[llvm] e653e03 - [X86] Add fcopysign vector test coverage
Simon Pilgrim via llvm-commits
llvm-commits at lists.llvm.org
Thu Nov 30 03:33:51 PST 2023
Author: Simon Pilgrim
Date: 2023-11-30T11:33:39Z
New Revision: e653e0303db024fe817e17c70c9d6d75ab9605b9
URL: https://github.com/llvm/llvm-project/commit/e653e0303db024fe817e17c70c9d6d75ab9605b9
DIFF: https://github.com/llvm/llvm-project/commit/e653e0303db024fe817e17c70c9d6d75ab9605b9.diff
LOG: [X86] Add fcopysign vector test coverage
Added:
llvm/test/CodeGen/X86/vec_fcopysign.ll
Modified:
Removed:
################################################################################
diff --git a/llvm/test/CodeGen/X86/vec_fcopysign.ll b/llvm/test/CodeGen/X86/vec_fcopysign.ll
new file mode 100644
index 000000000000000..631f6c2d0259bc1
--- /dev/null
+++ b/llvm/test/CodeGen/X86/vec_fcopysign.ll
@@ -0,0 +1,5721 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+sse2 | FileCheck %s --check-prefixes=X86,X86-SSE
+; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+avx | FileCheck %s --check-prefixes=X86,X86-AVX,X86-AVX1OR2,X86-AVX1
+; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefixes=X86,X86-AVX,X86-AVX1OR2,X86-AVX2
+; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+avx512vl | FileCheck %s --check-prefixes=X86,X86-AVX,X86-AVX512,X86-AVX512VL
+; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+avx512fp16 | FileCheck %s --check-prefixes=X86,X86-AVX,X86-AVX512,X86-AVX512FP16
+; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+avx512dq,+avx512vl | FileCheck %s --check-prefixes=X86,X86-AVX,X86-AVX512,X86-AVX512VLDQ
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse2 | FileCheck %s --check-prefixes=X64,X64-SSE
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s --check-prefixes=X64,X64-AVX,X64-AVX1OR2,X64-AVX1
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefixes=X64,X64-AVX,X64-AVX1OR2,X64-AVX2
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vl | FileCheck %s --check-prefixes=X64,X64-AVX,X64-AVX512,X64-AVX512VL
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512fp16 | FileCheck %s --check-prefixes=X64,X64-AVX,X64-AVX512,X64-AVX512FP16
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512dq,+avx512vl | FileCheck %s --check-prefixes=X64,X64-AVX,X64-AVX512,X64-AVX512VLDQ
+
+;
+; 128-bit Vectors
+;
+
+define <2 x double> @fcopysign_v2f64(<2 x double> %a0, <2 x double> %a1) nounwind {
+; X86-SSE-LABEL: fcopysign_v2f64:
+; X86-SSE: # %bb.0:
+; X86-SSE-NEXT: andps {{\.?LCPI[0-9]+_[0-9]+}}, %xmm1
+; X86-SSE-NEXT: andps {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0
+; X86-SSE-NEXT: orps %xmm1, %xmm0
+; X86-SSE-NEXT: retl
+;
+; X86-AVX1OR2-LABEL: fcopysign_v2f64:
+; X86-AVX1OR2: # %bb.0:
+; X86-AVX1OR2-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}, %xmm1, %xmm1
+; X86-AVX1OR2-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0
+; X86-AVX1OR2-NEXT: vorps %xmm1, %xmm0, %xmm0
+; X86-AVX1OR2-NEXT: retl
+;
+; X86-AVX512-LABEL: fcopysign_v2f64:
+; X86-AVX512: # %bb.0:
+; X86-AVX512-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}{1to2}, %xmm1, %xmm0
+; X86-AVX512-NEXT: retl
+;
+; X64-SSE-LABEL: fcopysign_v2f64:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: andps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; X64-SSE-NEXT: andps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; X64-SSE-NEXT: orps %xmm1, %xmm0
+; X64-SSE-NEXT: retq
+;
+; X64-AVX1OR2-LABEL: fcopysign_v2f64:
+; X64-AVX1OR2: # %bb.0:
+; X64-AVX1OR2-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1
+; X64-AVX1OR2-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
+; X64-AVX1OR2-NEXT: vorps %xmm1, %xmm0, %xmm0
+; X64-AVX1OR2-NEXT: retq
+;
+; X64-AVX512-LABEL: fcopysign_v2f64:
+; X64-AVX512: # %bb.0:
+; X64-AVX512-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to2}, %xmm1, %xmm0
+; X64-AVX512-NEXT: retq
+ %t = call <2 x double> @llvm.copysign.v2f64(<2 x double> %a0, <2 x double> %a1)
+ ret <2 x double> %t
+}
+declare <2 x double> @llvm.copysign.v2f64(<2 x double>, <2 x double>)
+
+define <4 x float> @fcopysign_v4f32(<4 x float> %a0, <4 x float> %a1) nounwind {
+; X86-SSE-LABEL: fcopysign_v4f32:
+; X86-SSE: # %bb.0:
+; X86-SSE-NEXT: andps {{\.?LCPI[0-9]+_[0-9]+}}, %xmm1
+; X86-SSE-NEXT: andps {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0
+; X86-SSE-NEXT: orps %xmm1, %xmm0
+; X86-SSE-NEXT: retl
+;
+; X86-AVX1-LABEL: fcopysign_v4f32:
+; X86-AVX1: # %bb.0:
+; X86-AVX1-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}, %xmm1, %xmm1
+; X86-AVX1-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0
+; X86-AVX1-NEXT: vorps %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: retl
+;
+; X86-AVX2-LABEL: fcopysign_v4f32:
+; X86-AVX2: # %bb.0:
+; X86-AVX2-NEXT: vbroadcastss {{.*#+}} xmm2 = [-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0]
+; X86-AVX2-NEXT: vandps %xmm2, %xmm1, %xmm1
+; X86-AVX2-NEXT: vbroadcastss {{.*#+}} xmm2 = [NaN,NaN,NaN,NaN]
+; X86-AVX2-NEXT: vandps %xmm2, %xmm0, %xmm0
+; X86-AVX2-NEXT: vorps %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: retl
+;
+; X86-AVX512-LABEL: fcopysign_v4f32:
+; X86-AVX512: # %bb.0:
+; X86-AVX512-NEXT: vpternlogd $228, {{\.?LCPI[0-9]+_[0-9]+}}{1to4}, %xmm1, %xmm0
+; X86-AVX512-NEXT: retl
+;
+; X64-SSE-LABEL: fcopysign_v4f32:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: andps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; X64-SSE-NEXT: andps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; X64-SSE-NEXT: orps %xmm1, %xmm0
+; X64-SSE-NEXT: retq
+;
+; X64-AVX1-LABEL: fcopysign_v4f32:
+; X64-AVX1: # %bb.0:
+; X64-AVX1-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1
+; X64-AVX1-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
+; X64-AVX1-NEXT: vorps %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: retq
+;
+; X64-AVX2-LABEL: fcopysign_v4f32:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: vbroadcastss {{.*#+}} xmm2 = [-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0]
+; X64-AVX2-NEXT: vandps %xmm2, %xmm1, %xmm1
+; X64-AVX2-NEXT: vbroadcastss {{.*#+}} xmm2 = [NaN,NaN,NaN,NaN]
+; X64-AVX2-NEXT: vandps %xmm2, %xmm0, %xmm0
+; X64-AVX2-NEXT: vorps %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: retq
+;
+; X64-AVX512-LABEL: fcopysign_v4f32:
+; X64-AVX512: # %bb.0:
+; X64-AVX512-NEXT: vpternlogd $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm1, %xmm0
+; X64-AVX512-NEXT: retq
+ %t = call <4 x float> @llvm.copysign.v4f32(<4 x float> %a0, <4 x float> %a1)
+ ret <4 x float> %t
+}
+declare <4 x float> @llvm.copysign.v4f32(<4 x float>, <4 x float>)
+
+define <8 x half> @fcopysign_v8f16(ptr %p0, ptr %p1) nounwind {
+; X86-SSE-LABEL: fcopysign_v8f16:
+; X86-SSE: # %bb.0:
+; X86-SSE-NEXT: movl {{[0-9]+}}(%esp), %eax
+; X86-SSE-NEXT: movl {{[0-9]+}}(%esp), %ecx
+; X86-SSE-NEXT: movdqa (%ecx), %xmm1
+; X86-SSE-NEXT: movdqa (%eax), %xmm2
+; X86-SSE-NEXT: movdqa %xmm2, %xmm0
+; X86-SSE-NEXT: psrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-SSE-NEXT: pextrw $0, %xmm0, %eax
+; X86-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-SSE-NEXT: movdqa %xmm1, %xmm0
+; X86-SSE-NEXT: psrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-SSE-NEXT: pextrw $0, %xmm0, %ecx
+; X86-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %eax, %ecx
+; X86-SSE-NEXT: pinsrw $0, %ecx, %xmm3
+; X86-SSE-NEXT: movdqa %xmm2, %xmm0
+; X86-SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm2[3,3]
+; X86-SSE-NEXT: pextrw $0, %xmm0, %eax
+; X86-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-SSE-NEXT: movdqa %xmm1, %xmm0
+; X86-SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm1[3,3]
+; X86-SSE-NEXT: pextrw $0, %xmm0, %ecx
+; X86-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %eax, %ecx
+; X86-SSE-NEXT: pinsrw $0, %ecx, %xmm0
+; X86-SSE-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
+; X86-SSE-NEXT: movdqa %xmm2, %xmm3
+; X86-SSE-NEXT: psrldq {{.*#+}} xmm3 = xmm3[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-SSE-NEXT: pextrw $0, %xmm3, %eax
+; X86-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-SSE-NEXT: movdqa %xmm1, %xmm3
+; X86-SSE-NEXT: psrldq {{.*#+}} xmm3 = xmm3[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-SSE-NEXT: pextrw $0, %xmm3, %ecx
+; X86-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %eax, %ecx
+; X86-SSE-NEXT: pinsrw $0, %ecx, %xmm4
+; X86-SSE-NEXT: movdqa %xmm2, %xmm3
+; X86-SSE-NEXT: punpckhqdq {{.*#+}} xmm3 = xmm3[1],xmm2[1]
+; X86-SSE-NEXT: pextrw $0, %xmm3, %eax
+; X86-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-SSE-NEXT: movdqa %xmm1, %xmm3
+; X86-SSE-NEXT: punpckhqdq {{.*#+}} xmm3 = xmm3[1],xmm1[1]
+; X86-SSE-NEXT: pextrw $0, %xmm3, %ecx
+; X86-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %eax, %ecx
+; X86-SSE-NEXT: pinsrw $0, %ecx, %xmm3
+; X86-SSE-NEXT: punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
+; X86-SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1]
+; X86-SSE-NEXT: movdqa %xmm2, %xmm0
+; X86-SSE-NEXT: psrlq $48, %xmm0
+; X86-SSE-NEXT: pextrw $0, %xmm0, %eax
+; X86-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-SSE-NEXT: movdqa %xmm1, %xmm0
+; X86-SSE-NEXT: psrlq $48, %xmm0
+; X86-SSE-NEXT: pextrw $0, %xmm0, %ecx
+; X86-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %eax, %ecx
+; X86-SSE-NEXT: pinsrw $0, %ecx, %xmm0
+; X86-SSE-NEXT: movaps %xmm2, %xmm4
+; X86-SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[1,1],xmm2[1,1]
+; X86-SSE-NEXT: pextrw $0, %xmm4, %eax
+; X86-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-SSE-NEXT: movaps %xmm1, %xmm4
+; X86-SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[1,1],xmm1[1,1]
+; X86-SSE-NEXT: pextrw $0, %xmm4, %ecx
+; X86-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %eax, %ecx
+; X86-SSE-NEXT: pinsrw $0, %ecx, %xmm4
+; X86-SSE-NEXT: punpcklwd {{.*#+}} xmm4 = xmm4[0],xmm0[0],xmm4[1],xmm0[1],xmm4[2],xmm0[2],xmm4[3],xmm0[3]
+; X86-SSE-NEXT: pextrw $0, %xmm2, %eax
+; X86-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-SSE-NEXT: pextrw $0, %xmm1, %ecx
+; X86-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %eax, %ecx
+; X86-SSE-NEXT: pinsrw $0, %ecx, %xmm0
+; X86-SSE-NEXT: psrld $16, %xmm2
+; X86-SSE-NEXT: pextrw $0, %xmm2, %eax
+; X86-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-SSE-NEXT: psrld $16, %xmm1
+; X86-SSE-NEXT: pextrw $0, %xmm1, %ecx
+; X86-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %eax, %ecx
+; X86-SSE-NEXT: pinsrw $0, %ecx, %xmm1
+; X86-SSE-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; X86-SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1]
+; X86-SSE-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm3[0]
+; X86-SSE-NEXT: retl
+;
+; X86-AVX1-LABEL: fcopysign_v8f16:
+; X86-AVX1: # %bb.0:
+; X86-AVX1-NEXT: pushl %esi
+; X86-AVX1-NEXT: movl {{[0-9]+}}(%esp), %ecx
+; X86-AVX1-NEXT: movl {{[0-9]+}}(%esp), %eax
+; X86-AVX1-NEXT: vmovdqa (%eax), %xmm0
+; X86-AVX1-NEXT: vmovdqa (%ecx), %xmm1
+; X86-AVX1-NEXT: vbroadcastss 12(%ecx), %xmm2
+; X86-AVX1-NEXT: vpextrw $0, %xmm2, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vbroadcastss 12(%eax), %xmm2
+; X86-AVX1-NEXT: vpextrw $0, %xmm2, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm2
+; X86-AVX1-NEXT: vpsrldq {{.*#+}} xmm3 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX1-NEXT: vpextrw $0, %xmm3, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vpsrldq {{.*#+}} xmm3 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX1-NEXT: vpextrw $0, %xmm3, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm3
+; X86-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
+; X86-AVX1-NEXT: vbroadcastss 8(%ecx), %xmm3
+; X86-AVX1-NEXT: vpextrw $0, %xmm3, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vbroadcastss 8(%eax), %xmm3
+; X86-AVX1-NEXT: vpextrw $0, %xmm3, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm3
+; X86-AVX1-NEXT: vpsrldq {{.*#+}} xmm4 = xmm1[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX1-NEXT: vpextrw $0, %xmm4, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vpsrldq {{.*#+}} xmm4 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX1-NEXT: vpextrw $0, %xmm4, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm4
+; X86-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
+; X86-AVX1-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]
+; X86-AVX1-NEXT: vbroadcastss 4(%ecx), %xmm3
+; X86-AVX1-NEXT: vpextrw $0, %xmm3, %ecx
