth=1 +; CHECK-NEXT: vmovups 4096(%rdi,%rax), %xmm1 +; CHECK-NEXT: vcmpltps {{.*}}(%rip){1to4}, %xmm1, %k1 +; CHECK-NEXT: vblendmps %xmm1, %xmm0, %xmm1 {%k1} +; CHECK-NEXT: vmovups %xmm1, 4096(%rdi,%rax) +; CHECK-NEXT: addq $16, %rax +; CHECK-NEXT: jne .LBB120_1 +; CHECK-NEXT: # %bb.2: # %bb10 +; CHECK-NEXT: retq +bb: + br label %bb1 + +bb1: ; preds = %bb1, %bb + %tmp = phi i64 [ 0, %bb ], [ %tmp8, %bb1 ] + %tmp2 = getelementptr inbounds float, float* %arg, i64 %tmp + %tmp3 = bitcast float* %tmp2 to <4 x float>* + %tmp4 = load <4 x float>, <4 x float>* %tmp3, align 4 + %tmp5 = fcmp olt <4 x float> %tmp4, + %tmp6 = select <4 x i1> %tmp5, <4 x float> %tmp4, <4 x float> + %tmp7 = bitcast float* %tmp2 to <4 x float>* + store <4 x float> %tmp6, <4 x float>* %tmp7, align 4 + %tmp8 = add i64 %tmp, 4 + %tmp9 = icmp eq i64 %tmp8, 1024 + br i1 %tmp9, label %bb10, label %bb1 + +bb10: ; preds = %bb1 + ret void +} + +define void @bcast_unfold_cmp_v8f32(float* %arg) { +; CHECK-LABEL: bcast_unfold_cmp_v8f32: +; CHECK: # %bb.0: # %bb +; CHECK-NEXT: movq $-4096, %rax # imm = 0xF000 +; CHECK-NEXT: vbroadcastss {{.*#+}} ymm0 = [3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0] +; CHECK-NEXT: .p2align 4, 0x90 +; CHECK-NEXT: .LBB121_1: # %bb1 +; CHECK-NEXT: # =>This Inner Loop Header: Depth=1 +; CHECK-NEXT: vmovups 4096(%rdi,%rax), %ymm1 +; CHECK-NEXT: vcmpltps {{.*}}(%rip){1to8}, %ymm1, %k1 +; CHECK-NEXT: vblendmps %ymm1, %ymm0, %ymm1 {%k1} +; CHECK-NEXT: vmovups %ymm1, 4096(%rdi,%rax) +; CHECK-NEXT: addq $32, %rax +; CHECK-NEXT: jne .LBB121_1 +; CHECK-NEXT: # %bb.2: # %bb10 +; CHECK-NEXT: vzeroupper +; CHECK-NEXT: retq +bb: + br label %bb1 + +bb1: ; preds = %bb1, %bb + %tmp = phi i64 [ 0, %bb ], [ %tmp8, %bb1 ] + %tmp2 = getelementptr inbounds float, float* %arg, i64 %tmp + %tmp3 = bitcast float* %tmp2 to <8 x float>* + %tmp4 = load <8 x float>, <8 x float>* %tmp3, align 4 + %tmp5 = fcmp olt <8 x float> %tmp4, + %tmp6 = select <8 x i1> %tmp5, <8 x float> %tmp4, <8 x float> + %tmp7 = bitcast float* %tmp2 to <8 x float>* + store <8 x float> %tmp6, <8 x float>* %tmp7, align 4 + %tmp8 = add i64 %tmp, 8 + %tmp9 = icmp eq i64 %tmp8, 1024 + br i1 %tmp9, label %bb10, label %bb1 + +bb10: ; preds = %bb1 + ret void +} + +define void @bcast_unfold_cmp_v16f32(float* %arg) { +; CHECK-LABEL: bcast_unfold_cmp_v16f32: +; CHECK: # %bb.0: # %bb +; CHECK-NEXT: movq $-4096, %rax # imm = 0xF000 +; CHECK-NEXT: vbroadcastss {{.