; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py ; RUN: llc < %s -mtriple=i686-apple-darwin -mattr=+avx | FileCheck %s --check-prefix=X32 --check-prefix=X32-AVX --check-prefix=X32-AVX1 ; RUN: llc < %s -mtriple=i686-apple-darwin -mattr=+avx2 | FileCheck %s --check-prefix=X32 --check-prefix=X32-AVX --check-prefix=X32-AVX2 ; RUN: llc < %s -mtriple=i686-apple-darwin -mattr=+avx512vl | FileCheck %s --check-prefix=X32 --check-prefix=X32-AVX512 --check-prefix=X32-AVX512F ; RUN: llc < %s -mtriple=i686-apple-darwin -mattr=+avx512bw,+avx512vl | FileCheck %s --check-prefix=X32 --check-prefix=X32-AVX512 --check-prefix=X32-AVX512BW ; RUN: llc < %s -mtriple=i686-apple-darwin -mattr=+avx512dq,+avx512vl | FileCheck %s --check-prefix=X32 --check-prefix=X32-AVX512 --check-prefix=X32-AVX512DQ ; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+avx | FileCheck %s --check-prefix=X64 --check-prefix=X64-AVX --check-prefix=X64-AVX1 ; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+avx2 | FileCheck %s --check-prefix=X64 --check-prefix=X64-AVX --check-prefix=X64-AVX2 ; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+avx512vl | FileCheck %s --check-prefix=X64 --check-prefix=X64-AVX512 --check-prefix=X64-AVX512F ; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+avx512bw,+avx512vl | FileCheck %s --check-prefix=X64 --check-prefix=X64-AVX512 --check-prefix=X64-AVX512BW ; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+avx512dq,+avx512vl | FileCheck %s --check-prefix=X64 --check-prefix=X64-AVX512 --check-prefix=X64-AVX512DQ ; ; Subvector Load + Broadcast ; define <4 x double> @test_broadcast_2f64_4f64(<2 x double> *%p) nounwind { ; X32-AVX-LABEL: test_broadcast_2f64_4f64: ; X32-AVX: ## BB#0: ; X32-AVX-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX-NEXT: vbroadcastf128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X32-AVX-NEXT: retl ; ; X32-AVX512F-LABEL: test_broadcast_2f64_4f64: ; X32-AVX512F: ## BB#0: ; X32-AVX512F-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512F-NEXT: vbroadcastf32x4 (%eax), %ymm0 ; X32-AVX512F-NEXT: retl ; ; X32-AVX512BW-LABEL: test_broadcast_2f64_4f64: ; X32-AVX512BW: ## BB#0: ; X32-AVX512BW-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512BW-NEXT: vbroadcastf32x4 (%eax), %ymm0 ; X32-AVX512BW-NEXT: retl ; ; X32-AVX512DQ-LABEL: test_broadcast_2f64_4f64: ; X32-AVX512DQ: ## BB#0: ; X32-AVX512DQ-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512DQ-NEXT: vbroadcastf64x2 (%eax), %ymm0 ; X32-AVX512DQ-NEXT: retl ; ; X64-AVX-LABEL: test_broadcast_2f64_4f64: ; X64-AVX: ## BB#0: ; X64-AVX-NEXT: vbroadcastf128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X64-AVX-NEXT: retq ; ; X64-AVX512F-LABEL: test_broadcast_2f64_4f64: ; X64-AVX512F: ## BB#0: ; X64-AVX512F-NEXT: vbroadcastf32x4 (%rdi), %ymm0 ; X64-AVX512F-NEXT: retq ; ; X64-AVX512BW-LABEL: test_broadcast_2f64_4f64: ; X64-AVX512BW: ## BB#0: ; X64-AVX512BW-NEXT: vbroadcastf32x4 (%rdi), %ymm0 ; X64-AVX512BW-NEXT: retq ; ; X64-AVX512DQ-LABEL: test_broadcast_2f64_4f64: ; X64-AVX512DQ: ## BB#0: ; X64-AVX512DQ-NEXT: vbroadcastf64x2 (%rdi), %ymm0 ; X64-AVX512DQ-NEXT: retq %1 = load <2 x double>, <2 x double> *%p %2 = shufflevector <2 x double> %1, <2 x double> undef, <4 x i32> ret <4 x double> %2 } define <8 x double> @test_broadcast_2f64_8f64(<2 x double> *%p) nounwind { ; X32-AVX-LABEL: test_broadcast_2f64_8f64: ; X32-AVX: ## BB#0: ; X32-AVX-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX-NEXT: vbroadcastf128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X32-AVX-NEXT: vmovdqa %ymm0, %ymm1 ; X32-AVX-NEXT: retl ; ; X32-AVX512F-LABEL: test_broadcast_2f64_8f64: ; X32-AVX512F: ## BB#0: ; X32-AVX512F-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512F-NEXT: vbroadcastf32x4 (%eax), %ymm0 ; X32-AVX512F-NEXT: vinsertf64x4 $1, %ymm0, %zmm0, %zmm0 ; X32-AVX512F-NEXT: retl ; ; X32-AVX512BW-LABEL: test_broadcast_2f64_8f64: ; X32-AVX512BW: ## BB#0: ; X32-AVX512BW-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512BW-NEXT: vbroadcastf32x4 (%eax), %ymm0 ; X32-AVX512BW-NEXT: vinsertf64x4 $1, %ymm0, %zmm0, %zmm0 ; X32-AVX512BW-NEXT: retl ; ; X32-AVX512DQ-LABEL: test_broadcast_2f64_8f64: ; X32-AVX512DQ: ## BB#0: ; X32-AVX512DQ-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512DQ-NEXT: vbroadcastf64x2 (%eax), %ymm0 ; X32-AVX512DQ-NEXT: vinsertf64x4 $1, %ymm0, %zmm0, %zmm0 ; X32-AVX512DQ-NEXT: retl ; ; X64-AVX-LABEL: test_broadcast_2f64_8f64: ; X64-AVX: ## BB#0: ; X64-AVX-NEXT: vbroadcastf128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X64-AVX-NEXT: vmovdqa %ymm0, %ymm1 ; X64-AVX-NEXT: retq ; ; X64-AVX512F-LABEL: test_broadcast_2f64_8f64: ; X64-AVX512F: ## BB#0: ; X64-AVX512F-NEXT: vbroadcastf32x4 (%rdi), %ymm0 ; X64-AVX512F-NEXT: vinsertf64x4 $1, %ymm0, %zmm0, %zmm0 ; X64-AVX512F-NEXT: retq ; ; X64-AVX512BW-LABEL: test_broadcast_2f64_8f64: ; X64-AVX512BW: ## BB#0: ; X64-AVX512BW-NEXT: vbroadcastf32x4 (%rdi), %ymm0 ; X64-AVX512BW-NEXT: vinsertf64x4 $1, %ymm0, %zmm0, %zmm0 ; X64-AVX512BW-NEXT: retq ; ; X64-AVX512DQ-LABEL: test_broadcast_2f64_8f64: ; X64-AVX512DQ: ## BB#0: ; X64-AVX512DQ-NEXT: vbroadcastf64x2 (%rdi), %ymm0 ; X64-AVX512DQ-NEXT: vinsertf64x4 $1, %ymm0, %zmm0, %zmm0 ; X64-AVX512DQ-NEXT: retq %1 = load <2 x double>, <2 x double> *%p %2 = shufflevector <2 x double> %1, <2 x double> undef, <8 x i32> ret <8 x double> %2 } define <8 x double> @test_broadcast_4f64_8f64(<4 x double> *%p) nounwind { ; X32-AVX-LABEL: test_broadcast_4f64_8f64: ; X32-AVX: ## BB#0: ; X32-AVX-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX-NEXT: vmovaps (%eax), %ymm0 ; X32-AVX-NEXT: vmovaps %ymm0, %ymm1 ; X32-AVX-NEXT: retl ; ; X32-AVX512-LABEL: test_broadcast_4f64_8f64: ; X32-AVX512: ## BB#0: ; X32-AVX512-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512-NEXT: vmovapd (%eax), %ymm0 ; X32-AVX512-NEXT: vinsertf64x4 $1, %ymm0, %zmm0, %zmm0 ; X32-AVX512-NEXT: retl ; ; X64-AVX-LABEL: test_broadcast_4f64_8f64: ; X64-AVX: ## BB#0: ; X64-AVX-NEXT: vmovaps (%rdi), %ymm0 ; X64-AVX-NEXT: vmovaps %ymm0, %ymm1 ; X64-AVX-NEXT: retq ; ; X64-AVX512-LABEL: test_broadcast_4f64_8f64: ; X64-AVX512: ## BB#0: ; X64-AVX512-NEXT: vmovapd (%rdi), %ymm0 ; X64-AVX512-NEXT: vinsertf64x4 $1, %ymm0, %zmm0, %zmm0 ; X64-AVX512-NEXT: retq %1 = load <4 x double>, <4 x double> *%p %2 = shufflevector <4 x double> %1, <4 x double> undef, <8 x i32> ret <8 x double> %2 } define <4 x i64> @test_broadcast_2i64_4i64(<2 x i64> *%p) nounwind { ; X32-AVX1-LABEL: test_broadcast_2i64_4i64: ; X32-AVX1: ## BB#0: ; X32-AVX1-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX1-NEXT: vbroadcastf128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X32-AVX1-NEXT: retl ; ; X32-AVX2-LABEL: test_broadcast_2i64_4i64: ; X32-AVX2: ## BB#0: ; X32-AVX2-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X32-AVX2-NEXT: retl ; ; X32-AVX512F-LABEL: test_broadcast_2i64_4i64: ; X32-AVX512F: ## BB#0: ; X32-AVX512F-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512F-NEXT: vbroadcasti32x4 (%eax), %ymm0 ; X32-AVX512F-NEXT: retl ; ; X32-AVX512BW-LABEL: test_broadcast_2i64_4i64: ; X32-AVX512BW: ## BB#0: ; X32-AVX512BW-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512BW-NEXT: vbroadcasti32x4 (%eax), %ymm0 ; X32-AVX512BW-NEXT: retl ; ; X32-AVX512DQ-LABEL: test_broadcast_2i64_4i64: ; X32-AVX512DQ: ## BB#0: ; X32-AVX512DQ-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512DQ-NEXT: vbroadcasti64x2 (%eax), %ymm0 ; X32-AVX512DQ-NEXT: retl ; ; X64-AVX1-LABEL: test_broadcast_2i64_4i64: ; X64-AVX1: ## BB#0: ; X64-AVX1-NEXT: vbroadcastf128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X64-AVX1-NEXT: retq ; ; X64-AVX2-LABEL: test_broadcast_2i64_4i64: ; X64-AVX2: ## BB#0: ; X64-AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X64-AVX2-NEXT: retq ; ; X64-AVX512F-LABEL: test_broadcast_2i64_4i64: ; X64-AVX512F: ## BB#0: ; X64-AVX512F-NEXT: vbroadcasti32x4 (%rdi), %ymm0 ; X64-AVX512F-NEXT: retq ; ; X64-AVX512BW-LABEL: test_broadcast_2i64_4i64: ; X64-AVX512BW: ## BB#0: ; X64-AVX512BW-NEXT: vbroadcasti32x4 (%rdi), %ymm0 ; X64-AVX512BW-NEXT: retq ; ; X64-AVX512DQ-LABEL: test_broadcast_2i64_4i64: ; X64-AVX512DQ: ## BB#0: ; X64-AVX512DQ-NEXT: vbroadcasti64x2 (%rdi), %ymm0 ; X64-AVX512DQ-NEXT: retq %1 = load <2 x i64>, <2 x i64> *%p %2 = shufflevector <2 x i64> %1, <2 x i64> undef, <4 x i32> ret <4 x i64> %2 } define <8 x i64> @test_broadcast_2i64_8i64(<2 x i64> *%p) nounwind { ; X32-AVX1-LABEL: test_broadcast_2i64_8i64: ; X32-AVX1: ## BB#0: ; X32-AVX1-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX1-NEXT: vbroadcastf128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X32-AVX1-NEXT: vmovdqa %ymm0, %ymm1 ; X32-AVX1-NEXT: retl ; ; X32-AVX2-LABEL: test_broadcast_2i64_8i64: ; X32-AVX2: ## BB#0: ; X32-AVX2-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X32-AVX2-NEXT: vmovdqa %ymm0, %ymm1 ; X32-AVX2-NEXT: retl ; ; X32-AVX512F-LABEL: test_broadcast_2i64_8i64: ; X32-AVX512F: ## BB#0: ; X32-AVX512F-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512F-NEXT: vbroadcasti32x4 (%eax), %ymm0 ; X32-AVX512F-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0 ; X32-AVX512F-NEXT: retl ; ; X32-AVX512BW-LABEL: test_broadcast_2i64_8i64: ; X32-AVX512BW: ## BB#0: ; X32-AVX512BW-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512BW-NEXT: vbroadcasti32x4 (%eax), %ymm0 ; X32-AVX512BW-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0 ; X32-AVX512BW-NEXT: retl ; ; X32-AVX512DQ-LABEL: test_broadcast_2i64_8i64: ; X32-AVX512DQ: ## BB#0: ; X32-AVX512DQ-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512DQ-NEXT: vbroadcasti64x2 (%eax), %ymm0 ; X32-AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0 ; X32-AVX512DQ-NEXT: retl ; ; X64-AVX1-LABEL: test_broadcast_2i64_8i64: ; X64-AVX1: ## BB#0: ; X64-AVX1-NEXT: vbroadcastf128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X64-AVX1-NEXT: vmovdqa %ymm0, %ymm1 ; X64-AVX1-NEXT: retq ; ; X64-AVX2-LABEL: test_broadcast_2i64_8i64: ; X64-AVX2: ## BB#0: ; X64-AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X64-AVX2-NEXT: vmovdqa %ymm0, %ymm1 ; X64-AVX2-NEXT: retq ; ; X64-AVX512F-LABEL: test_broadcast_2i64_8i64: ; X64-AVX512F: ## BB#0: ; X64-AVX512F-NEXT: vbroadcasti32x4 (%rdi), %ymm0 ; X64-AVX512F-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0 ; X64-AVX512F-NEXT: retq ; ; X64-AVX512BW-LABEL: test_broadcast_2i64_8i64: ; X64-AVX512BW: ## BB#0: ; X64-AVX512BW-NEXT: vbroadcasti32x4 (%rdi), %ymm0 ; X64-AVX512BW-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0 ; X64-AVX512BW-NEXT: retq ; ; X64-AVX512DQ-LABEL: test_broadcast_2i64_8i64: ; X64-AVX512DQ: ## BB#0: ; X64-AVX512DQ-NEXT: vbroadcasti64x2 (%rdi), %ymm0 ; X64-AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0 ; X64-AVX512DQ-NEXT: retq %1 = load <2 x i64>, <2 x i64> *%p %2 = shufflevector <2 x i64> %1, <2 x i64> undef, <8 x i32> ret <8 x i64> %2 } define <8 x i64> @test_broadcast_4i64_8i64(<4 x i64> *%p) nounwind { ; X32-AVX-LABEL: test_broadcast_4i64_8i64: ; X32-AVX: ## BB#0: ; X32-AVX-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX-NEXT: vmovaps (%eax), %ymm0 ; X32-AVX-NEXT: vmovaps %ymm0, %ymm1 ; X32-AVX-NEXT: retl ; ; X32-AVX512-LABEL: test_broadcast_4i64_8i64: ; X32-AVX512: ## BB#0: ; X32-AVX512-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512-NEXT: vmovdqa64 (%eax), %ymm0 ; X32-AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0 ; X32-AVX512-NEXT: retl ; ; X64-AVX-LABEL: test_broadcast_4i64_8i64: ; X64-AVX: ## BB#0: ; X64-AVX-NEXT: vmovaps (%rdi), %ymm0 ; X64-AVX-NEXT: vmovaps %ymm0, %ymm1 ; X64-AVX-NEXT: retq ; ; X64-AVX512-LABEL: test_broadcast_4i64_8i64: ; X64-AVX512: ## BB#0: ; X64-AVX512-NEXT: vmovdqa64 (%rdi), %ymm0 ; X64-AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0 ; X64-AVX512-NEXT: retq %1 = load <4 x i64>, <4 x i64> *%p %2 = shufflevector <4 x i64> %1, <4 x i64> undef, <8 x i32> ret <8 x i64> %2 } define <8 x float> @test_broadcast_4f32_8f32(<4 x float> *%p) nounwind { ; X32-AVX-LABEL: test_broadcast_4f32_8f32: ; X32-AVX: ## BB#0: ; X32-AVX-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX-NEXT: vbroadcastf128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X32-AVX-NEXT: retl ; ; X32-AVX512-LABEL: test_broadcast_4f32_8f32: ; X32-AVX512: ## BB#0: ; X32-AVX512-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512-NEXT: vbroadcastf32x4 (%eax), %ymm0 ; X32-AVX512-NEXT: retl ; ; X64-AVX-LABEL: test_broadcast_4f32_8f32: ; X64-AVX: ## BB#0: ; X64-AVX-NEXT: vbroadcastf128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X64-AVX-NEXT: retq ; ; X64-AVX512-LABEL: test_broadcast_4f32_8f32: ; X64-AVX512: ## BB#0: ; X64-AVX512-NEXT: vbroadcastf32x4 (%rdi), %ymm0 ; X64-AVX512-NEXT: retq %1 = load <4 x float>, <4 x float> *%p %2 = shufflevector <4 x float> %1, <4 x float> undef, <8 x i32> ret <8 x float> %2 } define <16 x float> @test_broadcast_4f32_16f32(<4 x float> *%p) nounwind { ; X32-AVX-LABEL: test_broadcast_4f32_16f32: ; X32-AVX: ## BB#0: ; X32-AVX-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX-NEXT: vbroadcastf128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X32-AVX-NEXT: vmovdqa %ymm0, %ymm1 ; X32-AVX-NEXT: retl ; ; X32-AVX512F-LABEL: test_broadcast_4f32_16f32: ; X32-AVX512F: ## BB#0: ; X32-AVX512F-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512F-NEXT: vbroadcastf32x4 (%eax), %ymm0 ; X32-AVX512F-NEXT: vinsertf64x4 $1, %ymm0, %zmm0, %zmm0 ; X32-AVX512F-NEXT: retl ; ; X32-AVX512BW-LABEL: test_broadcast_4f32_16f32: ; X32-AVX512BW: ## BB#0: ; X32-AVX512BW-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512BW-NEXT: vbroadcastf32x4 (%eax), %ymm0 ; X32-AVX512BW-NEXT: vinsertf64x4 $1, %ymm0, %zmm0, %zmm0 ; X32-AVX512BW-NEXT: retl ; ; X32-AVX512DQ-LABEL: test_broadcast_4f32_16f32: ; X32-AVX512DQ: ## BB#0: ; X32-AVX512DQ-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512DQ-NEXT: vbroadcastf32x4 (%eax), %ymm0 ; X32-AVX512DQ-NEXT: vinsertf32x8 $1, %ymm0, %zmm0, %zmm0 ; X32-AVX512DQ-NEXT: retl ; ; X64-AVX-LABEL: test_broadcast_4f32_16f32: ; X64-AVX: ## BB#0: ; X64-AVX-NEXT: vbroadcastf128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X64-AVX-NEXT: vmovdqa %ymm0, %ymm1 ; X64-AVX-NEXT: retq ; ; X64-AVX512F-LABEL: test_broadcast_4f32_16f32: ; X64-AVX512F: ## BB#0: ; X64-AVX512F-NEXT: vbroadcastf32x4 (%rdi), %ymm0 ; X64-AVX512F-NEXT: vinsertf64x4 $1, %ymm0, %zmm0, %zmm0 ; X64-AVX512F-NEXT: retq ; ; X64-AVX512BW-LABEL: test_broadcast_4f32_16f32: ; X64-AVX512BW: ## BB#0: ; X64-AVX512BW-NEXT: vbroadcastf32x4 (%rdi), %ymm0 ; X64-AVX512BW-NEXT: vinsertf64x4 $1, %ymm0, %zmm0, %zmm0 ; X64-AVX512BW-NEXT: retq ; ; X64-AVX512DQ-LABEL: test_broadcast_4f32_16f32: ; X64-AVX512DQ: ## BB#0: ; X64-AVX512DQ-NEXT: vbroadcastf32x4 (%rdi), %ymm0 ; X64-AVX512DQ-NEXT: vinsertf32x8 $1, %ymm0, %zmm0, %zmm0 ; X64-AVX512DQ-NEXT: retq %1 = load <4 x float>, <4 x float> *%p %2 = shufflevector <4 x float> %1, <4 x float> undef, <16 x i32> ret <16 x float> %2 } define <16 x float> @test_broadcast_8f32_16f32(<8 x float> *%p) nounwind { ; X32-AVX-LABEL: test_broadcast_8f32_16f32: ; X32-AVX: ## BB#0: ; X32-AVX-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX-NEXT: vmovaps (%eax), %ymm0 ; X32-AVX-NEXT: vmovaps %ymm0, %ymm1 ; X32-AVX-NEXT: retl ; ; X32-AVX512F-LABEL: test_broadcast_8f32_16f32: ; X32-AVX512F: ## BB#0: ; X32-AVX512F-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512F-NEXT: vmovapd (%eax), %ymm0 ; X32-AVX512F-NEXT: vinsertf64x4 $1, %ymm0, %zmm0, %zmm0 ; X32-AVX512F-NEXT: retl ; ; X32-AVX512BW-LABEL: test_broadcast_8f32_16f32: ; X32-AVX512BW: ## BB#0: ; X32-AVX512BW-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512BW-NEXT: vmovapd (%eax), %ymm0 ; X32-AVX512BW-NEXT: vinsertf64x4 $1, %ymm0, %zmm0, %zmm0 ; X32-AVX512BW-NEXT: retl ; ; X32-AVX512DQ-LABEL: test_broadcast_8f32_16f32: ; X32-AVX512DQ: ## BB#0: ; X32-AVX512DQ-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512DQ-NEXT: vmovaps (%eax), %ymm0 ; X32-AVX512DQ-NEXT: vinsertf32x8 $1, %ymm0, %zmm0, %zmm0 ; X32-AVX512DQ-NEXT: retl ; ; X64-AVX-LABEL: test_broadcast_8f32_16f32: ; X64-AVX: ## BB#0: ; X64-AVX-NEXT: vmovaps (%rdi), %ymm0 ; X64-AVX-NEXT: vmovaps %ymm0, %ymm1 ; X64-AVX-NEXT: retq ; ; X64-AVX512F-LABEL: test_broadcast_8f32_16f32: ; X64-AVX512F: ## BB#0: ; X64-AVX512F-NEXT: vmovapd (%rdi), %ymm0 ; X64-AVX512F-NEXT: vinsertf64x4 $1, %ymm0, %zmm0, %zmm0 ; X64-AVX512F-NEXT: retq ; ; X64-AVX512BW-LABEL: test_broadcast_8f32_16f32: ; X64-AVX512BW: ## BB#0: ; X64-AVX512BW-NEXT: vmovapd (%rdi), %ymm0 ; X64-AVX512BW-NEXT: vinsertf64x4 $1, %ymm0, %zmm0, %zmm0 ; X64-AVX512BW-NEXT: retq ; ; X64-AVX512DQ-LABEL: test_broadcast_8f32_16f32: ; X64-AVX512DQ: ## BB#0: ; X64-AVX512DQ-NEXT: vmovaps (%rdi), %ymm0 ; X64-AVX512DQ-NEXT: vinsertf32x8 $1, %ymm0, %zmm0, %zmm0 ; X64-AVX512DQ-NEXT: retq %1 = load <8 x float>, <8 x float> *%p %2 = shufflevector <8 x float> %1, <8 x float> undef, <16 x i32> ret <16 x float> %2 } define <8 x i32> @test_broadcast_4i32_8i32(<4 x i32> *%p) nounwind { ; X32-AVX1-LABEL: test_broadcast_4i32_8i32: ; X32-AVX1: ## BB#0: ; X32-AVX1-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX1-NEXT: vbroadcastf128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X32-AVX1-NEXT: retl ; ; X32-AVX2-LABEL: test_broadcast_4i32_8i32: ; X32-AVX2: ## BB#0: ; X32-AVX2-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X32-AVX2-NEXT: retl ; ; X32-AVX512-LABEL: test_broadcast_4i32_8i32: ; X32-AVX512: ## BB#0: ; X32-AVX512-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512-NEXT: vbroadcasti32x4 (%eax), %ymm0 ; X32-AVX512-NEXT: retl ; ; X64-AVX1-LABEL: test_broadcast_4i32_8i32: ; X64-AVX1: ## BB#0: ; X64-AVX1-NEXT: vbroadcastf128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X64-AVX1-NEXT: retq ; ; X64-AVX2-LABEL: test_broadcast_4i32_8i32: ; X64-AVX2: ## BB#0: ; X64-AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X64-AVX2-NEXT: retq ; ; X64-AVX512-LABEL: test_broadcast_4i32_8i32: ; X64-AVX512: ## BB#0: ; X64-AVX512-NEXT: vbroadcasti32x4 (%rdi), %ymm0 ; X64-AVX512-NEXT: retq %1 = load <4 x i32>, <4 x i32> *%p %2 = shufflevector <4 x i32> %1, <4 x i32> undef, <8 x i32> ret <8 x i32> %2 } define <16 x i32> @test_broadcast_4i32_16i32(<4 x i32> *%p) nounwind { ; X32-AVX1-LABEL: test_broadcast_4i32_16i32: ; X32-AVX1: ## BB#0: ; X32-AVX1-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX1-NEXT: vbroadcastf128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X32-AVX1-NEXT: vmovdqa %ymm0, %ymm1 ; X32-AVX1-NEXT: retl ; ; X32-AVX2-LABEL: test_broadcast_4i32_16i32: ; X32-AVX2: ## BB#0: ; X32-AVX2-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X32-AVX2-NEXT: vmovdqa %ymm0, %ymm1 ; X32-AVX2-NEXT: retl ; ; X32-AVX512F-LABEL: test_broadcast_4i32_16i32: ; X32-AVX512F: ## BB#0: ; X32-AVX512F-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512F-NEXT: vbroadcasti32x4 (%eax), %ymm0 ; X32-AVX512F-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0 ; X32-AVX512F-NEXT: retl ; ; X32-AVX512BW-LABEL: test_broadcast_4i32_16i32: ; X32-AVX512BW: ## BB#0: ; X32-AVX512BW-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512BW-NEXT: vbroadcasti32x4 (%eax), %ymm0 ; X32-AVX512BW-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0 ; X32-AVX512BW-NEXT: retl ; ; X32-AVX512DQ-LABEL: test_broadcast_4i32_16i32: ; X32-AVX512DQ: ## BB#0: ; X32-AVX512DQ-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512DQ-NEXT: vbroadcasti32x4 (%eax), %ymm0 ; X32-AVX512DQ-NEXT: vinserti32x8 $1, %ymm0, %zmm0, %zmm0 ; X32-AVX512DQ-NEXT: retl ; ; X64-AVX1-LABEL: test_broadcast_4i32_16i32: ; X64-AVX1: ## BB#0: ; X64-AVX1-NEXT: vbroadcastf128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X64-AVX1-NEXT: vmovdqa %ymm0, %ymm1 ; X64-AVX1-NEXT: retq ; ; X64-AVX2-LABEL: test_broadcast_4i32_16i32: ; X64-AVX2: ## BB#0: ; X64-AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X64-AVX2-NEXT: vmovdqa %ymm0, %ymm1 ; X64-AVX2-NEXT: retq ; ; X64-AVX512F-LABEL: test_broadcast_4i32_16i32: ; X64-AVX512F: ## BB#0: ; X64-AVX512F-NEXT: vbroadcasti32x4 (%rdi), %ymm0 ; X64-AVX512F-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0 ; X64-AVX512F-NEXT: retq ; ; X64-AVX512BW-LABEL: test_broadcast_4i32_16i32: ; X64-AVX512BW: ## BB#0: ; X64-AVX512BW-NEXT: vbroadcasti32x4 (%rdi), %ymm0 ; X64-AVX512BW-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0 ; X64-AVX512BW-NEXT: retq ; ; X64-AVX512DQ-LABEL: test_broadcast_4i32_16i32: ; X64-AVX512DQ: ## BB#0: ; X64-AVX512DQ-NEXT: vbroadcasti32x4 (%rdi), %ymm0 ; X64-AVX512DQ-NEXT: vinserti32x8 $1, %ymm0, %zmm0, %zmm0 ; X64-AVX512DQ-NEXT: retq %1 = load <4 x i32>, <4 x i32> *%p %2 = shufflevector <4 x i32> %1, <4 x i32> undef, <16 x i32> ret <16 x i32> %2 } define <16 x i32> @test_broadcast_8i32_16i32(<8 x i32> *%p) nounwind { ; X32-AVX-LABEL: test_broadcast_8i32_16i32: ; X32-AVX: ## BB#0: ; X32-AVX-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX-NEXT: vmovaps (%eax), %ymm0 ; X32-AVX-NEXT: vmovaps %ymm0, %ymm1 ; X32-AVX-NEXT: retl ; ; X32-AVX512F-LABEL: test_broadcast_8i32_16i32: ; X32-AVX512F: ## BB#0: ; X32-AVX512F-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512F-NEXT: vmovdqa32 (%eax), %ymm0 ; X32-AVX512F-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0 ; X32-AVX512F-NEXT: retl ; ; X32-AVX512BW-LABEL: test_broadcast_8i32_16i32: ; X32-AVX512BW: ## BB#0: ; X32-AVX512BW-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512BW-NEXT: vmovdqa32 (%eax), %ymm0 ; X32-AVX512BW-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0 ; X32-AVX512BW-NEXT: retl ; ; X32-AVX512DQ-LABEL: test_broadcast_8i32_16i32: ; X32-AVX512DQ: ## BB#0: ; X32-AVX512DQ-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512DQ-NEXT: vmovdqa32 (%eax), %ymm0 ; X32-AVX512DQ-NEXT: vinserti32x8 $1, %ymm0, %zmm0, %zmm0 ; X32-AVX512DQ-NEXT: retl ; ; X64-AVX-LABEL: test_broadcast_8i32_16i32: ; X64-AVX: ## BB#0: ; X64-AVX-NEXT: vmovaps (%rdi), %ymm0 ; X64-AVX-NEXT: vmovaps %ymm0, %ymm1 ; X64-AVX-NEXT: retq ; ; X64-AVX512F-LABEL: test_broadcast_8i32_16i32: ; X64-AVX512F: ## BB#0: ; X64-AVX512F-NEXT: vmovdqa32 (%rdi), %ymm0 ; X64-AVX512F-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0 ; X64-AVX512F-NEXT: retq ; ; X64-AVX512BW-LABEL: test_broadcast_8i32_16i32: ; X64-AVX512BW: ## BB#0: ; X64-AVX512BW-NEXT: vmovdqa32 (%rdi), %ymm0 ; X64-AVX512BW-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0 ; X64-AVX512BW-NEXT: retq ; ; X64-AVX512DQ-LABEL: test_broadcast_8i32_16i32: ; X64-AVX512DQ: ## BB#0: ; X64-AVX512DQ-NEXT: vmovdqa32 (%rdi), %ymm0 ; X64-AVX512DQ-NEXT: vinserti32x8 $1, %ymm0, %zmm0, %zmm0 ; X64-AVX512DQ-NEXT: retq %1 = load <8 x i32>, <8 x i32> *%p %2 = shufflevector <8 x i32> %1, <8 x i32> undef, <16 x i32> ret <16 x i32> %2 } define <16 x i16> @test_broadcast_8i16_16i16(<8 x i16> *%p) nounwind { ; X32-AVX1-LABEL: test_broadcast_8i16_16i16: ; X32-AVX1: ## BB#0: ; X32-AVX1-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX1-NEXT: vbroadcastf128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X32-AVX1-NEXT: retl ; ; X32-AVX2-LABEL: test_broadcast_8i16_16i16: ; X32-AVX2: ## BB#0: ; X32-AVX2-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X32-AVX2-NEXT: retl ; ; X32-AVX512-LABEL: test_broadcast_8i16_16i16: ; X32-AVX512: ## BB#0: ; X32-AVX512-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512-NEXT: vbroadcasti32x4 (%eax), %ymm0 ; X32-AVX512-NEXT: retl ; ; X64-AVX1-LABEL: test_broadcast_8i16_16i16: ; X64-AVX1: ## BB#0: ; X64-AVX1-NEXT: vbroadcastf128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X64-AVX1-NEXT: retq ; ; X64-AVX2-LABEL: test_broadcast_8i16_16i16: ; X64-AVX2: ## BB#0: ; X64-AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X64-AVX2-NEXT: retq ; ; X64-AVX512-LABEL: test_broadcast_8i16_16i16: ; X64-AVX512: ## BB#0: ; X64-AVX512-NEXT: vbroadcasti32x4 (%rdi), %ymm0 ; X64-AVX512-NEXT: retq %1 = load <8 x i16>, <8 x i16> *%p %2 = shufflevector <8 x i16> %1, <8 x i16> undef, <16 x i32> ret <16 x i16> %2 } define <32 x i16> @test_broadcast_8i16_32i16(<8 x i16> *%p) nounwind { ; X32-AVX1-LABEL: test_broadcast_8i16_32i16: ; X32-AVX1: ## BB#0: ; X32-AVX1-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX1-NEXT: vbroadcastf128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X32-AVX1-NEXT: vmovdqa %ymm0, %ymm1 ; X32-AVX1-NEXT: retl ; ; X32-AVX2-LABEL: test_broadcast_8i16_32i16: ; X32-AVX2: ## BB#0: ; X32-AVX2-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X32-AVX2-NEXT: vmovdqa %ymm0, %ymm1 ; X32-AVX2-NEXT: retl ; ; X32-AVX512F-LABEL: test_broadcast_8i16_32i16: ; X32-AVX512F: ## BB#0: ; X32-AVX512F-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512F-NEXT: vbroadcasti32x4 (%eax), %ymm0 ; X32-AVX512F-NEXT: vmovdqa64 %ymm0, %ymm1 ; X32-AVX512F-NEXT: retl ; ; X32-AVX512BW-LABEL: test_broadcast_8i16_32i16: ; X32-AVX512BW: ## BB#0: ; X32-AVX512BW-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512BW-NEXT: vbroadcasti32x4 (%eax), %ymm0 ; X32-AVX512BW-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0 ; X32-AVX512BW-NEXT: retl ; ; X32-AVX512DQ-LABEL: test_broadcast_8i16_32i16: ; X32-AVX512DQ: ## BB#0: ; X32-AVX512DQ-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512DQ-NEXT: vbroadcasti32x4 (%eax), %ymm0 ; X32-AVX512DQ-NEXT: vmovdqa64 %ymm0, %ymm1 ; X32-AVX512DQ-NEXT: retl ; ; X64-AVX1-LABEL: test_broadcast_8i16_32i16: ; X64-AVX1: ## BB#0: ; X64-AVX1-NEXT: vbroadcastf128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X64-AVX1-NEXT: vmovdqa %ymm0, %ymm1 ; X64-AVX1-NEXT: retq ; ; X64-AVX2-LABEL: test_broadcast_8i16_32i16: ; X64-AVX2: ## BB#0: ; X64-AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X64-AVX2-NEXT: vmovdqa %ymm0, %ymm1 ; X64-AVX2-NEXT: retq ; ; X64-AVX512F-LABEL: test_broadcast_8i16_32i16: ; X64-AVX512F: ## BB#0: ; X64-AVX512F-NEXT: vbroadcasti32x4 (%rdi), %ymm0 ; X64-AVX512F-NEXT: vmovdqa64 %ymm0, %ymm1 ; X64-AVX512F-NEXT: retq ; ; X64-AVX512BW-LABEL: test_broadcast_8i16_32i16: ; X64-AVX512BW: ## BB#0: ; X64-AVX512BW-NEXT: vbroadcasti32x4 (%rdi), %ymm0 ; X64-AVX512BW-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0 ; X64-AVX512BW-NEXT: retq ; ; X64-AVX512DQ-LABEL: test_broadcast_8i16_32i16: ; X64-AVX512DQ: ## BB#0: ; X64-AVX512DQ-NEXT: vbroadcasti32x4 (%rdi), %ymm0 ; X64-AVX512DQ-NEXT: vmovdqa64 %ymm0, %ymm1 ; X64-AVX512DQ-NEXT: retq %1 = load <8 x i16>, <8 x i16> *%p %2 = shufflevector <8 x i16> %1, <8 x i16> undef, <32 x i32> ret <32 x i16> %2 } define <32 x i16> @test_broadcast_16i16_32i16(<16 x i16> *%p) nounwind { ; X32-AVX-LABEL: test_broadcast_16i16_32i16: ; X32-AVX: ## BB#0: ; X32-AVX-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX-NEXT: vmovaps (%eax), %ymm0 ; X32-AVX-NEXT: vmovaps %ymm0, %ymm1 ; X32-AVX-NEXT: retl ; ; X32-AVX512F-LABEL: test_broadcast_16i16_32i16: ; X32-AVX512F: ## BB#0: ; X32-AVX512F-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512F-NEXT: vmovaps (%eax), %ymm0 ; X32-AVX512F-NEXT: vmovaps %ymm0, %ymm1 ; X32-AVX512F-NEXT: retl ; ; X32-AVX512BW-LABEL: test_broadcast_16i16_32i16: ; X32-AVX512BW: ## BB#0: ; X32-AVX512BW-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512BW-NEXT: vmovdqu16 (%eax), %ymm0 ; X32-AVX512BW-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0 ; X32-AVX512BW-NEXT: retl ; ; X32-AVX512DQ-LABEL: test_broadcast_16i16_32i16: ; X32-AVX512DQ: ## BB#0: ; X32-AVX512DQ-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512DQ-NEXT: vmovaps (%eax), %ymm0 ; X32-AVX512DQ-NEXT: vmovaps %ymm0, %ymm1 ; X32-AVX512DQ-NEXT: retl ; ; X64-AVX-LABEL: test_broadcast_16i16_32i16: ; X64-AVX: ## BB#0: ; X64-AVX-NEXT: vmovaps (%rdi), %ymm0 ; X64-AVX-NEXT: vmovaps %ymm0, %ymm1 ; X64-AVX-NEXT: retq ; ; X64-AVX512F-LABEL: test_broadcast_16i16_32i16: ; X64-AVX512F: ## BB#0: ; X64-AVX512F-NEXT: vmovaps (%rdi), %ymm0 ; X64-AVX512F-NEXT: vmovaps %ymm0, %ymm1 ; X64-AVX512F-NEXT: retq ; ; X64-AVX512BW-LABEL: test_broadcast_16i16_32i16: ; X64-AVX512BW: ## BB#0: ; X64-AVX512BW-NEXT: vmovdqu16 (%rdi), %ymm0 ; X64-AVX512BW-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0 ; X64-AVX512BW-NEXT: retq ; ; X64-AVX512DQ-LABEL: test_broadcast_16i16_32i16: ; X64-AVX512DQ: ## BB#0: ; X64-AVX512DQ-NEXT: vmovaps (%rdi), %ymm0 ; X64-AVX512DQ-NEXT: vmovaps %ymm0, %ymm1 ; X64-AVX512DQ-NEXT: retq %1 = load <16 x i16>, <16 x i16> *%p %2 = shufflevector <16 x i16> %1, <16 x i16> undef, <32 x i32> ret <32 x i16> %2 } define <32 x i8> @test_broadcast_16i8_32i8(<16 x i8> *%p) nounwind { ; X32-AVX1-LABEL: test_broadcast_16i8_32i8: ; X32-AVX1: ## BB#0: ; X32-AVX1-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX1-NEXT: vbroadcastf128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X32-AVX1-NEXT: retl ; ; X32-AVX2-LABEL: test_broadcast_16i8_32i8: ; X32-AVX2: ## BB#0: ; X32-AVX2-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X32-AVX2-NEXT: retl ; ; X32-AVX512-LABEL: test_broadcast_16i8_32i8: ; X32-AVX512: ## BB#0: ; X32-AVX512-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512-NEXT: vbroadcasti32x4 (%eax), %ymm0 ; X32-AVX512-NEXT: retl ; ; X64-AVX1-LABEL: test_broadcast_16i8_32i8: ; X64-AVX1: ## BB#0: ; X64-AVX1-NEXT: vbroadcastf128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X64-AVX1-NEXT: retq ; ; X64-AVX2-LABEL: test_broadcast_16i8_32i8: ; X64-AVX2: ## BB#0: ; X64-AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X64-AVX2-NEXT: retq ; ; X64-AVX512-LABEL: test_broadcast_16i8_32i8: ; X64-AVX512: ## BB#0: ; X64-AVX512-NEXT: vbroadcasti32x4 (%rdi), %ymm0 ; X64-AVX512-NEXT: retq %1 = load <16 x i8>, <16 x i8> *%p %2 = shufflevector <16 x i8> %1, <16 x i8> undef, <32 x i32> ret <32 x i8> %2 } define <64 x i8> @test_broadcast_16i8_64i8(<16 x i8> *%p) nounwind { ; X32-AVX1-LABEL: test_broadcast_16i8_64i8: ; X32-AVX1: ## BB#0: ; X32-AVX1-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX1-NEXT: vbroadcastf128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X32-AVX1-NEXT: vmovdqa %ymm0, %ymm1 ; X32-AVX1-NEXT: retl ; ; X32-AVX2-LABEL: test_broadcast_16i8_64i8: ; X32-AVX2: ## BB#0: ; X32-AVX2-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X32-AVX2-NEXT: vmovdqa %ymm0, %ymm1 ; X32-AVX2-NEXT: retl ; ; X32-AVX512F-LABEL: test_broadcast_16i8_64i8: ; X32-AVX512F: ## BB#0: ; X32-AVX512F-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512F-NEXT: vbroadcasti32x4 (%eax), %ymm0 ; X32-AVX512F-NEXT: vmovdqa64 %ymm0, %ymm1 ; X32-AVX512F-NEXT: retl ; ; X32-AVX512BW-LABEL: test_broadcast_16i8_64i8: ; X32-AVX512BW: ## BB#0: ; X32-AVX512BW-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512BW-NEXT: vbroadcasti32x4 (%eax), %ymm0 ; X32-AVX512BW-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0 ; X32-AVX512BW-NEXT: retl ; ; X32-AVX512DQ-LABEL: test_broadcast_16i8_64i8: ; X32-AVX512DQ: ## BB#0: ; X32-AVX512DQ-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512DQ-NEXT: vbroadcasti32x4 (%eax), %ymm0 ; X32-AVX512DQ-NEXT: vmovdqa64 %ymm0, %ymm1 ; X32-AVX512DQ-NEXT: retl ; ; X64-AVX1-LABEL: test_broadcast_16i8_64i8: ; X64-AVX1: ## BB#0: ; X64-AVX1-NEXT: vbroadcastf128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X64-AVX1-NEXT: vmovdqa %ymm0, %ymm1 ; X64-AVX1-NEXT: retq ; ; X64-AVX2-LABEL: test_broadcast_16i8_64i8: ; X64-AVX2: ## BB#0: ; X64-AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm0 = mem[0,1,0,1] ; X64-AVX2-NEXT: vmovdqa %ymm0, %ymm1 ; X64-AVX2-NEXT: retq ; ; X64-AVX512F-LABEL: test_broadcast_16i8_64i8: ; X64-AVX512F: ## BB#0: ; X64-AVX512F-NEXT: vbroadcasti32x4 (%rdi), %ymm0 ; X64-AVX512F-NEXT: vmovdqa64 %ymm0, %ymm1 ; X64-AVX512F-NEXT: retq ; ; X64-AVX512BW-LABEL: test_broadcast_16i8_64i8: ; X64-AVX512BW: ## BB#0: ; X64-AVX512BW-NEXT: vbroadcasti32x4 (%rdi), %ymm0 ; X64-AVX512BW-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0 ; X64-AVX512BW-NEXT: retq ; ; X64-AVX512DQ-LABEL: test_broadcast_16i8_64i8: ; X64-AVX512DQ: ## BB#0: ; X64-AVX512DQ-NEXT: vbroadcasti32x4 (%rdi), %ymm0 ; X64-AVX512DQ-NEXT: vmovdqa64 %ymm0, %ymm1 ; X64-AVX512DQ-NEXT: retq %1 = load <16 x i8>, <16 x i8> *%p %2 = shufflevector <16 x i8> %1, <16 x i8> undef, <64 x i32> ret <64 x i8> %2 } define <64 x i8> @test_broadcast_32i8_64i8(<32 x i8> *%p) nounwind { ; X32-AVX-LABEL: test_broadcast_32i8_64i8: ; X32-AVX: ## BB#0: ; X32-AVX-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX-NEXT: vmovaps (%eax), %ymm0 ; X32-AVX-NEXT: vmovaps %ymm0, %ymm1 ; X32-AVX-NEXT: retl ; ; X32-AVX512F-LABEL: test_broadcast_32i8_64i8: ; X32-AVX512F: ## BB#0: ; X32-AVX512F-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512F-NEXT: vmovaps (%eax), %ymm0 ; X32-AVX512F-NEXT: vmovaps %ymm0, %ymm1 ; X32-AVX512F-NEXT: retl ; ; X32-AVX512BW-LABEL: test_broadcast_32i8_64i8: ; X32-AVX512BW: ## BB#0: ; X32-AVX512BW-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512BW-NEXT: vmovdqu8 (%eax), %ymm0 ; X32-AVX512BW-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0 ; X32-AVX512BW-NEXT: retl ; ; X32-AVX512DQ-LABEL: test_broadcast_32i8_64i8: ; X32-AVX512DQ: ## BB#0: ; X32-AVX512DQ-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512DQ-NEXT: vmovaps (%eax), %ymm0 ; X32-AVX512DQ-NEXT: vmovaps %ymm0, %ymm1 ; X32-AVX512DQ-NEXT: retl ; ; X64-AVX-LABEL: test_broadcast_32i8_64i8: ; X64-AVX: ## BB#0: ; X64-AVX-NEXT: vmovaps (%rdi), %ymm0 ; X64-AVX-NEXT: vmovaps %ymm0, %ymm1 ; X64-AVX-NEXT: retq ; ; X64-AVX512F-LABEL: test_broadcast_32i8_64i8: ; X64-AVX512F: ## BB#0: ; X64-AVX512F-NEXT: vmovaps (%rdi), %ymm0 ; X64-AVX512F-NEXT: vmovaps %ymm0, %ymm1 ; X64-AVX512F-NEXT: retq ; ; X64-AVX512BW-LABEL: test_broadcast_32i8_64i8: ; X64-AVX512BW: ## BB#0: ; X64-AVX512BW-NEXT: vmovdqu8 (%rdi), %ymm0 ; X64-AVX512BW-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0 ; X64-AVX512BW-NEXT: retq ; ; X64-AVX512DQ-LABEL: test_broadcast_32i8_64i8: ; X64-AVX512DQ: ## BB#0: ; X64-AVX512DQ-NEXT: vmovaps (%rdi), %ymm0 ; X64-AVX512DQ-NEXT: vmovaps %ymm0, %ymm1 ; X64-AVX512DQ-NEXT: retq %1 = load <32 x i8>, <32 x i8> *%p %2 = shufflevector <32 x i8> %1, <32 x i8> undef, <64 x i32> ret <64 x i8> %2 } ; ; Subvector Load + Broadcast + Store ; define <4 x double> @test_broadcast_2f64_4f64_reuse(<2 x double>* %p0, <2 x double>* %p1) { ; X32-AVX-LABEL: test_broadcast_2f64_4f64_reuse: ; X32-AVX: ## BB#0: ; X32-AVX-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX-NEXT: movl {{[0-9]+}}(%esp), %ecx ; X32-AVX-NEXT: vmovaps (%ecx), %xmm0 ; X32-AVX-NEXT: vmovaps %xmm0, (%eax) ; X32-AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 ; X32-AVX-NEXT: retl ; ; X32-AVX512F-LABEL: test_broadcast_2f64_4f64_reuse: ; X32-AVX512F: ## BB#0: ; X32-AVX512F-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512F-NEXT: movl {{[0-9]+}}(%esp), %ecx ; X32-AVX512F-NEXT: vmovaps (%ecx), %xmm0 ; X32-AVX512F-NEXT: vmovaps %xmm0, (%eax) ; X32-AVX512F-NEXT: vinsertf32x4 $1, %xmm0, %ymm0, %ymm0 ; X32-AVX512F-NEXT: retl ; ; X32-AVX512BW-LABEL: test_broadcast_2f64_4f64_reuse: ; X32-AVX512BW: ## BB#0: ; X32-AVX512BW-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512BW-NEXT: movl {{[0-9]+}}(%esp), %ecx ; X32-AVX512BW-NEXT: vmovaps (%ecx), %xmm0 ; X32-AVX512BW-NEXT: vmovaps %xmm0, (%eax) ; X32-AVX512BW-NEXT: vinsertf32x4 $1, %xmm0, %ymm0, %ymm0 ; X32-AVX512BW-NEXT: retl ; ; X32-AVX512DQ-LABEL: test_broadcast_2f64_4f64_reuse: ; X32-AVX512DQ: ## BB#0: ; X32-AVX512DQ-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512DQ-NEXT: movl {{[0-9]+}}(%esp), %ecx ; X32-AVX512DQ-NEXT: vmovapd (%ecx), %xmm0 ; X32-AVX512DQ-NEXT: vmovapd %xmm0, (%eax) ; X32-AVX512DQ-NEXT: vinsertf64x2 $1, %xmm0, %ymm0, %ymm0 ; X32-AVX512DQ-NEXT: retl ; ; X64-AVX-LABEL: test_broadcast_2f64_4f64_reuse: ; X64-AVX: ## BB#0: ; X64-AVX-NEXT: vmovaps (%rdi), %xmm0 ; X64-AVX-NEXT: vmovaps %xmm0, (%rsi) ; X64-AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 ; X64-AVX-NEXT: retq ; ; X64-AVX512F-LABEL: test_broadcast_2f64_4f64_reuse: ; X64-AVX512F: ## BB#0: ; X64-AVX512F-NEXT: vmovaps (%rdi), %xmm0 ; X64-AVX512F-NEXT: vmovaps %xmm0, (%rsi) ; X64-AVX512F-NEXT: vinsertf32x4 $1, %xmm0, %ymm0, %ymm0 ; X64-AVX512F-NEXT: retq ; ; X64-AVX512BW-LABEL: test_broadcast_2f64_4f64_reuse: ; X64-AVX512BW: ## BB#0: ; X64-AVX512BW-NEXT: vmovaps (%rdi), %xmm0 ; X64-AVX512BW-NEXT: vmovaps %xmm0, (%rsi) ; X64-AVX512BW-NEXT: vinsertf32x4 $1, %xmm0, %ymm0, %ymm0 ; X64-AVX512BW-NEXT: retq ; ; X64-AVX512DQ-LABEL: test_broadcast_2f64_4f64_reuse: ; X64-AVX512DQ: ## BB#0: ; X64-AVX512DQ-NEXT: vmovapd (%rdi), %xmm0 ; X64-AVX512DQ-NEXT: vmovapd %xmm0, (%rsi) ; X64-AVX512DQ-NEXT: vinsertf64x2 $1, %xmm0, %ymm0, %ymm0 ; X64-AVX512DQ-NEXT: retq %1 = load <2 x double>, <2 x double>* %p0 store <2 x double> %1, <2 x double>* %p1 %2 = shufflevector <2 x double> %1, <2 x double> undef, <4 x i32> ret <4 x double> %2 } define <4 x i64> @test_broadcast_2i64_4i64_reuse(<2 x i64>* %p0, <2 x i64>* %p1) { ; X32-AVX-LABEL: test_broadcast_2i64_4i64_reuse: ; X32-AVX: ## BB#0: ; X32-AVX-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX-NEXT: movl {{[0-9]+}}(%esp), %ecx ; X32-AVX-NEXT: vmovaps (%ecx), %xmm0 ; X32-AVX-NEXT: vmovaps %xmm0, (%eax) ; X32-AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 ; X32-AVX-NEXT: retl ; ; X32-AVX512F-LABEL: test_broadcast_2i64_4i64_reuse: ; X32-AVX512F: ## BB#0: ; X32-AVX512F-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512F-NEXT: movl {{[0-9]+}}(%esp), %ecx ; X32-AVX512F-NEXT: vmovdqa64 (%ecx), %xmm0 ; X32-AVX512F-NEXT: vmovdqa64 %xmm0, (%eax) ; X32-AVX512F-NEXT: vinserti32x4 $1, %xmm0, %ymm0, %ymm0 ; X32-AVX512F-NEXT: retl ; ; X32-AVX512BW-LABEL: test_broadcast_2i64_4i64_reuse: ; X32-AVX512BW: ## BB#0: ; X32-AVX512BW-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512BW-NEXT: movl {{[0-9]+}}(%esp), %ecx ; X32-AVX512BW-NEXT: vmovdqa64 (%ecx), %xmm0 ; X32-AVX512BW-NEXT: vmovdqa64 %xmm0, (%eax) ; X32-AVX512BW-NEXT: vinserti32x4 $1, %xmm0, %ymm0, %ymm0 ; X32-AVX512BW-NEXT: retl ; ; X32-AVX512DQ-LABEL: test_broadcast_2i64_4i64_reuse: ; X32-AVX512DQ: ## BB#0: ; X32-AVX512DQ-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512DQ-NEXT: movl {{[0-9]+}}(%esp), %ecx ; X32-AVX512DQ-NEXT: vmovdqa64 (%ecx), %xmm0 ; X32-AVX512DQ-NEXT: vmovdqa64 %xmm0, (%eax) ; X32-AVX512DQ-NEXT: vinserti64x2 $1, %xmm0, %ymm0, %ymm0 ; X32-AVX512DQ-NEXT: retl ; ; X64-AVX-LABEL: test_broadcast_2i64_4i64_reuse: ; X64-AVX: ## BB#0: ; X64-AVX-NEXT: vmovaps (%rdi), %xmm0 ; X64-AVX-NEXT: vmovaps %xmm0, (%rsi) ; X64-AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 ; X64-AVX-NEXT: retq ; ; X64-AVX512F-LABEL: test_broadcast_2i64_4i64_reuse: ; X64-AVX512F: ## BB#0: ; X64-AVX512F-NEXT: vmovdqa64 (%rdi), %xmm0 ; X64-AVX512F-NEXT: vmovdqa64 %xmm0, (%rsi) ; X64-AVX512F-NEXT: vinserti32x4 $1, %xmm0, %ymm0, %ymm0 ; X64-AVX512F-NEXT: retq ; ; X64-AVX512BW-LABEL: test_broadcast_2i64_4i64_reuse: ; X64-AVX512BW: ## BB#0: ; X64-AVX512BW-NEXT: vmovdqa64 (%rdi), %xmm0 ; X64-AVX512BW-NEXT: vmovdqa64 %xmm0, (%rsi) ; X64-AVX512BW-NEXT: vinserti32x4 $1, %xmm0, %ymm0, %ymm0 ; X64-AVX512BW-NEXT: retq ; ; X64-AVX512DQ-LABEL: test_broadcast_2i64_4i64_reuse: ; X64-AVX512DQ: ## BB#0: ; X64-AVX512DQ-NEXT: vmovdqa64 (%rdi), %xmm0 ; X64-AVX512DQ-NEXT: vmovdqa64 %xmm0, (%rsi) ; X64-AVX512DQ-NEXT: vinserti64x2 $1, %xmm0, %ymm0, %ymm0 ; X64-AVX512DQ-NEXT: retq %1 = load <2 x i64>, <2 x i64>* %p0 store <2 x i64> %1, <2 x i64>* %p1 %2 = shufflevector <2 x i64> %1, <2 x i64> undef, <4 x i32> ret <4 x i64> %2 } define <8 x float> @test_broadcast_4f32_8f32_reuse(<4 x float>* %p0, <4 x float>* %p1) { ; X32-AVX-LABEL: test_broadcast_4f32_8f32_reuse: ; X32-AVX: ## BB#0: ; X32-AVX-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX-NEXT: movl {{[0-9]+}}(%esp), %ecx ; X32-AVX-NEXT: vmovaps (%ecx), %xmm0 ; X32-AVX-NEXT: vmovaps %xmm0, (%eax) ; X32-AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 ; X32-AVX-NEXT: retl ; ; X32-AVX512-LABEL: test_broadcast_4f32_8f32_reuse: ; X32-AVX512: ## BB#0: ; X32-AVX512-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512-NEXT: movl {{[0-9]+}}(%esp), %ecx ; X32-AVX512-NEXT: vmovaps (%ecx), %xmm0 ; X32-AVX512-NEXT: vmovaps %xmm0, (%eax) ; X32-AVX512-NEXT: vinsertf32x4 $1, %xmm0, %ymm0, %ymm0 ; X32-AVX512-NEXT: retl ; ; X64-AVX-LABEL: test_broadcast_4f32_8f32_reuse: ; X64-AVX: ## BB#0: ; X64-AVX-NEXT: vmovaps (%rdi), %xmm0 ; X64-AVX-NEXT: vmovaps %xmm0, (%rsi) ; X64-AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 ; X64-AVX-NEXT: retq ; ; X64-AVX512-LABEL: test_broadcast_4f32_8f32_reuse: ; X64-AVX512: ## BB#0: ; X64-AVX512-NEXT: vmovaps (%rdi), %xmm0 ; X64-AVX512-NEXT: vmovaps %xmm0, (%rsi) ; X64-AVX512-NEXT: vinsertf32x4 $1, %xmm0, %ymm0, %ymm0 ; X64-AVX512-NEXT: retq %1 = load <4 x float>, <4 x float>* %p0 store <4 x float> %1, <4 x float>* %p1 %2 = shufflevector <4 x float> %1, <4 x float> undef, <8 x i32> ret <8 x float> %2 } define <8 x i32> @test_broadcast_4i32_8i32_reuse(<4 x i32>* %p0, <4 x i32>* %p1) { ; X32-AVX-LABEL: test_broadcast_4i32_8i32_reuse: ; X32-AVX: ## BB#0: ; X32-AVX-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX-NEXT: movl {{[0-9]+}}(%esp), %ecx ; X32-AVX-NEXT: vmovaps (%ecx), %xmm0 ; X32-AVX-NEXT: vmovaps %xmm0, (%eax) ; X32-AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 ; X32-AVX-NEXT: retl ; ; X32-AVX512-LABEL: test_broadcast_4i32_8i32_reuse: ; X32-AVX512: ## BB#0: ; X32-AVX512-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512-NEXT: movl {{[0-9]+}}(%esp), %ecx ; X32-AVX512-NEXT: vmovdqa32 (%ecx), %xmm0 ; X32-AVX512-NEXT: vmovdqa32 %xmm0, (%eax) ; X32-AVX512-NEXT: vinserti32x4 $1, %xmm0, %ymm0, %ymm0 ; X32-AVX512-NEXT: retl ; ; X64-AVX-LABEL: test_broadcast_4i32_8i32_reuse: ; X64-AVX: ## BB#0: ; X64-AVX-NEXT: vmovaps (%rdi), %xmm0 ; X64-AVX-NEXT: vmovaps %xmm0, (%rsi) ; X64-AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 ; X64-AVX-NEXT: retq ; ; X64-AVX512-LABEL: test_broadcast_4i32_8i32_reuse: ; X64-AVX512: ## BB#0: ; X64-AVX512-NEXT: vmovdqa32 (%rdi), %xmm0 ; X64-AVX512-NEXT: vmovdqa32 %xmm0, (%rsi) ; X64-AVX512-NEXT: vinserti32x4 $1, %xmm0, %ymm0, %ymm0 ; X64-AVX512-NEXT: retq %1 = load <4 x i32>, <4 x i32>* %p0 store <4 x i32> %1, <4 x i32>* %p1 %2 = shufflevector <4 x i32> %1, <4 x i32> undef, <8 x i32> ret <8 x i32> %2 } define <16 x i16> @test_broadcast_8i16_16i16_reuse(<8 x i16> *%p0, <8 x i16> *%p1) nounwind { ; X32-AVX-LABEL: test_broadcast_8i16_16i16_reuse: ; X32-AVX: ## BB#0: ; X32-AVX-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX-NEXT: movl {{[0-9]+}}(%esp), %ecx ; X32-AVX-NEXT: vmovaps (%ecx), %xmm0 ; X32-AVX-NEXT: vmovaps %xmm0, (%eax) ; X32-AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 ; X32-AVX-NEXT: retl ; ; X32-AVX512F-LABEL: test_broadcast_8i16_16i16_reuse: ; X32-AVX512F: ## BB#0: ; X32-AVX512F-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512F-NEXT: movl {{[0-9]+}}(%esp), %ecx ; X32-AVX512F-NEXT: vmovdqa64 (%ecx), %xmm0 ; X32-AVX512F-NEXT: vmovdqa32 %xmm0, (%eax) ; X32-AVX512F-NEXT: vinserti32x4 $1, %xmm0, %ymm0, %ymm0 ; X32-AVX512F-NEXT: retl ; ; X32-AVX512BW-LABEL: test_broadcast_8i16_16i16_reuse: ; X32-AVX512BW: ## BB#0: ; X32-AVX512BW-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512BW-NEXT: movl {{[0-9]+}}(%esp), %ecx ; X32-AVX512BW-NEXT: vmovdqu16 (%ecx), %xmm0 ; X32-AVX512BW-NEXT: vmovdqu16 %xmm0, (%eax) ; X32-AVX512BW-NEXT: vinserti32x4 $1, %xmm0, %ymm0, %ymm0 ; X32-AVX512BW-NEXT: retl ; ; X32-AVX512DQ-LABEL: test_broadcast_8i16_16i16_reuse: ; X32-AVX512DQ: ## BB#0: ; X32-AVX512DQ-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512DQ-NEXT: movl {{[0-9]+}}(%esp), %ecx ; X32-AVX512DQ-NEXT: vmovdqa64 (%ecx), %xmm0 ; X32-AVX512DQ-NEXT: vmovdqa32 %xmm0, (%eax) ; X32-AVX512DQ-NEXT: vinserti32x4 $1, %xmm0, %ymm0, %ymm0 ; X32-AVX512DQ-NEXT: retl ; ; X64-AVX-LABEL: test_broadcast_8i16_16i16_reuse: ; X64-AVX: ## BB#0: ; X64-AVX-NEXT: vmovaps (%rdi), %xmm0 ; X64-AVX-NEXT: vmovaps %xmm0, (%rsi) ; X64-AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 ; X64-AVX-NEXT: retq ; ; X64-AVX512F-LABEL: test_broadcast_8i16_16i16_reuse: ; X64-AVX512F: ## BB#0: ; X64-AVX512F-NEXT: vmovdqa64 (%rdi), %xmm0 ; X64-AVX512F-NEXT: vmovdqa32 %xmm0, (%rsi) ; X64-AVX512F-NEXT: vinserti32x4 $1, %xmm0, %ymm0, %ymm0 ; X64-AVX512F-NEXT: retq ; ; X64-AVX512BW-LABEL: test_broadcast_8i16_16i16_reuse: ; X64-AVX512BW: ## BB#0: ; X64-AVX512BW-NEXT: vmovdqu16 (%rdi), %xmm0 ; X64-AVX512BW-NEXT: vmovdqu16 %xmm0, (%rsi) ; X64-AVX512BW-NEXT: vinserti32x4 $1, %xmm0, %ymm0, %ymm0 ; X64-AVX512BW-NEXT: retq ; ; X64-AVX512DQ-LABEL: test_broadcast_8i16_16i16_reuse: ; X64-AVX512DQ: ## BB#0: ; X64-AVX512DQ-NEXT: vmovdqa64 (%rdi), %xmm0 ; X64-AVX512DQ-NEXT: vmovdqa32 %xmm0, (%rsi) ; X64-AVX512DQ-NEXT: vinserti32x4 $1, %xmm0, %ymm0, %ymm0 ; X64-AVX512DQ-NEXT: retq %1 = load <8 x i16>, <8 x i16> *%p0 store <8 x i16> %1, <8 x i16>* %p1 %2 = shufflevector <8 x i16> %1, <8 x i16> undef, <16 x i32> ret <16 x i16> %2 } define <32 x i8> @test_broadcast_16i8_32i8_reuse(<16 x i8> *%p0, <16 x i8> *%p1) nounwind { ; X32-AVX-LABEL: test_broadcast_16i8_32i8_reuse: ; X32-AVX: ## BB#0: ; X32-AVX-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX-NEXT: movl {{[0-9]+}}(%esp), %ecx ; X32-AVX-NEXT: vmovaps (%ecx), %xmm0 ; X32-AVX-NEXT: vmovaps %xmm0, (%eax) ; X32-AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 ; X32-AVX-NEXT: retl ; ; X32-AVX512F-LABEL: test_broadcast_16i8_32i8_reuse: ; X32-AVX512F: ## BB#0: ; X32-AVX512F-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512F-NEXT: movl {{[0-9]+}}(%esp), %ecx ; X32-AVX512F-NEXT: vmovdqa64 (%ecx), %xmm0 ; X32-AVX512F-NEXT: vmovdqa32 %xmm0, (%eax) ; X32-AVX512F-NEXT: vinserti32x4 $1, %xmm0, %ymm0, %ymm0 ; X32-AVX512F-NEXT: retl ; ; X32-AVX512BW-LABEL: test_broadcast_16i8_32i8_reuse: ; X32-AVX512BW: ## BB#0: ; X32-AVX512BW-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512BW-NEXT: movl {{[0-9]+}}(%esp), %ecx ; X32-AVX512BW-NEXT: vmovdqu8 (%ecx), %xmm0 ; X32-AVX512BW-NEXT: vmovdqu8 %xmm0, (%eax) ; X32-AVX512BW-NEXT: vinserti32x4 $1, %xmm0, %ymm0, %ymm0 ; X32-AVX512BW-NEXT: retl ; ; X32-AVX512DQ-LABEL: test_broadcast_16i8_32i8_reuse: ; X32-AVX512DQ: ## BB#0: ; X32-AVX512DQ-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512DQ-NEXT: movl {{[0-9]+}}(%esp), %ecx ; X32-AVX512DQ-NEXT: vmovdqa64 (%ecx), %xmm0 ; X32-AVX512DQ-NEXT: vmovdqa32 %xmm0, (%eax) ; X32-AVX512DQ-NEXT: vinserti32x4 $1, %xmm0, %ymm0, %ymm0 ; X32-AVX512DQ-NEXT: retl ; ; X64-AVX-LABEL: test_broadcast_16i8_32i8_reuse: ; X64-AVX: ## BB#0: ; X64-AVX-NEXT: vmovaps (%rdi), %xmm0 ; X64-AVX-NEXT: vmovaps %xmm0, (%rsi) ; X64-AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 ; X64-AVX-NEXT: retq ; ; X64-AVX512F-LABEL: test_broadcast_16i8_32i8_reuse: ; X64-AVX512F: ## BB#0: ; X64-AVX512F-NEXT: vmovdqa64 (%rdi), %xmm0 ; X64-AVX512F-NEXT: vmovdqa32 %xmm0, (%rsi) ; X64-AVX512F-NEXT: vinserti32x4 $1, %xmm0, %ymm0, %ymm0 ; X64-AVX512F-NEXT: retq ; ; X64-AVX512BW-LABEL: test_broadcast_16i8_32i8_reuse: ; X64-AVX512BW: ## BB#0: ; X64-AVX512BW-NEXT: vmovdqu8 (%rdi), %xmm0 ; X64-AVX512BW-NEXT: vmovdqu8 %xmm0, (%rsi) ; X64-AVX512BW-NEXT: vinserti32x4 $1, %xmm0, %ymm0, %ymm0 ; X64-AVX512BW-NEXT: retq ; ; X64-AVX512DQ-LABEL: test_broadcast_16i8_32i8_reuse: ; X64-AVX512DQ: ## BB#0: ; X64-AVX512DQ-NEXT: vmovdqa64 (%rdi), %xmm0 ; X64-AVX512DQ-NEXT: vmovdqa32 %xmm0, (%rsi) ; X64-AVX512DQ-NEXT: vinserti32x4 $1, %xmm0, %ymm0, %ymm0 ; X64-AVX512DQ-NEXT: retq %1 = load <16 x i8>, <16 x i8> *%p0 store <16 x i8> %1, <16 x i8>* %p1 %2 = shufflevector <16 x i8> %1, <16 x i8> undef, <32 x i32> ret <32 x i8> %2 } ; ; Subvector Load + Broadcast with Separate Store ; define <8 x i32> @test_broadcast_4i32_8i32_chain(<4 x i32>* %p0, <4 x float>* %p1) { ; X32-AVX-LABEL: test_broadcast_4i32_8i32_chain: ; X32-AVX: ## BB#0: ; X32-AVX-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX-NEXT: movl {{[0-9]+}}(%esp), %ecx ; X32-AVX-NEXT: vmovaps (%ecx), %xmm0 ; X32-AVX-NEXT: vxorps %xmm1, %xmm1, %xmm1 ; X32-AVX-NEXT: vmovaps %xmm1, (%eax) ; X32-AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 ; X32-AVX-NEXT: retl ; ; X32-AVX512F-LABEL: test_broadcast_4i32_8i32_chain: ; X32-AVX512F: ## BB#0: ; X32-AVX512F-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512F-NEXT: movl {{[0-9]+}}(%esp), %ecx ; X32-AVX512F-NEXT: vmovdqa64 (%ecx), %xmm0 ; X32-AVX512F-NEXT: vpxord %xmm1, %xmm1, %xmm1 ; X32-AVX512F-NEXT: vmovdqa32 %xmm1, (%eax) ; X32-AVX512F-NEXT: vinserti32x4 $1, %xmm0, %ymm0, %ymm0 ; X32-AVX512F-NEXT: retl ; ; X32-AVX512BW-LABEL: test_broadcast_4i32_8i32_chain: ; X32-AVX512BW: ## BB#0: ; X32-AVX512BW-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512BW-NEXT: movl {{[0-9]+}}(%esp), %ecx ; X32-AVX512BW-NEXT: vmovdqa64 (%ecx), %xmm0 ; X32-AVX512BW-NEXT: vpxord %xmm1, %xmm1, %xmm1 ; X32-AVX512BW-NEXT: vmovdqa32 %xmm1, (%eax) ; X32-AVX512BW-NEXT: vinserti32x4 $1, %xmm0, %ymm0, %ymm0 ; X32-AVX512BW-NEXT: retl ; ; X32-AVX512DQ-LABEL: test_broadcast_4i32_8i32_chain: ; X32-AVX512DQ: ## BB#0: ; X32-AVX512DQ-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512DQ-NEXT: movl {{[0-9]+}}(%esp), %ecx ; X32-AVX512DQ-NEXT: vmovdqa64 (%ecx), %xmm0 ; X32-AVX512DQ-NEXT: vxorps %xmm1, %xmm1, %xmm1 ; X32-AVX512DQ-NEXT: vmovaps %xmm1, (%eax) ; X32-AVX512DQ-NEXT: vinserti32x4 $1, %xmm0, %ymm0, %ymm0 ; X32-AVX512DQ-NEXT: retl ; ; X64-AVX-LABEL: test_broadcast_4i32_8i32_chain: ; X64-AVX: ## BB#0: ; X64-AVX-NEXT: vmovaps (%rdi), %xmm0 ; X64-AVX-NEXT: vxorps %xmm1, %xmm1, %xmm1 ; X64-AVX-NEXT: vmovaps %xmm1, (%rsi) ; X64-AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 ; X64-AVX-NEXT: retq ; ; X64-AVX512F-LABEL: test_broadcast_4i32_8i32_chain: ; X64-AVX512F: ## BB#0: ; X64-AVX512F-NEXT: vmovdqa64 (%rdi), %xmm0 ; X64-AVX512F-NEXT: vpxord %xmm1, %xmm1, %xmm1 ; X64-AVX512F-NEXT: vmovdqa32 %xmm1, (%rsi) ; X64-AVX512F-NEXT: vinserti32x4 $1, %xmm0, %ymm0, %ymm0 ; X64-AVX512F-NEXT: retq ; ; X64-AVX512BW-LABEL: test_broadcast_4i32_8i32_chain: ; X64-AVX512BW: ## BB#0: ; X64-AVX512BW-NEXT: vmovdqa64 (%rdi), %xmm0 ; X64-AVX512BW-NEXT: vpxord %xmm1, %xmm1, %xmm1 ; X64-AVX512BW-NEXT: vmovdqa32 %xmm1, (%rsi) ; X64-AVX512BW-NEXT: vinserti32x4 $1, %xmm0, %ymm0, %ymm0 ; X64-AVX512BW-NEXT: retq ; ; X64-AVX512DQ-LABEL: test_broadcast_4i32_8i32_chain: ; X64-AVX512DQ: ## BB#0: ; X64-AVX512DQ-NEXT: vmovdqa64 (%rdi), %xmm0 ; X64-AVX512DQ-NEXT: vxorps %xmm1, %xmm1, %xmm1 ; X64-AVX512DQ-NEXT: vmovaps %xmm1, (%rsi) ; X64-AVX512DQ-NEXT: vinserti32x4 $1, %xmm0, %ymm0, %ymm0 ; X64-AVX512DQ-NEXT: retq %1 = load <4 x i32>, <4 x i32>* %p0 store <4 x float> zeroinitializer, <4 x float>* %p1 %2 = shufflevector <4 x i32> %1, <4 x i32> undef, <8 x i32> ret <8 x i32> %2 } define <16 x i32> @test_broadcast_4i32_16i32_chain(<4 x i32>* %p0, <4 x float>* %p1) { ; X32-AVX-LABEL: test_broadcast_4i32_16i32_chain: ; X32-AVX: ## BB#0: ; X32-AVX-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX-NEXT: movl {{[0-9]+}}(%esp), %ecx ; X32-AVX-NEXT: vmovaps (%ecx), %xmm0 ; X32-AVX-NEXT: vxorps %xmm1, %xmm1, %xmm1 ; X32-AVX-NEXT: vmovaps %xmm1, (%eax) ; X32-AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 ; X32-AVX-NEXT: vmovaps %ymm0, %ymm1 ; X32-AVX-NEXT: retl ; ; X32-AVX512F-LABEL: test_broadcast_4i32_16i32_chain: ; X32-AVX512F: ## BB#0: ; X32-AVX512F-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512F-NEXT: movl {{[0-9]+}}(%esp), %ecx ; X32-AVX512F-NEXT: vmovdqa64 (%ecx), %xmm0 ; X32-AVX512F-NEXT: vpxord %xmm1, %xmm1, %xmm1 ; X32-AVX512F-NEXT: vmovdqa32 %xmm1, (%eax) ; X32-AVX512F-NEXT: vinserti32x4 $1, %xmm0, %ymm0, %ymm0 ; X32-AVX512F-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0 ; X32-AVX512F-NEXT: retl ; ; X32-AVX512BW-LABEL: test_broadcast_4i32_16i32_chain: ; X32-AVX512BW: ## BB#0: ; X32-AVX512BW-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512BW-NEXT: movl {{[0-9]+}}(%esp), %ecx ; X32-AVX512BW-NEXT: vmovdqa64 (%ecx), %xmm0 ; X32-AVX512BW-NEXT: vpxord %xmm1, %xmm1, %xmm1 ; X32-AVX512BW-NEXT: vmovdqa32 %xmm1, (%eax) ; X32-AVX512BW-NEXT: vinserti32x4 $1, %xmm0, %ymm0, %ymm0 ; X32-AVX512BW-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0 ; X32-AVX512BW-NEXT: retl ; ; X32-AVX512DQ-LABEL: test_broadcast_4i32_16i32_chain: ; X32-AVX512DQ: ## BB#0: ; X32-AVX512DQ-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-AVX512DQ-NEXT: movl {{[0-9]+}}(%esp), %ecx ; X32-AVX512DQ-NEXT: vmovdqa64 (%ecx), %xmm0 ; X32-AVX512DQ-NEXT: vxorps %xmm1, %xmm1, %xmm1 ; X32-AVX512DQ-NEXT: vmovaps %xmm1, (%eax) ; X32-AVX512DQ-NEXT: vinserti32x4 $1, %xmm0, %ymm0, %ymm0 ; X32-AVX512DQ-NEXT: vinserti32x8 $1, %ymm0, %zmm0, %zmm0 ; X32-AVX512DQ-NEXT: retl ; ; X64-AVX-LABEL: test_broadcast_4i32_16i32_chain: ; X64-AVX: ## BB#0: ; X64-AVX-NEXT: vmovaps (%rdi), %xmm0 ; X64-AVX-NEXT: vxorps %xmm1, %xmm1, %xmm1 ; X64-AVX-NEXT: vmovaps %xmm1, (%rsi) ; X64-AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0 ; X64-AVX-NEXT: vmovaps %ymm0, %ymm1 ; X64-AVX-NEXT: retq ; ; X64-AVX512F-LABEL: test_broadcast_4i32_16i32_chain: ; X64-AVX512F: ## BB#0: ; X64-AVX512F-NEXT: vmovdqa64 (%rdi), %xmm0 ; X64-AVX512F-NEXT: vpxord %xmm1, %xmm1, %xmm1 ; X64-AVX512F-NEXT: vmovdqa32 %xmm1, (%rsi) ; X64-AVX512F-NEXT: vinserti32x4 $1, %xmm0, %ymm0, %ymm0 ; X64-AVX512F-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0 ; X64-AVX512F-NEXT: retq ; ; X64-AVX512BW-LABEL: test_broadcast_4i32_16i32_chain: ; X64-AVX512BW: ## BB#0: ; X64-AVX512BW-NEXT: vmovdqa64 (%rdi), %xmm0 ; X64-AVX512BW-NEXT: vpxord %xmm1, %xmm1, %xmm1 ; X64-AVX512BW-NEXT: vmovdqa32 %xmm1, (%rsi) ; X64-AVX512BW-NEXT: vinserti32x4 $1, %xmm0, %ymm0, %ymm0 ; X64-AVX512BW-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0 ; X64-AVX512BW-NEXT: retq ; ; X64-AVX512DQ-LABEL: test_broadcast_4i32_16i32_chain: ; X64-AVX512DQ: ## BB#0: ; X64-AVX512DQ-NEXT: vmovdqa64 (%rdi), %xmm0 ; X64-AVX512DQ-NEXT: vxorps %xmm1, %xmm1, %xmm1 ; X64-AVX512DQ-NEXT: vmovaps %xmm1, (%rsi) ; X64-AVX512DQ-NEXT: vinserti32x4 $1, %xmm0, %ymm0, %ymm0 ; X64-AVX512DQ-NEXT: vinserti32x8 $1, %ymm0, %zmm0, %zmm0 ; X64-AVX512DQ-NEXT: retq %1 = load <4 x i32>, <4 x i32>* %p0 store <4 x float> zeroinitializer, <4 x float>* %p1 %2 = shufflevector <4 x i32> %1, <4 x i32> undef, <16 x i32> ret <16 x i32> %2 }