+; X86-AVX1-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X86-AVX1-NEXT: vbroadcastss 4(%eax), %xmm3
+; X86-AVX1-NEXT: vpextrw $0, %xmm3, %eax
+; X86-AVX1-NEXT: andl $32767, %eax # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %ecx, %eax
+; X86-AVX1-NEXT: vpinsrw $0, %eax, %xmm0, %xmm3
+; X86-AVX1-NEXT: vpsrlq $48, %xmm1, %xmm4
+; X86-AVX1-NEXT: vpextrw $0, %xmm4, %eax
+; X86-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-AVX1-NEXT: vpsrlq $48, %xmm0, %xmm4
+; X86-AVX1-NEXT: vpextrw $0, %xmm4, %ecx
+; X86-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %eax, %ecx
+; X86-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm4
+; X86-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
+; X86-AVX1-NEXT: vpextrw $0, %xmm1, %eax
+; X86-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-AVX1-NEXT: vpextrw $0, %xmm0, %ecx
+; X86-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %eax, %ecx
+; X86-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm4
+; X86-AVX1-NEXT: vpsrld $16, %xmm1, %xmm1
+; X86-AVX1-NEXT: vpextrw $0, %xmm1, %eax
+; X86-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-AVX1-NEXT: vpsrld $16, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpextrw $0, %xmm0, %ecx
+; X86-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %eax, %ecx
+; X86-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm4[0],xmm0[0],xmm4[1],xmm0[1],xmm4[2],xmm0[2],xmm4[3],xmm0[3]
+; X86-AVX1-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1]
+; X86-AVX1-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm2[0]
+; X86-AVX1-NEXT: popl %esi
+; X86-AVX1-NEXT: retl
+;
+; X86-AVX2-LABEL: fcopysign_v8f16:
+; X86-AVX2: # %bb.0:
+; X86-AVX2-NEXT: pushl %edi
+; X86-AVX2-NEXT: pushl %esi
+; X86-AVX2-NEXT: movl {{[0-9]+}}(%esp), %eax
+; X86-AVX2-NEXT: movl {{[0-9]+}}(%esp), %ecx
+; X86-AVX2-NEXT: vmovdqa (%ecx), %xmm0
+; X86-AVX2-NEXT: vmovdqa (%eax), %xmm1
+; X86-AVX2-NEXT: vpsrldq {{.*#+}} xmm2 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX2-NEXT: vpextrw $0, %xmm2, %edx
+; X86-AVX2-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX2-NEXT: vpsrldq {{.*#+}} xmm2 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX2-NEXT: vpextrw $0, %xmm2, %esi
+; X86-AVX2-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %edx, %esi
+; X86-AVX2-NEXT: vpinsrw $0, %esi, %xmm0, %xmm2
+; X86-AVX2-NEXT: movl $-32768, %edx # imm = 0x8000
+; X86-AVX2-NEXT: movl 12(%eax), %esi
+; X86-AVX2-NEXT: andl %edx, %esi
+; X86-AVX2-NEXT: movzwl 12(%ecx), %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm3
+; X86-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; X86-AVX2-NEXT: vpsrldq {{.*#+}} xmm3 = xmm1[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX2-NEXT: vpextrw $0, %xmm3, %esi
+; X86-AVX2-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX2-NEXT: vpsrldq {{.*#+}} xmm3 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX2-NEXT: vpextrw $0, %xmm3, %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm3
+; X86-AVX2-NEXT: movl 8(%eax), %esi
+; X86-AVX2-NEXT: andl %edx, %esi
+; X86-AVX2-NEXT: movzwl 8(%ecx), %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm4
+; X86-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; X86-AVX2-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]
+; X86-AVX2-NEXT: vpextrw $0, %xmm1, %esi
+; X86-AVX2-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX2-NEXT: vpextrw $0, %xmm0, %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm3
+; X86-AVX2-NEXT: vpsrld $16, %xmm1, %xmm4
+; X86-AVX2-NEXT: vpextrw $0, %xmm4, %esi
+; X86-AVX2-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX2-NEXT: vpsrld $16, %xmm0, %xmm4
+; X86-AVX2-NEXT: vpextrw $0, %xmm4, %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm4
+; X86-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
+; X86-AVX2-NEXT: vpsrlq $48, %xmm1, %xmm1
+; X86-AVX2-NEXT: vpextrw $0, %xmm1, %esi
+; X86-AVX2-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX2-NEXT: vpsrlq $48, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpextrw $0, %xmm0, %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm0
+; X86-AVX2-NEXT: andl 4(%eax), %edx
+; X86-AVX2-NEXT: movzwl 4(%ecx), %eax
+; X86-AVX2-NEXT: andl $32767, %eax # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %edx, %eax
+; X86-AVX2-NEXT: vpinsrw $0, %eax, %xmm0, %xmm1
+; X86-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; X86-AVX2-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm3[0],xmm0[0],xmm3[1],xmm0[1]
+; X86-AVX2-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm2[0]
+; X86-AVX2-NEXT: popl %esi
+; X86-AVX2-NEXT: popl %edi
+; X86-AVX2-NEXT: retl
+;
+; X86-AVX512VL-LABEL: fcopysign_v8f16:
+; X86-AVX512VL: # %bb.0:
+; X86-AVX512VL-NEXT: pushl %edi
+; X86-AVX512VL-NEXT: pushl %esi
+; X86-AVX512VL-NEXT: movl {{[0-9]+}}(%esp), %eax
+; X86-AVX512VL-NEXT: movl {{[0-9]+}}(%esp), %ecx
+; X86-AVX512VL-NEXT: vmovdqa (%ecx), %xmm0
+; X86-AVX512VL-NEXT: vmovdqa (%eax), %xmm1
+; X86-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm2 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm2, %edx
+; X86-AVX512VL-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm2 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm2, %esi
+; X86-AVX512VL-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %edx, %esi
+; X86-AVX512VL-NEXT: vpinsrw $0, %esi, %xmm0, %xmm2
+; X86-AVX512VL-NEXT: movl $-32768, %edx # imm = 0x8000
+; X86-AVX512VL-NEXT: movl 12(%eax), %esi
+; X86-AVX512VL-NEXT: andl %edx, %esi
+; X86-AVX512VL-NEXT: movzwl 12(%ecx), %edi
+; X86-AVX512VL-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %esi, %edi
+; X86-AVX512VL-NEXT: vpinsrw $0, %edi, %xmm0, %xmm3
+; X86-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; X86-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm3 = xmm1[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm3, %esi
+; X86-AVX512VL-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm3 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm3, %edi
+; X86-AVX512VL-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %esi, %edi
+; X86-AVX512VL-NEXT: vpinsrw $0, %edi, %xmm0, %xmm3
+; X86-AVX512VL-NEXT: movl 8(%eax), %esi
+; X86-AVX512VL-NEXT: andl %edx, %esi
+; X86-AVX512VL-NEXT: movzwl 8(%ecx), %edi
+; X86-AVX512VL-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %esi, %edi
+; X86-AVX512VL-NEXT: vpinsrw $0, %edi, %xmm0, %xmm4
+; X86-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; X86-AVX512VL-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm1, %esi
+; X86-AVX512VL-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm0, %edi
+; X86-AVX512VL-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %esi, %edi
+; X86-AVX512VL-NEXT: vpinsrw $0, %edi, %xmm0, %xmm3
+; X86-AVX512VL-NEXT: vpsrld $16, %xmm1, %xmm4
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm4, %esi
+; X86-AVX512VL-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX512VL-NEXT: vpsrld $16, %xmm0, %xmm4
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm4, %edi
+; X86-AVX512VL-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %esi, %edi
+; X86-AVX512VL-NEXT: vpinsrw $0, %edi, %xmm0, %xmm4
+; X86-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
+; X86-AVX512VL-NEXT: vpsrlq $48, %xmm1, %xmm1
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm1, %esi
+; X86-AVX512VL-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX512VL-NEXT: vpsrlq $48, %xmm0, %xmm0
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm0, %edi
+; X86-AVX512VL-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %esi, %edi
+; X86-AVX512VL-NEXT: vpinsrw $0, %edi, %xmm0, %xmm0
+; X86-AVX512VL-NEXT: andl 4(%eax), %edx
+; X86-AVX512VL-NEXT: movzwl 4(%ecx), %eax
+; X86-AVX512VL-NEXT: andl $32767, %eax # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %edx, %eax
+; X86-AVX512VL-NEXT: vpinsrw $0, %eax, %xmm0, %xmm1
+; X86-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; X86-AVX512VL-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm3[0],xmm0[0],xmm3[1],xmm0[1]
+; X86-AVX512VL-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm2[0]
+; X86-AVX512VL-NEXT: popl %esi
+; X86-AVX512VL-NEXT: popl %edi
+; X86-AVX512VL-NEXT: retl
+;
+; X86-AVX512FP16-LABEL: fcopysign_v8f16:
+; X86-AVX512FP16: # %bb.0:
+; X86-AVX512FP16-NEXT: movl {{[0-9]+}}(%esp), %eax
+; X86-AVX512FP16-NEXT: movl {{[0-9]+}}(%esp), %ecx
+; X86-AVX512FP16-NEXT: vmovdqa (%ecx), %xmm1
+; X86-AVX512FP16-NEXT: vpbroadcastw {{.*#+}} xmm0 = [NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN]
+; X86-AVX512FP16-NEXT: vpternlogq $202, (%eax), %xmm1, %xmm0
+; X86-AVX512FP16-NEXT: retl
+;
+; X86-AVX512VLDQ-LABEL: fcopysign_v8f16:
+; X86-AVX512VLDQ: # %bb.0:
+; X86-AVX512VLDQ-NEXT: pushl %edi
+; X86-AVX512VLDQ-NEXT: pushl %esi
+; X86-AVX512VLDQ-NEXT: movl {{[0-9]+}}(%esp), %eax
+; X86-AVX512VLDQ-NEXT: movl {{[0-9]+}}(%esp), %ecx
+; X86-AVX512VLDQ-NEXT: vmovdqa (%ecx), %xmm0
+; X86-AVX512VLDQ-NEXT: vmovdqa (%eax), %xmm1
+; X86-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm2 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm2, %edx
+; X86-AVX512VLDQ-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm2 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm2, %esi
+; X86-AVX512VLDQ-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %edx, %esi
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %esi, %xmm0, %xmm2
+; X86-AVX512VLDQ-NEXT: movl $-32768, %edx # imm = 0x8000
+; X86-AVX512VLDQ-NEXT: movl 12(%eax), %esi
+; X86-AVX512VLDQ-NEXT: andl %edx, %esi
+; X86-AVX512VLDQ-NEXT: movzwl 12(%ecx), %edi
+; X86-AVX512VLDQ-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %esi, %edi
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %edi, %xmm0, %xmm3
+; X86-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; X86-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm3 = xmm1[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm3, %esi
+; X86-AVX512VLDQ-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm3 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm3, %edi
+; X86-AVX512VLDQ-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %esi, %edi
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %edi, %xmm0, %xmm3
+; X86-AVX512VLDQ-NEXT: movl 8(%eax), %esi
+; X86-AVX512VLDQ-NEXT: andl %edx, %esi
+; X86-AVX512VLDQ-NEXT: movzwl 8(%ecx), %edi
+; X86-AVX512VLDQ-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %esi, %edi
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %edi, %xmm0, %xmm4
+; X86-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; X86-AVX512VLDQ-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm1, %esi
+; X86-AVX512VLDQ-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm0, %edi
+; X86-AVX512VLDQ-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %esi, %edi
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %edi, %xmm0, %xmm3
+; X86-AVX512VLDQ-NEXT: vpsrld $16, %xmm1, %xmm4
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm4, %esi
+; X86-AVX512VLDQ-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX512VLDQ-NEXT: vpsrld $16, %xmm0, %xmm4
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm4, %edi
+; X86-AVX512VLDQ-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %esi, %edi
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %edi, %xmm0, %xmm4
+; X86-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
+; X86-AVX512VLDQ-NEXT: vpsrlq $48, %xmm1, %xmm1
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm1, %esi
+; X86-AVX512VLDQ-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX512VLDQ-NEXT: vpsrlq $48, %xmm0, %xmm0
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm0, %edi
+; X86-AVX512VLDQ-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %esi, %edi
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %edi, %xmm0, %xmm0
+; X86-AVX512VLDQ-NEXT: andl 4(%eax), %edx
+; X86-AVX512VLDQ-NEXT: movzwl 4(%ecx), %eax
+; X86-AVX512VLDQ-NEXT: andl $32767, %eax # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %edx, %eax
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %eax, %xmm0, %xmm1
+; X86-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; X86-AVX512VLDQ-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm3[0],xmm0[0],xmm3[1],xmm0[1]
+; X86-AVX512VLDQ-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm2[0]
+; X86-AVX512VLDQ-NEXT: popl %esi
+; X86-AVX512VLDQ-NEXT: popl %edi
+; X86-AVX512VLDQ-NEXT: retl
+;
+; X64-SSE-LABEL: fcopysign_v8f16:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: movdqa (%rdi), %xmm1