*#+}} zmm0 = [3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0] +; CHECK-NEXT: .p2align 4, 0x90 +; CHECK-NEXT: .LBB122_1: # %bb1 +; CHECK-NEXT: # =>This Inner Loop Header: Depth=1 +; CHECK-NEXT: vmovups 4096(%rdi,%rax), %zmm1 +; CHECK-NEXT: vcmpltps {{.*}}(%rip){1to16}, %zmm1, %k1 +; CHECK-NEXT: vblendmps %zmm1, %zmm0, %zmm1 {%k1} +; CHECK-NEXT: vmovups %zmm1, 4096(%rdi,%rax) +; CHECK-NEXT: addq $64, %rax +; CHECK-NEXT: jne .LBB122_1 +; CHECK-NEXT: # %bb.2: # %bb10 +; CHECK-NEXT: vzeroupper +; CHECK-NEXT: retq +bb: + br label %bb1 + +bb1: ; preds = %bb1, %bb + %tmp = phi i64 [ 0, %bb ], [ %tmp8, %bb1 ] + %tmp2 = getelementptr inbounds float, float* %arg, i64 %tmp + %tmp3 = bitcast float* %tmp2 to <16 x float>* + %tmp4 = load <16 x float>, <16 x float>* %tmp3, align 4 + %tmp5 = fcmp olt <16 x float> %tmp4, + %tmp6 = select <16 x i1> %tmp5, <16 x float> %tmp4, <16 x float> + %tmp7 = bitcast float* %tmp2 to <16 x float>* + store <16 x float> %tmp6, <16 x float>* %tmp7, align 4 + %tmp8 = add i64 %tmp, 16 + %tmp9 = icmp eq i64 %tmp8, 1024 + br i1 %tmp9, label %bb10, label %bb1 + +bb10: ; preds = %bb1 + ret void +} + +define void @bcast_unfold_cmp_v2f64(double* %arg) { +; CHECK-LABEL: bcast_unfold_cmp_v2f64: +; CHECK: # %bb.0: # %bb +; CHECK-NEXT: movq $-8192, %rax # imm = 0xE000 +; CHECK-NEXT: vmovapd {{.*#+}} xmm0 = [2.0E+0,2.0E+0] +; CHECK-NEXT: vmovapd {{.*#+}} xmm1 = [3.0E+0,3.0E+0] +; CHECK-NEXT: .p2align 4, 0x90 +; CHECK-NEXT: .LBB123_1: # %bb1 +; CHECK-NEXT: # =>This Inner Loop Header: Depth=1 +; CHECK-NEXT: vmovupd 8192(%rdi,%rax), %xmm2 +; CHECK-NEXT: vcmpltpd %xmm0, %xmm2, %k1 +; CHECK-NEXT: vblendmpd %xmm2, %xmm1, %xmm2 {%k1} +; CHECK-NEXT: vmovupd %xmm2, 8192(%rdi,%rax) +; CHECK-NEXT: addq $16, %rax +; CHECK-NEXT: jne .LBB123_1 +; CHECK-NEXT: # %bb.2: # %bb10 +; CHECK-NEXT: retq +bb: + br label %bb1 + +bb1: ; preds = %bb1, %bb + %tmp = phi i64 [ 0, %bb ], [ %tmp8, %bb1 ] + %tmp2 = getelementptr inbounds double, double* %arg, i64 %tmp + %tmp3 = bitcast double* %tmp2 to <2 x double>* + %tmp4 = load <2 x double>, <2 x double>* %tmp3, align 8 + %tmp5 = fcmp olt <2 x double> %tmp4, + %tmp6 = select <2 x i1> %tmp5, <2 x double> %tmp4, <2 x double> + %tmp7 = bitcast double* %tmp2 to <2 x double>* + store <2 x double> %tmp6, <2 x double>* %tmp7, align 8 + %tmp8 = add i64 %tmp, 2 + %tmp9 = icmp eq i64 %tmp8, 1024 + br i1 %tmp9, label %bb10, label %bb1 + +bb10: ; preds = %bb1 + ret void +} + +define void @bcast_unfold_cmp_v4f64(double* %arg) { +; CHECK-LABEL: bcast_unfold_cmp_v4f64: +; CHECK: # %bb.0: # %bb +; CHECK-NEXT: movq $-8192, %rax # imm = 0xE000 +; CHECK-NEXT: vbroadcastsd {{.