+; X64-SSE-NEXT: movdqa (%rsi), %xmm2
+; X64-SSE-NEXT: movdqa %xmm2, %xmm0
+; X64-SSE-NEXT: psrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-SSE-NEXT: pextrw $0, %xmm0, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movdqa %xmm1, %xmm0
+; X64-SSE-NEXT: psrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-SSE-NEXT: pextrw $0, %xmm0, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm3
+; X64-SSE-NEXT: movdqa %xmm2, %xmm0
+; X64-SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm2[3,3]
+; X64-SSE-NEXT: pextrw $0, %xmm0, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movdqa %xmm1, %xmm0
+; X64-SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm1[3,3]
+; X64-SSE-NEXT: pextrw $0, %xmm0, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm0
+; X64-SSE-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
+; X64-SSE-NEXT: movdqa %xmm2, %xmm3
+; X64-SSE-NEXT: psrldq {{.*#+}} xmm3 = xmm3[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-SSE-NEXT: pextrw $0, %xmm3, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movdqa %xmm1, %xmm3
+; X64-SSE-NEXT: psrldq {{.*#+}} xmm3 = xmm3[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-SSE-NEXT: pextrw $0, %xmm3, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm4
+; X64-SSE-NEXT: movdqa %xmm2, %xmm3
+; X64-SSE-NEXT: punpckhqdq {{.*#+}} xmm3 = xmm3[1],xmm2[1]
+; X64-SSE-NEXT: pextrw $0, %xmm3, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movdqa %xmm1, %xmm3
+; X64-SSE-NEXT: punpckhqdq {{.*#+}} xmm3 = xmm3[1],xmm1[1]
+; X64-SSE-NEXT: pextrw $0, %xmm3, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm3
+; X64-SSE-NEXT: punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
+; X64-SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1]
+; X64-SSE-NEXT: movdqa %xmm2, %xmm0
+; X64-SSE-NEXT: psrlq $48, %xmm0
+; X64-SSE-NEXT: pextrw $0, %xmm0, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movdqa %xmm1, %xmm0
+; X64-SSE-NEXT: psrlq $48, %xmm0
+; X64-SSE-NEXT: pextrw $0, %xmm0, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm0
+; X64-SSE-NEXT: movaps %xmm2, %xmm4
+; X64-SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[1,1],xmm2[1,1]
+; X64-SSE-NEXT: pextrw $0, %xmm4, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movaps %xmm1, %xmm4
+; X64-SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[1,1],xmm1[1,1]
+; X64-SSE-NEXT: pextrw $0, %xmm4, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm4
+; X64-SSE-NEXT: punpcklwd {{.*#+}} xmm4 = xmm4[0],xmm0[0],xmm4[1],xmm0[1],xmm4[2],xmm0[2],xmm4[3],xmm0[3]
+; X64-SSE-NEXT: pextrw $0, %xmm2, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: pextrw $0, %xmm1, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm0
+; X64-SSE-NEXT: psrld $16, %xmm2
+; X64-SSE-NEXT: pextrw $0, %xmm2, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: psrld $16, %xmm1
+; X64-SSE-NEXT: pextrw $0, %xmm1, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm1
+; X64-SSE-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; X64-SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1]
+; X64-SSE-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm3[0]
+; X64-SSE-NEXT: retq
+;
+; X64-AVX1-LABEL: fcopysign_v8f16:
+; X64-AVX1: # %bb.0:
+; X64-AVX1-NEXT: vmovdqa (%rdi), %xmm0
+; X64-AVX1-NEXT: vmovdqa (%rsi), %xmm1
+; X64-AVX1-NEXT: vbroadcastss 12(%rsi), %xmm2
+; X64-AVX1-NEXT: vpextrw $0, %xmm2, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vbroadcastss 12(%rdi), %xmm2
+; X64-AVX1-NEXT: vpextrw $0, %xmm2, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm2
+; X64-AVX1-NEXT: vpsrldq {{.*#+}} xmm3 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX1-NEXT: vpextrw $0, %xmm3, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vpsrldq {{.*#+}} xmm3 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX1-NEXT: vpextrw $0, %xmm3, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm3
+; X64-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
+; X64-AVX1-NEXT: vbroadcastss 8(%rsi), %xmm3
+; X64-AVX1-NEXT: vpextrw $0, %xmm3, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vbroadcastss 8(%rdi), %xmm3
+; X64-AVX1-NEXT: vpextrw $0, %xmm3, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm3
+; X64-AVX1-NEXT: vpsrldq {{.*#+}} xmm4 = xmm1[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX1-NEXT: vpextrw $0, %xmm4, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vpsrldq {{.*#+}} xmm4 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX1-NEXT: vpextrw $0, %xmm4, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm4
+; X64-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
+; X64-AVX1-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]
+; X64-AVX1-NEXT: vbroadcastss 4(%rsi), %xmm3
+; X64-AVX1-NEXT: vpextrw $0, %xmm3, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vbroadcastss 4(%rdi), %xmm3
+; X64-AVX1-NEXT: vpextrw $0, %xmm3, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm3
+; X64-AVX1-NEXT: vpsrlq $48, %xmm1, %xmm4
+; X64-AVX1-NEXT: vpextrw $0, %xmm4, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vpsrlq $48, %xmm0, %xmm4
+; X64-AVX1-NEXT: vpextrw $0, %xmm4, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm4
+; X64-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
+; X64-AVX1-NEXT: vpextrw $0, %xmm1, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vpextrw $0, %xmm0, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm4
+; X64-AVX1-NEXT: vpsrld $16, %xmm1, %xmm1
+; X64-AVX1-NEXT: vpextrw $0, %xmm1, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vpsrld $16, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpextrw $0, %xmm0, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm4[0],xmm0[0],xmm4[1],xmm0[1],xmm4[2],xmm0[2],xmm4[3],xmm0[3]
+; X64-AVX1-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1]
+; X64-AVX1-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm2[0]
+; X64-AVX1-NEXT: retq
+;
+; X64-AVX2-LABEL: fcopysign_v8f16:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: vmovdqa (%rdi), %xmm0
+; X64-AVX2-NEXT: vmovdqa (%rsi), %xmm1
+; X64-AVX2-NEXT: vpsrldq {{.*#+}} xmm2 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX2-NEXT: vpextrw $0, %xmm2, %eax
+; X64-AVX2-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX2-NEXT: vpsrldq {{.*#+}} xmm2 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX2-NEXT: vpextrw $0, %xmm2, %ecx
+; X64-AVX2-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %eax, %ecx
+; X64-AVX2-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm2
+; X64-AVX2-NEXT: movl $-32768, %eax # imm = 0x8000
+; X64-AVX2-NEXT: movl 12(%rsi), %ecx
+; X64-AVX2-NEXT: andl %eax, %ecx
+; X64-AVX2-NEXT: movzwl 12(%rdi), %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm3
+; X64-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; X64-AVX2-NEXT: vpsrldq {{.*#+}} xmm3 = xmm1[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX2-NEXT: vpextrw $0, %xmm3, %ecx
+; X64-AVX2-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX2-NEXT: vpsrldq {{.*#+}} xmm3 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX2-NEXT: vpextrw $0, %xmm3, %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm3
+; X64-AVX2-NEXT: movl 8(%rsi), %ecx
+; X64-AVX2-NEXT: andl %eax, %ecx
+; X64-AVX2-NEXT: movzwl 8(%rdi), %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm4
+; X64-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; X64-AVX2-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]
+; X64-AVX2-NEXT: vpextrw $0, %xmm1, %ecx
+; X64-AVX2-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX2-NEXT: vpextrw $0, %xmm0, %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm3
+; X64-AVX2-NEXT: vpsrld $16, %xmm1, %xmm4
+; X64-AVX2-NEXT: vpextrw $0, %xmm4, %ecx
+; X64-AVX2-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX2-NEXT: vpsrld $16, %xmm0, %xmm4
+; X64-AVX2-NEXT: vpextrw $0, %xmm4, %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm4
+; X64-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
+; X64-AVX2-NEXT: vpsrlq $48, %xmm1, %xmm1
+; X64-AVX2-NEXT: vpextrw $0, %xmm1, %ecx
+; X64-AVX2-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX2-NEXT: vpsrlq $48, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpextrw $0, %xmm0, %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm0
+; X64-AVX2-NEXT: andl 4(%rsi), %eax
+; X64-AVX2-NEXT: movzwl 4(%rdi), %ecx
+; X64-AVX2-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %eax, %ecx
+; X64-AVX2-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm1
+; X64-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; X64-AVX2-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm3[0],xmm0[0],xmm3[1],xmm0[1]
+; X64-AVX2-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm2[0]
+; X64-AVX2-NEXT: retq
+;
+; X64-AVX512VL-LABEL: fcopysign_v8f16:
+; X64-AVX512VL: # %bb.0:
+; X64-AVX512VL-NEXT: vmovdqa (%rdi), %xmm0
+; X64-AVX512VL-NEXT: vmovdqa (%rsi), %xmm1
+; X64-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm2 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm2, %eax
+; X64-AVX512VL-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm2 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm2, %ecx
+; X64-AVX512VL-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %eax, %ecx
+; X64-AVX512VL-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm2
+; X64-AVX512VL-NEXT: movl $-32768, %eax # imm = 0x8000
+; X64-AVX512VL-NEXT: movl 12(%rsi), %ecx
+; X64-AVX512VL-NEXT: andl %eax, %ecx
+; X64-AVX512VL-NEXT: movzwl 12(%rdi), %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm3
+; X64-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; X64-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm3 = xmm1[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm3, %ecx
+; X64-AVX512VL-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm3 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm3, %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm3
+; X64-AVX512VL-NEXT: movl 8(%rsi), %ecx
+; X64-AVX512VL-NEXT: andl %eax, %ecx
+; X64-AVX512VL-NEXT: movzwl 8(%rdi), %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm4
+; X64-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; X64-AVX512VL-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm1, %ecx
+; X64-AVX512VL-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm0, %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm3
+; X64-AVX512VL-NEXT: vpsrld $16, %xmm1, %xmm4
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm4, %ecx
+; X64-AVX512VL-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VL-NEXT: vpsrld $16, %xmm0, %xmm4
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm4, %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm4
+; X64-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
+; X64-AVX512VL-NEXT: vpsrlq $48, %xmm1, %xmm1
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm1, %ecx
+; X64-AVX512VL-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VL-NEXT: vpsrlq $48, %xmm0, %xmm0
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm0, %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm0
+; X64-AVX512VL-NEXT: andl 4(%rsi), %eax
+; X64-AVX512VL-NEXT: movzwl 4(%rdi), %ecx
+; X64-AVX512VL-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %eax, %ecx
+; X64-AVX512VL-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm1
+; X64-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; X64-AVX512VL-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm3[0],xmm0[0],xmm3[1],xmm0[1]
+; X64-AVX512VL-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm2[0]
+; X64-AVX512VL-NEXT: retq
+;
+; X64-AVX512FP16-LABEL: fcopysign_v8f16:
+; X64-AVX512FP16: # %bb.0:
+; X64-AVX512FP16-NEXT: vmovdqa (%rdi), %xmm1
+; X64-AVX512FP16-NEXT: vpbroadcastq {{.*#+}} xmm0 = [9223231297218904063,9223231297218904063]
+; X64-AVX512FP16-NEXT: vpternlogq $202, (%rsi), %xmm1, %xmm0
+; X64-AVX512FP16-NEXT: retq
+;
+; X64-AVX512VLDQ-LABEL: fcopysign_v8f16:
+; X64-AVX512VLDQ: # %bb.0:
+; X64-AVX512VLDQ-NEXT: vmovdqa (%rdi), %xmm0
+; X64-AVX512VLDQ-NEXT: vmovdqa (%rsi), %xmm1
+; X64-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm2 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm2, %eax
+; X64-AVX512VLDQ-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm2 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm2, %ecx
+; X64-AVX512VLDQ-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %eax, %ecx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm2
+; X64-AVX512VLDQ-NEXT: movl $-32768, %eax # imm = 0x8000
+; X64-AVX512VLDQ-NEXT: movl 12(%rsi), %ecx
+; X64-AVX512VLDQ-NEXT: andl %eax, %ecx
+; X64-AVX512VLDQ-NEXT: movzwl 12(%rdi), %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm3
+; X64-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; X64-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm3 = xmm1[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm3, %ecx
+; X64-AVX512VLDQ-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm3 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm3, %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm3
+; X64-AVX512VLDQ-NEXT: movl 8(%rsi), %ecx
+; X64-AVX512VLDQ-NEXT: andl %eax, %ecx
+; X64-AVX512VLDQ-NEXT: movzwl 8(%rdi), %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm4
+; X64-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; X64-AVX512VLDQ-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm1, %ecx
+; X64-AVX512VLDQ-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm0, %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm3
+; X64-AVX512VLDQ-NEXT: vpsrld $16, %xmm1, %xmm4
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm4, %ecx
+; X64-AVX512VLDQ-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VLDQ-NEXT: vpsrld $16, %xmm0, %xmm4
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm4, %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm4
+; X64-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
+; X64-AVX512VLDQ-NEXT: vpsrlq $48, %xmm1, %xmm1
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm1, %ecx
+; X64-AVX512VLDQ-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VLDQ-NEXT: vpsrlq $48, %xmm0, %xmm0
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm0, %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm0
+; X64-AVX512VLDQ-NEXT: andl 4(%rsi), %eax
+; X64-AVX512VLDQ-NEXT: movzwl 4(%rdi), %ecx
+; X64-AVX512VLDQ-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %eax, %ecx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm1
+; X64-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; X64-AVX512VLDQ-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm3[0],xmm0[0],xmm3[1],xmm0[1]
+; X64-AVX512VLDQ-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm2[0]
+; X64-AVX512VLDQ-NEXT: retq
+ %a0 = load <8 x half>, ptr %p0, align 16
+ %a1 = load <8 x half>, ptr %p1, align 16
+ %t = call <8 x half> @llvm.copysign.v8f16(<8 x half> %a0, <8 x half> %a1)
+ ret <8 x half> %t
+}
+declare <8 x half> @llvm.copysign.v8f16(<8 x half>, <8 x half>)
+
+;
+; 256-bit Vectors
+;
+
+define <4 x double> @fcopysign_v4f64(<4 x double> %a0, <4 x double> %a1) nounwind {
+; X86-SSE-LABEL: fcopysign_v4f64:
+; X86-SSE: # %bb.0:
+; X86-SSE-NEXT: pushl %ebp
+; X86-SSE-NEXT: movl %esp, %ebp
+; X86-SSE-NEXT: andl $-16, %esp
+; X86-SSE-NEXT: subl $16, %esp
+; X86-SSE-NEXT: movaps {{.*#+}} xmm3 = [NaN,NaN]
+; X86-SSE-NEXT: movaps %xmm3, %xmm4
+; X86-SSE-NEXT: andnps %xmm2, %xmm4
+; X86-SSE-NEXT: andps %xmm3, %xmm0
+; X86-SSE-NEXT: orps %xmm4, %xmm0
+; X86-SSE-NEXT: andps %xmm3, %xmm1
+; X86-SSE-NEXT: andnps 8(%ebp), %xmm3
+; X86-SSE-NEXT: orps %xmm3, %xmm1
+; X86-SSE-NEXT: movl %ebp, %esp
+; X86-SSE-NEXT: popl %ebp
+; X86-SSE-NEXT: retl
+;
+; X86-AVX1-LABEL: fcopysign_v4f64:
+; X86-AVX1: # %bb.0:
+; X86-AVX1-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}, %ymm1, %ymm1
+; X86-AVX1-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}, %ymm0, %ymm0
+; X86-AVX1-NEXT: vorps %ymm1, %ymm0, %ymm0
+; X86-AVX1-NEXT: retl
+;
+; X86-AVX2-LABEL: fcopysign_v4f64:
+; X86-AVX2: # %bb.0:
+; X86-AVX2-NEXT: vbroadcastsd {{.*#+}} ymm2 = [-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0]
+; X86-AVX2-NEXT: vandps %ymm2, %ymm1, %ymm1
+; X86-AVX2-NEXT: vbroadcastsd {{.*#+}} ymm2 = [NaN,NaN,NaN,NaN]
+; X86-AVX2-NEXT: vandps %ymm2, %ymm0, %ymm0
+; X86-AVX2-NEXT: vorps %ymm1, %ymm0, %ymm0
+; X86-AVX2-NEXT: retl
+;
+; X86-AVX512-LABEL: fcopysign_v4f64:
+; X86-AVX512: # %bb.0:
+; X86-AVX512-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}{1to4}, %ymm1, %ymm0
+; X86-AVX512-NEXT: retl
+;
+; X64-SSE-LABEL: fcopysign_v4f64:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: movaps {{.*#+}} xmm4 = [NaN,NaN]
+; X64-SSE-NEXT: movaps %xmm4, %xmm5
+; X64-SSE-NEXT: andnps %xmm2, %xmm5
+; X64-SSE-NEXT: andps %xmm4, %xmm0
+; X64-SSE-NEXT: orps %xmm5, %xmm0
+; X64-SSE-NEXT: andps %xmm4, %xmm1
+; X64-SSE-NEXT: andnps %xmm3, %xmm4
+; X64-SSE-NEXT: orps %xmm4, %xmm1
+; X64-SSE-NEXT: retq
+;
+; X64-AVX1-LABEL: fcopysign_v4f64:
+; X64-AVX1: # %bb.0:
+; X64-AVX1-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm1
+; X64-AVX1-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
+; X64-AVX1-NEXT: vorps %ymm1, %ymm0, %ymm0
+; X64-AVX1-NEXT: retq
+;
+; X64-AVX2-LABEL: fcopysign_v4f64:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: vbroadcastsd {{.*#+}} ymm2 = [-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0]
+; X64-AVX2-NEXT: vandps %ymm2, %ymm1, %ymm1
+; X64-AVX2-NEXT: vbroadcastsd {{.*#+}} ymm2 = [NaN,NaN,NaN,NaN]
+; X64-AVX2-NEXT: vandps %ymm2, %ymm0, %ymm0
+; X64-AVX2-NEXT: vorps %ymm1, %ymm0, %ymm0
+; X64-AVX2-NEXT: retq
+;
+; X64-AVX512-LABEL: fcopysign_v4f64:
+; X64-AVX512: # %bb.0:
+; X64-AVX512-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %ymm1, %ymm0
+; X64-AVX512-NEXT: retq
+ %t = call <4 x double> @llvm.copysign.v4f64(<4 x double> %a0, <4 x double> %a1)
+ ret <4 x double> %t
+}
+declare <4 x double> @llvm.copysign.v4f64(<4 x double>, <4 x double>)
+
+define <8 x float> @fcopysign_v8f32(<8 x float> %a0, <8 x float> %a1) nounwind {
+; X86-SSE-LABEL: fcopysign_v8f32:
+; X86-SSE: # %bb.0:
+; X86-SSE-NEXT: pushl %ebp
+; X86-SSE-NEXT: movl %esp, %ebp
+; X86-SSE-NEXT: andl $-16, %esp
+; X86-SSE-NEXT: subl $16, %esp
+; X86-SSE-NEXT: movaps {{.*#+}} xmm3 = [NaN,NaN,NaN,NaN]
+; X86-SSE-NEXT: movaps %xmm3, %xmm4
+; X86-SSE-NEXT: andnps %xmm2, %xmm4
+; X86-SSE-NEXT: andps %xmm3, %xmm0
+; X86-SSE-NEXT: orps %xmm4, %xmm0
+; X86-SSE-NEXT: andps %xmm3, %xmm1
+; X86-SSE-NEXT: andnps 8(%ebp), %xmm3
+; X86-SSE-NEXT: orps %xmm3, %xmm1
+; X86-SSE-NEXT: movl %ebp, %esp
+; X86-SSE-NEXT: popl %ebp
+; X86-SSE-NEXT: retl
+;
+; X86-AVX1-LABEL: fcopysign_v8f32:
+; X86-AVX1: # %bb.0:
+; X86-AVX1-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}, %ymm1, %ymm1
+; X86-AVX1-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}, %ymm0, %ymm0
+; X86-AVX1-NEXT: vorps %ymm1, %ymm0, %ymm0
+; X86-AVX1-NEXT: retl
+;
+; X86-AVX2-LABEL: fcopysign_v8f32:
+; X86-AVX2: # %bb.0:
+; X86-AVX2-NEXT: vbroadcastss {{.*#+}} ymm2 = [-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0]
+; X86-AVX2-NEXT: vandps %ymm2, %ymm1, %ymm1
+; X86-AVX2-NEXT: vbroadcastss {{.*#+}} ymm2 = [NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN]
+; X86-AVX2-NEXT: vandps %ymm2, %ymm0, %ymm0
+; X86-AVX2-NEXT: vorps %ymm1, %ymm0, %ymm0
+; X86-AVX2-NEXT: retl
+;
+; X86-AVX512-LABEL: fcopysign_v8f32:
+; X86-AVX512: # %bb.0:
+; X86-AVX512-NEXT: vpternlogd $228, {{\.?LCPI[0-9]+_[0-9]+}}{1to8}, %ymm1, %ymm0
+; X86-AVX512-NEXT: retl
+;
+; X64-SSE-LABEL: fcopysign_v8f32:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: movaps {{.*#+}} xmm4 = [NaN,NaN,NaN,NaN]
+; X64-SSE-NEXT: movaps %xmm4, %xmm5
+; X64-SSE-NEXT: andnps %xmm2, %xmm5
+; X64-SSE-NEXT: andps %xmm4, %xmm0
+; X64-SSE-NEXT: orps %xmm5, %xmm0
+; X64-SSE-NEXT: andps %xmm4, %xmm1
+; X64-SSE-NEXT: andnps %xmm3, %xmm4
+; X64-SSE-NEXT: orps %xmm4, %xmm1
+; X64-SSE-NEXT: retq
+;
+; X64-AVX1-LABEL: fcopysign_v8f32:
+; X64-AVX1: # %bb.0:
+; X64-AVX1-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm1
+; X64-AVX1-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
+; X64-AVX1-NEXT: vorps %ymm1, %ymm0, %ymm0
+; X64-AVX1-NEXT: retq
+;
+; X64-AVX2-LABEL: fcopysign_v8f32:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: vbroadcastss {{.*#+}} ymm2 = [-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0]
+; X64-AVX2-NEXT: vandps %ymm2, %ymm1, %ymm1
+; X64-AVX2-NEXT: vbroadcastss {{.*#+}} ymm2 = [NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN]
+; X64-AVX2-NEXT: vandps %ymm2, %ymm0, %ymm0
+; X64-AVX2-NEXT: vorps %ymm1, %ymm0, %ymm0
+; X64-AVX2-NEXT: retq
+;
+; X64-AVX512-LABEL: fcopysign_v8f32:
+; X64-AVX512: # %bb.0:
+; X64-AVX512-NEXT: vpternlogd $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to8}, %ymm1, %ymm0
+; X64-AVX512-NEXT: retq
+ %t = call <8 x float> @llvm.copysign.v8f32(<8 x float> %a0, <8 x float> %a1)
+ ret <8 x float> %t
+}
+declare <8 x float> @llvm.copysign.v8f32(<8 x float>, <8 x float>)
+
+define <16 x half> @fcopysign_v16f16(ptr %p0, ptr %p1) nounwind {
+; X86-SSE-LABEL: fcopysign_v16f16:
+; X86-SSE: # %bb.0:
+; X86-SSE-NEXT: movl {{[0-9]+}}(%esp), %eax
+; X86-SSE-NEXT: movl {{[0-9]+}}(%esp), %ecx
+; X86-SSE-NEXT: movdqa (%ecx), %xmm1
+; X86-SSE-NEXT: movdqa 16(%ecx), %xmm2
+; X86-SSE-NEXT: movdqa (%eax), %xmm4
+; X86-SSE-NEXT: movdqa 16(%eax), %xmm3
+; X86-SSE-NEXT: movdqa %xmm4, %xmm0
+; X86-SSE-NEXT: psrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-SSE-NEXT: pextrw $0, %xmm0, %eax
+; X86-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-SSE-NEXT: movdqa %xmm1, %xmm0
+; X86-SSE-NEXT: psrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-SSE-NEXT: pextrw $0, %xmm0, %ecx
+; X86-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %eax, %ecx
+; X86-SSE-NEXT: pinsrw $0, %ecx, %xmm5
+; X86-SSE-NEXT: movdqa %xmm4, %xmm0
+; X86-SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm4[3,3]
+; X86-SSE-NEXT: pextrw $0, %xmm0, %eax
+; X86-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-SSE-NEXT: movdqa %xmm1, %xmm0
+; X86-SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm1[3,3]
+; X86-SSE-NEXT: pextrw $0, %xmm0, %ecx
+; X86-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %eax, %ecx
+; X86-SSE-NEXT: pinsrw $0, %ecx, %xmm0
+; X86-SSE-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm5[0],xmm0[1],xmm5[1],xmm0[2],xmm5[2],xmm0[3],xmm5[3]
+; X86-SSE-NEXT: movdqa %xmm4, %xmm5
+; X86-SSE-NEXT: psrldq {{.*#+}} xmm5 = xmm5[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-SSE-NEXT: pextrw $0, %xmm5, %eax
+; X86-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-SSE-NEXT: movdqa %xmm1, %xmm5
+; X86-SSE-NEXT: psrldq {{.*#+}} xmm5 = xmm5[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-SSE-NEXT: pextrw $0, %xmm5, %ecx
+; X86-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %eax, %ecx
+; X86-SSE-NEXT: pinsrw $0, %ecx, %xmm6
+; X86-SSE-NEXT: movdqa %xmm4, %xmm5
+; X86-SSE-NEXT: punpckhqdq {{.*#+}} xmm5 = xmm5[1],xmm4[1]
+; X86-SSE-NEXT: pextrw $0, %xmm5, %eax
+; X86-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-SSE-NEXT: movdqa %xmm1, %xmm5
+; X86-SSE-NEXT: punpckhqdq {{.*#+}} xmm5 = xmm5[1],xmm1[1]
+; X86-SSE-NEXT: pextrw $0, %xmm5, %ecx
+; X86-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %eax, %ecx
+; X86-SSE-NEXT: pinsrw $0, %ecx, %xmm5
+; X86-SSE-NEXT: punpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
+; X86-SSE-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm0[0],xmm5[1],xmm0[1]
+; X86-SSE-NEXT: movdqa %xmm4, %xmm0
+; X86-SSE-NEXT: psrlq $48, %xmm0
+; X86-SSE-NEXT: pextrw $0, %xmm0, %eax
+; X86-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-SSE-NEXT: movdqa %xmm1, %xmm0
+; X86-SSE-NEXT: psrlq $48, %xmm0
+; X86-SSE-NEXT: pextrw $0, %xmm0, %ecx
+; X86-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %eax, %ecx
+; X86-SSE-NEXT: pinsrw $0, %ecx, %xmm0
+; X86-SSE-NEXT: movaps %xmm4, %xmm6
+; X86-SSE-NEXT: shufps {{.*#+}} xmm6 = xmm6[1,1],xmm4[1,1]
+; X86-SSE-NEXT: pextrw $0, %xmm6, %eax
+; X86-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-SSE-NEXT: movaps %xmm1, %xmm6
+; X86-SSE-NEXT: shufps {{.*#+}} xmm6 = xmm6[1,1],xmm1[1,1]
+; X86-SSE-NEXT: pextrw $0, %xmm6, %ecx
+; X86-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %eax, %ecx
+; X86-SSE-NEXT: pinsrw $0, %ecx, %xmm6
+; X86-SSE-NEXT: punpcklwd {{.*#+}} xmm6 = xmm6[0],xmm0[0],xmm6[1],xmm0[1],xmm6[2],xmm0[2],xmm6[3],xmm0[3]
+; X86-SSE-NEXT: pextrw $0, %xmm4, %eax
+; X86-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-SSE-NEXT: pextrw $0, %xmm1, %ecx
+; X86-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %eax, %ecx
+; X86-SSE-NEXT: pinsrw $0, %ecx, %xmm0
+; X86-SSE-NEXT: psrld $16, %xmm4
+; X86-SSE-NEXT: pextrw $0, %xmm4, %eax
+; X86-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-SSE-NEXT: psrld $16, %xmm1
+; X86-SSE-NEXT: pextrw $0, %xmm1, %ecx