*#+}} ymm0 = [3.0E+0,3.0E+0,3.0E+0,3.0E+0] +; CHECK-NEXT: .p2align 4, 0x90 +; CHECK-NEXT: .LBB124_1: # %bb1 +; CHECK-NEXT: # =>This Inner Loop Header: Depth=1 +; CHECK-NEXT: vmovupd 8192(%rdi,%rax), %ymm1 +; CHECK-NEXT: vcmpltpd {{.*}}(%rip){1to4}, %ymm1, %k1 +; CHECK-NEXT: vblendmpd %ymm1, %ymm0, %ymm1 {%k1} +; CHECK-NEXT: vmovupd %ymm1, 8192(%rdi,%rax) +; CHECK-NEXT: addq $32, %rax +; CHECK-NEXT: jne .LBB124_1 +; CHECK-NEXT: # %bb.2: # %bb10 +; CHECK-NEXT: vzeroupper +; CHECK-NEXT: retq +bb: + br label %bb1 + +bb1: ; preds = %bb1, %bb + %tmp = phi i64 [ 0, %bb ], [ %tmp8, %bb1 ] + %tmp2 = getelementptr inbounds double, double* %arg, i64 %tmp + %tmp3 = bitcast double* %tmp2 to <4 x double>* + %tmp4 = load <4 x double>, <4 x double>* %tmp3, align 8 + %tmp5 = fcmp olt <4 x double> %tmp4, + %tmp6 = select <4 x i1> %tmp5, <4 x double> %tmp4, <4 x double> + %tmp7 = bitcast double* %tmp2 to <4 x double>* + store <4 x double> %tmp6, <4 x double>* %tmp7, align 8 + %tmp8 = add i64 %tmp, 4 + %tmp9 = icmp eq i64 %tmp8, 1024 + br i1 %tmp9, label %bb10, label %bb1 + +bb10: ; preds = %bb1 + ret void +} + +define void @bcast_unfold_cmp_v8f64(double* %arg) { +; CHECK-LABEL: bcast_unfold_cmp_v8f64: +; CHECK: # %bb.0: # %bb +; CHECK-NEXT: movq $-8192, %rax # imm = 0xE000 +; CHECK-NEXT: vbroadcastsd {{.*#+}} zmm0 = [3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0] +; CHECK-NEXT: .p2align 4, 0x90 +; CHECK-NEXT: .LBB125_1: # %bb1 +; CHECK-NEXT: # =>This Inner Loop Header: Depth=1 +; CHECK-NEXT: vmovupd 8192(%rdi,%rax), %zmm1 +; CHECK-NEXT: vcmpltpd {{.*}}(%rip){1to8}, %zmm1, %k1 +; CHECK-NEXT: vblendmpd %zmm1, %zmm0, %zmm1 {%k1} +; CHECK-NEXT: vmovupd %zmm1, 8192(%rdi,%rax) +; CHECK-NEXT: addq $64, %rax +; CHECK-NEXT: jne .LBB125_1 +; CHECK-NEXT: # %bb.2: # %bb10 +; CHECK-NEXT: vzeroupper +; CHECK-NEXT: retq +bb: + br label %bb1 + +bb1: ; preds = %bb1, %bb + %tmp = phi i64 [ 0, %bb ], [ %tmp8, %bb1 ] + %tmp2 = getelementptr inbounds double, double* %arg, i64 %tmp + %tmp3 = bitcast double* %tmp2 to <8 x double>* + %tmp4 = load <8 x double>, <8 x double>* %tmp3, align 8 + %tmp5 = fcmp olt <8 x double> %tmp4, + %tmp6 = select <8 x i1> %tmp5, <8 x double> %tmp4, <8 x double> + %tmp7 = bitcast double* %tmp2 to <8 x double>* + store <8 x double> %tmp6, <8 x double>* %tmp7, align 8 + %tmp8 = add i64 %tmp, 8 + %tmp9 = icmp eq i64 %tmp8, 1024 + br i1 %tmp9, label %bb10, label %bb1 + +bb10: ; preds = %bb1 + ret void +} -- 2.7.4