+; X86-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %eax, %ecx
+; X86-SSE-NEXT: pinsrw $0, %ecx, %xmm1
+; X86-SSE-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; X86-SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm6[0],xmm0[1],xmm6[1]
+; X86-SSE-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm5[0]
+; X86-SSE-NEXT: movdqa %xmm3, %xmm1
+; X86-SSE-NEXT: psrldq {{.*#+}} xmm1 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-SSE-NEXT: pextrw $0, %xmm1, %eax
+; X86-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-SSE-NEXT: movdqa %xmm2, %xmm1
+; X86-SSE-NEXT: psrldq {{.*#+}} xmm1 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-SSE-NEXT: pextrw $0, %xmm1, %ecx
+; X86-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %eax, %ecx
+; X86-SSE-NEXT: pinsrw $0, %ecx, %xmm4
+; X86-SSE-NEXT: movdqa %xmm3, %xmm1
+; X86-SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[3,3],xmm3[3,3]
+; X86-SSE-NEXT: pextrw $0, %xmm1, %eax
+; X86-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-SSE-NEXT: movdqa %xmm2, %xmm1
+; X86-SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[3,3],xmm2[3,3]
+; X86-SSE-NEXT: pextrw $0, %xmm1, %ecx
+; X86-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %eax, %ecx
+; X86-SSE-NEXT: pinsrw $0, %ecx, %xmm1
+; X86-SSE-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3]
+; X86-SSE-NEXT: movdqa %xmm3, %xmm4
+; X86-SSE-NEXT: psrldq {{.*#+}} xmm4 = xmm4[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-SSE-NEXT: pextrw $0, %xmm4, %eax
+; X86-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-SSE-NEXT: movdqa %xmm2, %xmm4
+; X86-SSE-NEXT: psrldq {{.*#+}} xmm4 = xmm4[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-SSE-NEXT: pextrw $0, %xmm4, %ecx
+; X86-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %eax, %ecx
+; X86-SSE-NEXT: pinsrw $0, %ecx, %xmm5
+; X86-SSE-NEXT: movdqa %xmm3, %xmm4
+; X86-SSE-NEXT: punpckhqdq {{.*#+}} xmm4 = xmm4[1],xmm3[1]
+; X86-SSE-NEXT: pextrw $0, %xmm4, %eax
+; X86-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-SSE-NEXT: movdqa %xmm2, %xmm4
+; X86-SSE-NEXT: punpckhqdq {{.*#+}} xmm4 = xmm4[1],xmm2[1]
+; X86-SSE-NEXT: pextrw $0, %xmm4, %ecx
+; X86-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %eax, %ecx
+; X86-SSE-NEXT: pinsrw $0, %ecx, %xmm4
+; X86-SSE-NEXT: punpcklwd {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
+; X86-SSE-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm1[0],xmm4[1],xmm1[1]
+; X86-SSE-NEXT: movdqa %xmm3, %xmm1
+; X86-SSE-NEXT: psrlq $48, %xmm1
+; X86-SSE-NEXT: pextrw $0, %xmm1, %eax
+; X86-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-SSE-NEXT: movdqa %xmm2, %xmm1
+; X86-SSE-NEXT: psrlq $48, %xmm1
+; X86-SSE-NEXT: pextrw $0, %xmm1, %ecx
+; X86-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %eax, %ecx
+; X86-SSE-NEXT: pinsrw $0, %ecx, %xmm1
+; X86-SSE-NEXT: movaps %xmm3, %xmm5
+; X86-SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[1,1],xmm3[1,1]
+; X86-SSE-NEXT: pextrw $0, %xmm5, %eax
+; X86-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-SSE-NEXT: movaps %xmm2, %xmm5
+; X86-SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[1,1],xmm2[1,1]
+; X86-SSE-NEXT: pextrw $0, %xmm5, %ecx
+; X86-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %eax, %ecx
+; X86-SSE-NEXT: pinsrw $0, %ecx, %xmm5
+; X86-SSE-NEXT: punpcklwd {{.*#+}} xmm5 = xmm5[0],xmm1[0],xmm5[1],xmm1[1],xmm5[2],xmm1[2],xmm5[3],xmm1[3]
+; X86-SSE-NEXT: pextrw $0, %xmm3, %eax
+; X86-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-SSE-NEXT: pextrw $0, %xmm2, %ecx
+; X86-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %eax, %ecx
+; X86-SSE-NEXT: pinsrw $0, %ecx, %xmm1
+; X86-SSE-NEXT: psrld $16, %xmm3
+; X86-SSE-NEXT: pextrw $0, %xmm3, %eax
+; X86-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-SSE-NEXT: psrld $16, %xmm2
+; X86-SSE-NEXT: pextrw $0, %xmm2, %ecx
+; X86-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %eax, %ecx
+; X86-SSE-NEXT: pinsrw $0, %ecx, %xmm2
+; X86-SSE-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
+; X86-SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm5[0],xmm1[1],xmm5[1]
+; X86-SSE-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm4[0]
+; X86-SSE-NEXT: retl
+;
+; X86-AVX1-LABEL: fcopysign_v16f16:
+; X86-AVX1: # %bb.0:
+; X86-AVX1-NEXT: pushl %esi
+; X86-AVX1-NEXT: movl {{[0-9]+}}(%esp), %eax
+; X86-AVX1-NEXT: movl {{[0-9]+}}(%esp), %ecx
+; X86-AVX1-NEXT: vbroadcastss 28(%ecx), %xmm0
+; X86-AVX1-NEXT: vpextrw $0, %xmm0, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vbroadcastss 28(%eax), %xmm0
+; X86-AVX1-NEXT: vpextrw $0, %xmm0, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm4
+; X86-AVX1-NEXT: vmovdqa (%ecx), %xmm0
+; X86-AVX1-NEXT: vmovdqa 16(%ecx), %xmm2
+; X86-AVX1-NEXT: vpsrldq {{.*#+}} xmm1 = xmm2[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX1-NEXT: vpextrw $0, %xmm1, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vmovdqa (%eax), %xmm1
+; X86-AVX1-NEXT: vmovdqa 16(%eax), %xmm3
+; X86-AVX1-NEXT: vpsrldq {{.*#+}} xmm5 = xmm3[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX1-NEXT: vpextrw $0, %xmm5, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm5
+; X86-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
+; X86-AVX1-NEXT: vbroadcastss 24(%ecx), %xmm5
+; X86-AVX1-NEXT: vpextrw $0, %xmm5, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vbroadcastss 24(%eax), %xmm5
+; X86-AVX1-NEXT: vpextrw $0, %xmm5, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm5
+; X86-AVX1-NEXT: vpsrldq {{.*#+}} xmm6 = xmm2[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX1-NEXT: vpextrw $0, %xmm6, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vpsrldq {{.*#+}} xmm6 = xmm3[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX1-NEXT: vpextrw $0, %xmm6, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm6
+; X86-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
+; X86-AVX1-NEXT: vpunpckldq {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]
+; X86-AVX1-NEXT: vbroadcastss 20(%ecx), %xmm5
+; X86-AVX1-NEXT: vpextrw $0, %xmm5, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vbroadcastss 20(%eax), %xmm5
+; X86-AVX1-NEXT: vpextrw $0, %xmm5, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm5
+; X86-AVX1-NEXT: vpsrlq $48, %xmm2, %xmm6
+; X86-AVX1-NEXT: vpextrw $0, %xmm6, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vpsrlq $48, %xmm3, %xmm6
+; X86-AVX1-NEXT: vpextrw $0, %xmm6, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm6
+; X86-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
+; X86-AVX1-NEXT: vpextrw $0, %xmm2, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vpextrw $0, %xmm3, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm6
+; X86-AVX1-NEXT: vpsrld $16, %xmm2, %xmm2
+; X86-AVX1-NEXT: vpextrw $0, %xmm2, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vpsrld $16, %xmm3, %xmm2
+; X86-AVX1-NEXT: vpextrw $0, %xmm2, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm2
+; X86-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm6[0],xmm2[0],xmm6[1],xmm2[1],xmm6[2],xmm2[2],xmm6[3],xmm2[3]
+; X86-AVX1-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm2[0],xmm5[0],xmm2[1],xmm5[1]
+; X86-AVX1-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm2[0],xmm4[0]
+; X86-AVX1-NEXT: vbroadcastss 12(%ecx), %xmm3
+; X86-AVX1-NEXT: vpextrw $0, %xmm3, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vbroadcastss 12(%eax), %xmm3
+; X86-AVX1-NEXT: vpextrw $0, %xmm3, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm3
+; X86-AVX1-NEXT: vpsrldq {{.*#+}} xmm4 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX1-NEXT: vpextrw $0, %xmm4, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vpsrldq {{.*#+}} xmm4 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX1-NEXT: vpextrw $0, %xmm4, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm4
+; X86-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
+; X86-AVX1-NEXT: vbroadcastss 8(%ecx), %xmm4
+; X86-AVX1-NEXT: vpextrw $0, %xmm4, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vbroadcastss 8(%eax), %xmm4
+; X86-AVX1-NEXT: vpextrw $0, %xmm4, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm4
+; X86-AVX1-NEXT: vpsrldq {{.*#+}} xmm5 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX1-NEXT: vpextrw $0, %xmm5, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vpsrldq {{.*#+}} xmm5 = xmm1[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX1-NEXT: vpextrw $0, %xmm5, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm5
+; X86-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
+; X86-AVX1-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]
+; X86-AVX1-NEXT: vbroadcastss 4(%ecx), %xmm4
+; X86-AVX1-NEXT: vpextrw $0, %xmm4, %ecx
+; X86-AVX1-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X86-AVX1-NEXT: vbroadcastss 4(%eax), %xmm4
+; X86-AVX1-NEXT: vpextrw $0, %xmm4, %eax
+; X86-AVX1-NEXT: andl $32767, %eax # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %ecx, %eax
+; X86-AVX1-NEXT: vpinsrw $0, %eax, %xmm0, %xmm4
+; X86-AVX1-NEXT: vpsrlq $48, %xmm0, %xmm5
+; X86-AVX1-NEXT: vpextrw $0, %xmm5, %eax
+; X86-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-AVX1-NEXT: vpsrlq $48, %xmm1, %xmm5
+; X86-AVX1-NEXT: vpextrw $0, %xmm5, %ecx
+; X86-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %eax, %ecx
+; X86-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm5
+; X86-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
+; X86-AVX1-NEXT: vpextrw $0, %xmm0, %eax
+; X86-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-AVX1-NEXT: vpextrw $0, %xmm1, %ecx
+; X86-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %eax, %ecx
+; X86-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm5
+; X86-AVX1-NEXT: vpsrld $16, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpextrw $0, %xmm0, %eax
+; X86-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-AVX1-NEXT: vpsrld $16, %xmm1, %xmm0
+; X86-AVX1-NEXT: vpextrw $0, %xmm0, %ecx
+; X86-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %eax, %ecx
+; X86-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm5[0],xmm0[0],xmm5[1],xmm0[1],xmm5[2],xmm0[2],xmm5[3],xmm0[3]
+; X86-AVX1-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1]
+; X86-AVX1-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm3[0]
+; X86-AVX1-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm0
+; X86-AVX1-NEXT: popl %esi
+; X86-AVX1-NEXT: retl
+;
+; X86-AVX2-LABEL: fcopysign_v16f16:
+; X86-AVX2: # %bb.0:
+; X86-AVX2-NEXT: pushl %edi
+; X86-AVX2-NEXT: pushl %esi
+; X86-AVX2-NEXT: movl {{[0-9]+}}(%esp), %eax
+; X86-AVX2-NEXT: movl {{[0-9]+}}(%esp), %ecx
+; X86-AVX2-NEXT: vmovdqa (%ecx), %xmm0
+; X86-AVX2-NEXT: vmovdqa 16(%ecx), %xmm2
+; X86-AVX2-NEXT: vpsrldq {{.*#+}} xmm1 = xmm2[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX2-NEXT: vpextrw $0, %xmm1, %edx
+; X86-AVX2-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX2-NEXT: vmovdqa (%eax), %xmm1
+; X86-AVX2-NEXT: vmovdqa 16(%eax), %xmm3
+; X86-AVX2-NEXT: vpsrldq {{.*#+}} xmm4 = xmm3[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX2-NEXT: vpextrw $0, %xmm4, %esi
+; X86-AVX2-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %edx, %esi
+; X86-AVX2-NEXT: vpinsrw $0, %esi, %xmm0, %xmm4
+; X86-AVX2-NEXT: movl $-32768, %edx # imm = 0x8000
+; X86-AVX2-NEXT: movl 28(%ecx), %esi
+; X86-AVX2-NEXT: andl %edx, %esi
+; X86-AVX2-NEXT: movzwl 28(%eax), %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm5
+; X86-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; X86-AVX2-NEXT: vpsrldq {{.*#+}} xmm5 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX2-NEXT: vpextrw $0, %xmm5, %esi
+; X86-AVX2-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX2-NEXT: vpsrldq {{.*#+}} xmm5 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX2-NEXT: vpextrw $0, %xmm5, %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm5
+; X86-AVX2-NEXT: movl 12(%ecx), %esi
+; X86-AVX2-NEXT: andl %edx, %esi
+; X86-AVX2-NEXT: movzwl 12(%eax), %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm6
+; X86-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; X86-AVX2-NEXT: vinserti128 $1, %xmm4, %ymm5, %ymm4
+; X86-AVX2-NEXT: vpsrldq {{.*#+}} xmm5 = xmm2[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX2-NEXT: vpextrw $0, %xmm5, %esi
+; X86-AVX2-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX2-NEXT: vpsrldq {{.*#+}} xmm5 = xmm3[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX2-NEXT: vpextrw $0, %xmm5, %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm5
+; X86-AVX2-NEXT: movl 24(%ecx), %esi
+; X86-AVX2-NEXT: andl %edx, %esi
+; X86-AVX2-NEXT: movzwl 24(%eax), %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm6
+; X86-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; X86-AVX2-NEXT: vpsrldq {{.*#+}} xmm6 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX2-NEXT: vpextrw $0, %xmm6, %esi
+; X86-AVX2-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX2-NEXT: vpsrldq {{.*#+}} xmm6 = xmm1[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX2-NEXT: vpextrw $0, %xmm6, %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm6
+; X86-AVX2-NEXT: movl 8(%ecx), %esi
+; X86-AVX2-NEXT: andl %edx, %esi
+; X86-AVX2-NEXT: movzwl 8(%eax), %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm7
+; X86-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
+; X86-AVX2-NEXT: vinserti128 $1, %xmm5, %ymm6, %ymm5
+; X86-AVX2-NEXT: vpunpckldq {{.*#+}} ymm4 = ymm5[0],ymm4[0],ymm5[1],ymm4[1],ymm5[4],ymm4[4],ymm5[5],ymm4[5]
+; X86-AVX2-NEXT: vpextrw $0, %xmm2, %esi
+; X86-AVX2-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX2-NEXT: vpextrw $0, %xmm3, %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm5
+; X86-AVX2-NEXT: vpsrld $16, %xmm2, %xmm6
+; X86-AVX2-NEXT: vpextrw $0, %xmm6, %esi
+; X86-AVX2-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX2-NEXT: vpsrld $16, %xmm3, %xmm6
+; X86-AVX2-NEXT: vpextrw $0, %xmm6, %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm6
+; X86-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
+; X86-AVX2-NEXT: vpextrw $0, %xmm0, %esi
+; X86-AVX2-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX2-NEXT: vpextrw $0, %xmm1, %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm6
+; X86-AVX2-NEXT: vpsrld $16, %xmm0, %xmm7
+; X86-AVX2-NEXT: vpextrw $0, %xmm7, %esi
+; X86-AVX2-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX2-NEXT: vpsrld $16, %xmm1, %xmm7
+; X86-AVX2-NEXT: vpextrw $0, %xmm7, %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm7
+; X86-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
+; X86-AVX2-NEXT: vinserti128 $1, %xmm5, %ymm6, %ymm5
+; X86-AVX2-NEXT: vpsrlq $48, %xmm2, %xmm2
+; X86-AVX2-NEXT: vpextrw $0, %xmm2, %esi
+; X86-AVX2-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX2-NEXT: vpsrlq $48, %xmm3, %xmm2
+; X86-AVX2-NEXT: vpextrw $0, %xmm2, %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm2
+; X86-AVX2-NEXT: movl 20(%ecx), %esi
+; X86-AVX2-NEXT: andl %edx, %esi
+; X86-AVX2-NEXT: movzwl 20(%eax), %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm3
+; X86-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; X86-AVX2-NEXT: vpsrlq $48, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpextrw $0, %xmm0, %esi
+; X86-AVX2-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX2-NEXT: vpsrlq $48, %xmm1, %xmm0
+; X86-AVX2-NEXT: vpextrw $0, %xmm0, %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm0
+; X86-AVX2-NEXT: andl 4(%ecx), %edx
+; X86-AVX2-NEXT: movzwl 4(%eax), %eax
+; X86-AVX2-NEXT: andl $32767, %eax # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %edx, %eax
+; X86-AVX2-NEXT: vpinsrw $0, %eax, %xmm0, %xmm1
+; X86-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; X86-AVX2-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm0
+; X86-AVX2-NEXT: vpunpckldq {{.*#+}} ymm0 = ymm5[0],ymm0[0],ymm5[1],ymm0[1],ymm5[4],ymm0[4],ymm5[5],ymm0[5]
+; X86-AVX2-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm4[0],ymm0[2],ymm4[2]
+; X86-AVX2-NEXT: popl %esi
+; X86-AVX2-NEXT: popl %edi
+; X86-AVX2-NEXT: retl
+;
+; X86-AVX512VL-LABEL: fcopysign_v16f16:
+; X86-AVX512VL: # %bb.0:
+; X86-AVX512VL-NEXT: pushl %edi
+; X86-AVX512VL-NEXT: pushl %esi
+; X86-AVX512VL-NEXT: movl {{[0-9]+}}(%esp), %eax
+; X86-AVX512VL-NEXT: movl {{[0-9]+}}(%esp), %ecx
+; X86-AVX512VL-NEXT: vmovdqa (%ecx), %xmm0
+; X86-AVX512VL-NEXT: vmovdqa 16(%ecx), %xmm2
+; X86-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm1 = xmm2[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm1, %edx
+; X86-AVX512VL-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX512VL-NEXT: vmovdqa (%eax), %xmm1
+; X86-AVX512VL-NEXT: vmovdqa 16(%eax), %xmm3
+; X86-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm4 = xmm3[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm4, %esi
+; X86-AVX512VL-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %edx, %esi
+; X86-AVX512VL-NEXT: vpinsrw $0, %esi, %xmm0, %xmm4
+; X86-AVX512VL-NEXT: movl $-32768, %edx # imm = 0x8000
+; X86-AVX512VL-NEXT: movl 28(%ecx), %esi
+; X86-AVX512VL-NEXT: andl %edx, %esi
+; X86-AVX512VL-NEXT: movzwl 28(%eax), %edi
+; X86-AVX512VL-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %esi, %edi
+; X86-AVX512VL-NEXT: vpinsrw $0, %edi, %xmm0, %xmm5
+; X86-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; X86-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm5 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm5, %esi
+; X86-AVX512VL-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm5 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm5, %edi
+; X86-AVX512VL-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %esi, %edi
+; X86-AVX512VL-NEXT: vpinsrw $0, %edi, %xmm0, %xmm5
+; X86-AVX512VL-NEXT: movl 12(%ecx), %esi
+; X86-AVX512VL-NEXT: andl %edx, %esi
+; X86-AVX512VL-NEXT: movzwl 12(%eax), %edi
+; X86-AVX512VL-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %esi, %edi
+; X86-AVX512VL-NEXT: vpinsrw $0, %edi, %xmm0, %xmm6
+; X86-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; X86-AVX512VL-NEXT: vinserti128 $1, %xmm4, %ymm5, %ymm4
+; X86-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm5 = xmm2[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm5, %esi
+; X86-AVX512VL-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm5 = xmm3[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm5, %edi
+; X86-AVX512VL-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %esi, %edi
+; X86-AVX512VL-NEXT: vpinsrw $0, %edi, %xmm0, %xmm5
+; X86-AVX512VL-NEXT: movl 24(%ecx), %esi
+; X86-AVX512VL-NEXT: andl %edx, %esi
+; X86-AVX512VL-NEXT: movzwl 24(%eax), %edi
+; X86-AVX512VL-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %esi, %edi
+; X86-AVX512VL-NEXT: vpinsrw $0, %edi, %xmm0, %xmm6
+; X86-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; X86-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm6 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm6, %esi
+; X86-AVX512VL-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm6 = xmm1[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm6, %edi
+; X86-AVX512VL-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %esi, %edi
+; X86-AVX512VL-NEXT: vpinsrw $0, %edi, %xmm0, %xmm6
+; X86-AVX512VL-NEXT: movl 8(%ecx), %esi
+; X86-AVX512VL-NEXT: andl %edx, %esi
+; X86-AVX512VL-NEXT: movzwl 8(%eax), %edi
+; X86-AVX512VL-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %esi, %edi
+; X86-AVX512VL-NEXT: vpinsrw $0, %edi, %xmm0, %xmm7
+; X86-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
+; X86-AVX512VL-NEXT: vinserti128 $1, %xmm5, %ymm6, %ymm5
+; X86-AVX512VL-NEXT: vpunpckldq {{.*#+}} ymm4 = ymm5[0],ymm4[0],ymm5[1],ymm4[1],ymm5[4],ymm4[4],ymm5[5],ymm4[5]
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm2, %esi
+; X86-AVX512VL-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm3, %edi
+; X86-AVX512VL-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %esi, %edi
+; X86-AVX512VL-NEXT: vpinsrw $0, %edi, %xmm0, %xmm5
+; X86-AVX512VL-NEXT: vpsrld $16, %xmm2, %xmm6
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm6, %esi
+; X86-AVX512VL-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX512VL-NEXT: vpsrld $16, %xmm3, %xmm6
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm6, %edi
+; X86-AVX512VL-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %esi, %edi
+; X86-AVX512VL-NEXT: vpinsrw $0, %edi, %xmm0, %xmm6
+; X86-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm0, %esi
+; X86-AVX512VL-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm1, %edi
+; X86-AVX512VL-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %esi, %edi
+; X86-AVX512VL-NEXT: vpinsrw $0, %edi, %xmm0, %xmm6
+; X86-AVX512VL-NEXT: vpsrld $16, %xmm0, %xmm7
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm7, %esi
+; X86-AVX512VL-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX512VL-NEXT: vpsrld $16, %xmm1, %xmm7
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm7, %edi
+; X86-AVX512VL-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %esi, %edi
+; X86-AVX512VL-NEXT: vpinsrw $0, %edi, %xmm0, %xmm7
+; X86-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
+; X86-AVX512VL-NEXT: vinserti128 $1, %xmm5, %ymm6, %ymm5
+; X86-AVX512VL-NEXT: vpsrlq $48, %xmm2, %xmm2
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm2, %esi
+; X86-AVX512VL-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX512VL-NEXT: vpsrlq $48, %xmm3, %xmm2
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm2, %edi
+; X86-AVX512VL-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %esi, %edi
+; X86-AVX512VL-NEXT: vpinsrw $0, %edi, %xmm0, %xmm2
+; X86-AVX512VL-NEXT: movl 20(%ecx), %esi
+; X86-AVX512VL-NEXT: andl %edx, %esi
+; X86-AVX512VL-NEXT: movzwl 20(%eax), %edi
+; X86-AVX512VL-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %esi, %edi
+; X86-AVX512VL-NEXT: vpinsrw $0, %edi, %xmm0, %xmm3
+; X86-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; X86-AVX512VL-NEXT: vpsrlq $48, %xmm0, %xmm0
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm0, %esi
+; X86-AVX512VL-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX512VL-NEXT: vpsrlq $48, %xmm1, %xmm0
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm0, %edi
+; X86-AVX512VL-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %esi, %edi
+; X86-AVX512VL-NEXT: vpinsrw $0, %edi, %xmm0, %xmm0
+; X86-AVX512VL-NEXT: andl 4(%ecx), %edx
+; X86-AVX512VL-NEXT: movzwl 4(%eax), %eax
+; X86-AVX512VL-NEXT: andl $32767, %eax # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %edx, %eax
+; X86-AVX512VL-NEXT: vpinsrw $0, %eax, %xmm0, %xmm1
+; X86-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; X86-AVX512VL-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm0
+; X86-AVX512VL-NEXT: vpunpckldq {{.*#+}} ymm0 = ymm5[0],ymm0[0],ymm5[1],ymm0[1],ymm5[4],ymm0[4],ymm5[5],ymm0[5]
+; X86-AVX512VL-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm4[0],ymm0[2],ymm4[2]
+; X86-AVX512VL-NEXT: popl %esi
+; X86-AVX512VL-NEXT: popl %edi
+; X86-AVX512VL-NEXT: retl
+;
+; X86-AVX512FP16-LABEL: fcopysign_v16f16:
+; X86-AVX512FP16: # %bb.0:
+; X86-AVX512FP16-NEXT: movl {{[0-9]+}}(%esp), %eax
+; X86-AVX512FP16-NEXT: movl {{[0-9]+}}(%esp), %ecx
+; X86-AVX512FP16-NEXT: vmovdqu (%ecx), %ymm1
+; X86-AVX512FP16-NEXT: vpbroadcastw {{.*#+}} ymm0 = [NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN]
+; X86-AVX512FP16-NEXT: vpternlogq $202, (%eax), %ymm1, %ymm0
+; X86-AVX512FP16-NEXT: retl
+;
+; X86-AVX512VLDQ-LABEL: fcopysign_v16f16:
+; X86-AVX512VLDQ: # %bb.0:
+; X86-AVX512VLDQ-NEXT: pushl %edi
+; X86-AVX512VLDQ-NEXT: pushl %esi
+; X86-AVX512VLDQ-NEXT: movl {{[0-9]+}}(%esp), %eax
+; X86-AVX512VLDQ-NEXT: movl {{[0-9]+}}(%esp), %ecx
+; X86-AVX512VLDQ-NEXT: vmovdqa (%ecx), %xmm0
+; X86-AVX512VLDQ-NEXT: vmovdqa 16(%ecx), %xmm2
+; X86-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm1 = xmm2[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm1, %edx
+; X86-AVX512VLDQ-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX512VLDQ-NEXT: vmovdqa (%eax), %xmm1
+; X86-AVX512VLDQ-NEXT: vmovdqa 16(%eax), %xmm3
+; X86-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm4 = xmm3[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm4, %esi
+; X86-AVX512VLDQ-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %edx, %esi
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %esi, %xmm0, %xmm4
+; X86-AVX512VLDQ-NEXT: movl $-32768, %edx # imm = 0x8000
+; X86-AVX512VLDQ-NEXT: movl 28(%ecx), %esi
+; X86-AVX512VLDQ-NEXT: andl %edx, %esi
+; X86-AVX512VLDQ-NEXT: movzwl 28(%eax), %edi
+; X86-AVX512VLDQ-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %esi, %edi
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %edi, %xmm0, %xmm5
+; X86-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; X86-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm5 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm5, %esi
+; X86-AVX512VLDQ-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm5 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm5, %edi
+; X86-AVX512VLDQ-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %esi, %edi
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %edi, %xmm0, %xmm5
+; X86-AVX512VLDQ-NEXT: movl 12(%ecx), %esi
+; X86-AVX512VLDQ-NEXT: andl %edx, %esi
+; X86-AVX512VLDQ-NEXT: movzwl 12(%eax), %edi
+; X86-AVX512VLDQ-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %esi, %edi
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %edi, %xmm0, %xmm6
+; X86-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; X86-AVX512VLDQ-NEXT: vinserti128 $1, %xmm4, %ymm5, %ymm4
+; X86-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm5 = xmm2[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm5, %esi
+; X86-AVX512VLDQ-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm5 = xmm3[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm5, %edi
+; X86-AVX512VLDQ-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %esi, %edi
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %edi, %xmm0, %xmm5
+; X86-AVX512VLDQ-NEXT: movl 24(%ecx), %esi
+; X86-AVX512VLDQ-NEXT: andl %edx, %esi
+; X86-AVX512VLDQ-NEXT: movzwl 24(%eax), %edi
+; X86-AVX512VLDQ-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %esi, %edi
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %edi, %xmm0, %xmm6
+; X86-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; X86-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm6 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm6, %esi
+; X86-AVX512VLDQ-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm6 = xmm1[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm6, %edi
+; X86-AVX512VLDQ-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %esi, %edi
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %edi, %xmm0, %xmm6
+; X86-AVX512VLDQ-NEXT: movl 8(%ecx), %esi
+; X86-AVX512VLDQ-NEXT: andl %edx, %esi
+; X86-AVX512VLDQ-NEXT: movzwl 8(%eax), %edi
+; X86-AVX512VLDQ-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %esi, %edi
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %edi, %xmm0, %xmm7
+; X86-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
+; X86-AVX512VLDQ-NEXT: vinserti128 $1, %xmm5, %ymm6, %ymm5
+; X86-AVX512VLDQ-NEXT: vpunpckldq {{.*#+}} ymm4 = ymm5[0],ymm4[0],ymm5[1],ymm4[1],ymm5[4],ymm4[4],ymm5[5],ymm4[5]
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm2, %esi
+; X86-AVX512VLDQ-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm3, %edi
+; X86-AVX512VLDQ-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %esi, %edi
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %edi, %xmm0, %xmm5
+; X86-AVX512VLDQ-NEXT: vpsrld $16, %xmm2, %xmm6
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm6, %esi
+; X86-AVX512VLDQ-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX512VLDQ-NEXT: vpsrld $16, %xmm3, %xmm6
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm6, %edi
+; X86-AVX512VLDQ-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %esi, %edi
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %edi, %xmm0, %xmm6
+; X86-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm0, %esi
+; X86-AVX512VLDQ-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm1, %edi
+; X86-AVX512VLDQ-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %esi, %edi
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %edi, %xmm0, %xmm6
+; X86-AVX512VLDQ-NEXT: vpsrld $16, %xmm0, %xmm7
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm7, %esi
+; X86-AVX512VLDQ-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX512VLDQ-NEXT: vpsrld $16, %xmm1, %xmm7
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm7, %edi
+; X86-AVX512VLDQ-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %esi, %edi
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %edi, %xmm0, %xmm7
+; X86-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
+; X86-AVX512VLDQ-NEXT: vinserti128 $1, %xmm5, %ymm6, %ymm5
+; X86-AVX512VLDQ-NEXT: vpsrlq $48, %xmm2, %xmm2
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm2, %esi
+; X86-AVX512VLDQ-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX512VLDQ-NEXT: vpsrlq $48, %xmm3, %xmm2
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm2, %edi
+; X86-AVX512VLDQ-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %esi, %edi
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %edi, %xmm0, %xmm2
+; X86-AVX512VLDQ-NEXT: movl 20(%ecx), %esi
+; X86-AVX512VLDQ-NEXT: andl %edx, %esi
+; X86-AVX512VLDQ-NEXT: movzwl 20(%eax), %edi
+; X86-AVX512VLDQ-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %esi, %edi
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %edi, %xmm0, %xmm3
+; X86-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; X86-AVX512VLDQ-NEXT: vpsrlq $48, %xmm0, %xmm0
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm0, %esi
+; X86-AVX512VLDQ-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX512VLDQ-NEXT: vpsrlq $48, %xmm1, %xmm0
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm0, %edi
+; X86-AVX512VLDQ-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %esi, %edi
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %edi, %xmm0, %xmm0
+; X86-AVX512VLDQ-NEXT: andl 4(%ecx), %edx
+; X86-AVX512VLDQ-NEXT: movzwl 4(%eax), %eax
+; X86-AVX512VLDQ-NEXT: andl $32767, %eax # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %edx, %eax
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %eax, %xmm0, %xmm1
+; X86-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; X86-AVX512VLDQ-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm0
+; X86-AVX512VLDQ-NEXT: vpunpckldq {{.*#+}} ymm0 = ymm5[0],ymm0[0],ymm5[1],ymm0[1],ymm5[4],ymm0[4],ymm5[5],ymm0[5]
+; X86-AVX512VLDQ-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm4[0],ymm0[2],ymm4[2]
+; X86-AVX512VLDQ-NEXT: popl %esi
+; X86-AVX512VLDQ-NEXT: popl %edi
+; X86-AVX512VLDQ-NEXT: retl
+;
+; X64-SSE-LABEL: fcopysign_v16f16:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: movdqa (%rdi), %xmm1
+; X64-SSE-NEXT: movdqa 16(%rdi), %xmm2
+; X64-SSE-NEXT: movdqa (%rsi), %xmm4
+; X64-SSE-NEXT: movdqa 16(%rsi), %xmm3
+; X64-SSE-NEXT: movdqa %xmm4, %xmm0
+; X64-SSE-NEXT: psrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-SSE-NEXT: pextrw $0, %xmm0, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movdqa %xmm1, %xmm0
+; X64-SSE-NEXT: psrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-SSE-NEXT: pextrw $0, %xmm0, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm5
+; X64-SSE-NEXT: movdqa %xmm4, %xmm0
+; X64-SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm4[3,3]
+; X64-SSE-NEXT: pextrw $0, %xmm0, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movdqa %xmm1, %xmm0
+; X64-SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm1[3,3]
+; X64-SSE-NEXT: pextrw $0, %xmm0, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm0
+; X64-SSE-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm5[0],xmm0[1],xmm5[1],xmm0[2],xmm5[2],xmm0[3],xmm5[3]
+; X64-SSE-NEXT: movdqa %xmm4, %xmm5
+; X64-SSE-NEXT: psrldq {{.*#+}} xmm5 = xmm5[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-SSE-NEXT: pextrw $0, %xmm5, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movdqa %xmm1, %xmm5
+; X64-SSE-NEXT: psrldq {{.*#+}} xmm5 = xmm5[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-SSE-NEXT: pextrw $0, %xmm5, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm6
+; X64-SSE-NEXT: movdqa %xmm4, %xmm5
+; X64-SSE-NEXT: punpckhqdq {{.*#+}} xmm5 = xmm5[1],xmm4[1]
+; X64-SSE-NEXT: pextrw $0, %xmm5, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movdqa %xmm1, %xmm5
+; X64-SSE-NEXT: punpckhqdq {{.*#+}} xmm5 = xmm5[1],xmm1[1]
+; X64-SSE-NEXT: pextrw $0, %xmm5, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm5
+; X64-SSE-NEXT: punpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
+; X64-SSE-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm0[0],xmm5[1],xmm0[1]
+; X64-SSE-NEXT: movdqa %xmm4, %xmm0
+; X64-SSE-NEXT: psrlq $48, %xmm0
+; X64-SSE-NEXT: pextrw $0, %xmm0, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movdqa %xmm1, %xmm0
+; X64-SSE-NEXT: psrlq $48, %xmm0
+; X64-SSE-NEXT: pextrw $0, %xmm0, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm0
+; X64-SSE-NEXT: movaps %xmm4, %xmm6
+; X64-SSE-NEXT: shufps {{.*#+}} xmm6 = xmm6[1,1],xmm4[1,1]
+; X64-SSE-NEXT: pextrw $0, %xmm6, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movaps %xmm1, %xmm6
+; X64-SSE-NEXT: shufps {{.*#+}} xmm6 = xmm6[1,1],xmm1[1,1]
+; X64-SSE-NEXT: pextrw $0, %xmm6, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm6
+; X64-SSE-NEXT: punpcklwd {{.*#+}} xmm6 = xmm6[0],xmm0[0],xmm6[1],xmm0[1],xmm6[2],xmm0[2],xmm6[3],xmm0[3]
+; X64-SSE-NEXT: pextrw $0, %xmm4, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: pextrw $0, %xmm1, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm0
+; X64-SSE-NEXT: psrld $16, %xmm4
+; X64-SSE-NEXT: pextrw $0, %xmm4, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: psrld $16, %xmm1
+; X64-SSE-NEXT: pextrw $0, %xmm1, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm1
+; X64-SSE-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; X64-SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm6[0],xmm0[1],xmm6[1]
+; X64-SSE-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm5[0]
+; X64-SSE-NEXT: movdqa %xmm3, %xmm1
+; X64-SSE-NEXT: psrldq {{.*#+}} xmm1 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-SSE-NEXT: pextrw $0, %xmm1, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movdqa %xmm2, %xmm1
+; X64-SSE-NEXT: psrldq {{.*#+}} xmm1 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-SSE-NEXT: pextrw $0, %xmm1, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm4
+; X64-SSE-NEXT: movdqa %xmm3, %xmm1
+; X64-SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[3,3],xmm3[3,3]
+; X64-SSE-NEXT: pextrw $0, %xmm1, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movdqa %xmm2, %xmm1
+; X64-SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[3,3],xmm2[3,3]
+; X64-SSE-NEXT: pextrw $0, %xmm1, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm1
+; X64-SSE-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3]
+; X64-SSE-NEXT: movdqa %xmm3, %xmm4
+; X64-SSE-NEXT: psrldq {{.*#+}} xmm4 = xmm4[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-SSE-NEXT: pextrw $0, %xmm4, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movdqa %xmm2, %xmm4
+; X64-SSE-NEXT: psrldq {{.*#+}} xmm4 = xmm4[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-SSE-NEXT: pextrw $0, %xmm4, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm5
+; X64-SSE-NEXT: movdqa %xmm3, %xmm4
+; X64-SSE-NEXT: punpckhqdq {{.*#+}} xmm4 = xmm4[1],xmm3[1]
+; X64-SSE-NEXT: pextrw $0, %xmm4, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movdqa %xmm2, %xmm4
+; X64-SSE-NEXT: punpckhqdq {{.*#+}} xmm4 = xmm4[1],xmm2[1]
+; X64-SSE-NEXT: pextrw $0, %xmm4, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm4
+; X64-SSE-NEXT: punpcklwd {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
+; X64-SSE-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm1[0],xmm4[1],xmm1[1]
+; X64-SSE-NEXT: movdqa %xmm3, %xmm1
+; X64-SSE-NEXT: psrlq $48, %xmm1
+; X64-SSE-NEXT: pextrw $0, %xmm1, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movdqa %xmm2, %xmm1
+; X64-SSE-NEXT: psrlq $48, %xmm1
+; X64-SSE-NEXT: pextrw $0, %xmm1, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm1
+; X64-SSE-NEXT: movaps %xmm3, %xmm5
+; X64-SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[1,1],xmm3[1,1]
+; X64-SSE-NEXT: pextrw $0, %xmm5, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movaps %xmm2, %xmm5
+; X64-SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[1,1],xmm2[1,1]
+; X64-SSE-NEXT: pextrw $0, %xmm5, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm5
+; X64-SSE-NEXT: punpcklwd {{.*#+}} xmm5 = xmm5[0],xmm1[0],xmm5[1],xmm1[1],xmm5[2],xmm1[2],xmm5[3],xmm1[3]
+; X64-SSE-NEXT: pextrw $0, %xmm3, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: pextrw $0, %xmm2, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm1
+; X64-SSE-NEXT: psrld $16, %xmm3
+; X64-SSE-NEXT: pextrw $0, %xmm3, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: psrld $16, %xmm2
+; X64-SSE-NEXT: pextrw $0, %xmm2, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm2
+; X64-SSE-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
+; X64-SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm5[0],xmm1[1],xmm5[1]
+; X64-SSE-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm4[0]
+; X64-SSE-NEXT: retq
+;
+; X64-AVX1-LABEL: fcopysign_v16f16:
+; X64-AVX1: # %bb.0:
+; X64-AVX1-NEXT: vbroadcastss 28(%rsi), %xmm0
+; X64-AVX1-NEXT: vpextrw $0, %xmm0, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vbroadcastss 28(%rdi), %xmm0
+; X64-AVX1-NEXT: vpextrw $0, %xmm0, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm4
+; X64-AVX1-NEXT: vmovdqa (%rsi), %xmm0
+; X64-AVX1-NEXT: vmovdqa 16(%rsi), %xmm2
+; X64-AVX1-NEXT: vpsrldq {{.*#+}} xmm1 = xmm2[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX1-NEXT: vpextrw $0, %xmm1, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vmovdqa (%rdi), %xmm1
+; X64-AVX1-NEXT: vmovdqa 16(%rdi), %xmm3
+; X64-AVX1-NEXT: vpsrldq {{.*#+}} xmm5 = xmm3[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX1-NEXT: vpextrw $0, %xmm5, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm5
+; X64-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
+; X64-AVX1-NEXT: vbroadcastss 24(%rsi), %xmm5
+; X64-AVX1-NEXT: vpextrw $0, %xmm5, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vbroadcastss 24(%rdi), %xmm5
+; X64-AVX1-NEXT: vpextrw $0, %xmm5, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm5
+; X64-AVX1-NEXT: vpsrldq {{.*#+}} xmm6 = xmm2[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX1-NEXT: vpextrw $0, %xmm6, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vpsrldq {{.*#+}} xmm6 = xmm3[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX1-NEXT: vpextrw $0, %xmm6, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm6
+; X64-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
+; X64-AVX1-NEXT: vpunpckldq {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]
+; X64-AVX1-NEXT: vbroadcastss 20(%rsi), %xmm5
+; X64-AVX1-NEXT: vpextrw $0, %xmm5, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vbroadcastss 20(%rdi), %xmm5
+; X64-AVX1-NEXT: vpextrw $0, %xmm5, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm5
+; X64-AVX1-NEXT: vpsrlq $48, %xmm2, %xmm6
+; X64-AVX1-NEXT: vpextrw $0, %xmm6, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vpsrlq $48, %xmm3, %xmm6
+; X64-AVX1-NEXT: vpextrw $0, %xmm6, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm6
+; X64-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
+; X64-AVX1-NEXT: vpextrw $0, %xmm2, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vpextrw $0, %xmm3, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm6
+; X64-AVX1-NEXT: vpsrld $16, %xmm2, %xmm2
+; X64-AVX1-NEXT: vpextrw $0, %xmm2, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vpsrld $16, %xmm3, %xmm2
+; X64-AVX1-NEXT: vpextrw $0, %xmm2, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm2
+; X64-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm6[0],xmm2[0],xmm6[1],xmm2[1],xmm6[2],xmm2[2],xmm6[3],xmm2[3]
+; X64-AVX1-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm2[0],xmm5[0],xmm2[1],xmm5[1]
+; X64-AVX1-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm2[0],xmm4[0]
+; X64-AVX1-NEXT: vbroadcastss 12(%rsi), %xmm3
+; X64-AVX1-NEXT: vpextrw $0, %xmm3, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vbroadcastss 12(%rdi), %xmm3
+; X64-AVX1-NEXT: vpextrw $0, %xmm3, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm3
+; X64-AVX1-NEXT: vpsrldq {{.*#+}} xmm4 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX1-NEXT: vpextrw $0, %xmm4, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vpsrldq {{.*#+}} xmm4 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX1-NEXT: vpextrw $0, %xmm4, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm4
+; X64-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
+; X64-AVX1-NEXT: vbroadcastss 8(%rsi), %xmm4
+; X64-AVX1-NEXT: vpextrw $0, %xmm4, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vbroadcastss 8(%rdi), %xmm4
+; X64-AVX1-NEXT: vpextrw $0, %xmm4, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm4
+; X64-AVX1-NEXT: vpsrldq {{.*#+}} xmm5 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX1-NEXT: vpextrw $0, %xmm5, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vpsrldq {{.*#+}} xmm5 = xmm1[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX1-NEXT: vpextrw $0, %xmm5, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm5
+; X64-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
+; X64-AVX1-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]
+; X64-AVX1-NEXT: vbroadcastss 4(%rsi), %xmm4
+; X64-AVX1-NEXT: vpextrw $0, %xmm4, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vbroadcastss 4(%rdi), %xmm4
+; X64-AVX1-NEXT: vpextrw $0, %xmm4, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm4
+; X64-AVX1-NEXT: vpsrlq $48, %xmm0, %xmm5
+; X64-AVX1-NEXT: vpextrw $0, %xmm5, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vpsrlq $48, %xmm1, %xmm5
+; X64-AVX1-NEXT: vpextrw $0, %xmm5, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm5
+; X64-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
+; X64-AVX1-NEXT: vpextrw $0, %xmm0, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vpextrw $0, %xmm1, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm5
+; X64-AVX1-NEXT: vpsrld $16, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpextrw $0, %xmm0, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vpsrld $16, %xmm1, %xmm0
+; X64-AVX1-NEXT: vpextrw $0, %xmm0, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm5[0],xmm0[0],xmm5[1],xmm0[1],xmm5[2],xmm0[2],xmm5[3],xmm0[3]
+; X64-AVX1-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1]
+; X64-AVX1-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm3[0]
+; X64-AVX1-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm0
+; X64-AVX1-NEXT: retq
+;
+; X64-AVX2-LABEL: fcopysign_v16f16:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: vmovdqa (%rsi), %xmm0
+; X64-AVX2-NEXT: vmovdqa 16(%rsi), %xmm2
+; X64-AVX2-NEXT: vpsrldq {{.*#+}} xmm1 = xmm2[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX2-NEXT: vpextrw $0, %xmm1, %eax
+; X64-AVX2-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX2-NEXT: vmovdqa (%rdi), %xmm1
+; X64-AVX2-NEXT: vmovdqa 16(%rdi), %xmm3
+; X64-AVX2-NEXT: vpsrldq {{.*#+}} xmm4 = xmm3[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX2-NEXT: vpextrw $0, %xmm4, %ecx
+; X64-AVX2-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %eax, %ecx
+; X64-AVX2-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm4
+; X64-AVX2-NEXT: movl $-32768, %eax # imm = 0x8000
+; X64-AVX2-NEXT: movl 28(%rsi), %ecx
+; X64-AVX2-NEXT: andl %eax, %ecx
+; X64-AVX2-NEXT: movzwl 28(%rdi), %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm5
+; X64-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; X64-AVX2-NEXT: vpsrldq {{.*#+}} xmm5 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX2-NEXT: vpextrw $0, %xmm5, %ecx
+; X64-AVX2-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX2-NEXT: vpsrldq {{.*#+}} xmm5 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX2-NEXT: vpextrw $0, %xmm5, %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm5
+; X64-AVX2-NEXT: movl 12(%rsi), %ecx
+; X64-AVX2-NEXT: andl %eax, %ecx
+; X64-AVX2-NEXT: movzwl 12(%rdi), %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm6
+; X64-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; X64-AVX2-NEXT: vinserti128 $1, %xmm4, %ymm5, %ymm4
+; X64-AVX2-NEXT: vpsrldq {{.*#+}} xmm5 = xmm2[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX2-NEXT: vpextrw $0, %xmm5, %ecx
+; X64-AVX2-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX2-NEXT: vpsrldq {{.*#+}} xmm5 = xmm3[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX2-NEXT: vpextrw $0, %xmm5, %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm5
+; X64-AVX2-NEXT: movl 24(%rsi), %ecx
+; X64-AVX2-NEXT: andl %eax, %ecx
+; X64-AVX2-NEXT: movzwl 24(%rdi), %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm6
+; X64-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; X64-AVX2-NEXT: vpsrldq {{.*#+}} xmm6 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX2-NEXT: vpextrw $0, %xmm6, %ecx
+; X64-AVX2-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX2-NEXT: vpsrldq {{.*#+}} xmm6 = xmm1[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX2-NEXT: vpextrw $0, %xmm6, %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm6
+; X64-AVX2-NEXT: movl 8(%rsi), %ecx
+; X64-AVX2-NEXT: andl %eax, %ecx
+; X64-AVX2-NEXT: movzwl 8(%rdi), %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm7
+; X64-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
+; X64-AVX2-NEXT: vinserti128 $1, %xmm5, %ymm6, %ymm5
+; X64-AVX2-NEXT: vpunpckldq {{.*#+}} ymm4 = ymm5[0],ymm4[0],ymm5[1],ymm4[1],ymm5[4],ymm4[4],ymm5[5],ymm4[5]
+; X64-AVX2-NEXT: vpextrw $0, %xmm2, %ecx
+; X64-AVX2-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX2-NEXT: vpextrw $0, %xmm3, %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm5
+; X64-AVX2-NEXT: vpsrld $16, %xmm2, %xmm6
+; X64-AVX2-NEXT: vpextrw $0, %xmm6, %ecx
+; X64-AVX2-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX2-NEXT: vpsrld $16, %xmm3, %xmm6
+; X64-AVX2-NEXT: vpextrw $0, %xmm6, %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm6
+; X64-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
+; X64-AVX2-NEXT: vpextrw $0, %xmm0, %ecx
+; X64-AVX2-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX2-NEXT: vpextrw $0, %xmm1, %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm6
+; X64-AVX2-NEXT: vpsrld $16, %xmm0, %xmm7
+; X64-AVX2-NEXT: vpextrw $0, %xmm7, %ecx
+; X64-AVX2-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX2-NEXT: vpsrld $16, %xmm1, %xmm7
+; X64-AVX2-NEXT: vpextrw $0, %xmm7, %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm7
+; X64-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
+; X64-AVX2-NEXT: vinserti128 $1, %xmm5, %ymm6, %ymm5
+; X64-AVX2-NEXT: vpsrlq $48, %xmm2, %xmm2
+; X64-AVX2-NEXT: vpextrw $0, %xmm2, %ecx
+; X64-AVX2-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX2-NEXT: vpsrlq $48, %xmm3, %xmm2
+; X64-AVX2-NEXT: vpextrw $0, %xmm2, %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm2
+; X64-AVX2-NEXT: movl 20(%rsi), %ecx
+; X64-AVX2-NEXT: andl %eax, %ecx
+; X64-AVX2-NEXT: movzwl 20(%rdi), %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm3
+; X64-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; X64-AVX2-NEXT: vpsrlq $48, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpextrw $0, %xmm0, %ecx
+; X64-AVX2-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX2-NEXT: vpsrlq $48, %xmm1, %xmm0
+; X64-AVX2-NEXT: vpextrw $0, %xmm0, %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm0
+; X64-AVX2-NEXT: andl 4(%rsi), %eax
+; X64-AVX2-NEXT: movzwl 4(%rdi), %ecx
+; X64-AVX2-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %eax, %ecx
+; X64-AVX2-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm1
+; X64-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; X64-AVX2-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm0
+; X64-AVX2-NEXT: vpunpckldq {{.*#+}} ymm0 = ymm5[0],ymm0[0],ymm5[1],ymm0[1],ymm5[4],ymm0[4],ymm5[5],ymm0[5]
+; X64-AVX2-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm4[0],ymm0[2],ymm4[2]
+; X64-AVX2-NEXT: retq
+;
+; X64-AVX512VL-LABEL: fcopysign_v16f16:
+; X64-AVX512VL: # %bb.0:
+; X64-AVX512VL-NEXT: vmovdqa (%rsi), %xmm0
+; X64-AVX512VL-NEXT: vmovdqa 16(%rsi), %xmm2
+; X64-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm1 = xmm2[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm1, %eax
+; X64-AVX512VL-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX512VL-NEXT: vmovdqa (%rdi), %xmm1
+; X64-AVX512VL-NEXT: vmovdqa 16(%rdi), %xmm3
+; X64-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm4 = xmm3[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm4, %ecx
+; X64-AVX512VL-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %eax, %ecx
+; X64-AVX512VL-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm4
+; X64-AVX512VL-NEXT: movl $-32768, %eax # imm = 0x8000
+; X64-AVX512VL-NEXT: movl 28(%rsi), %ecx
+; X64-AVX512VL-NEXT: andl %eax, %ecx
+; X64-AVX512VL-NEXT: movzwl 28(%rdi), %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm5
+; X64-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; X64-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm5 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm5, %ecx
+; X64-AVX512VL-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm5 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm5, %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm5
+; X64-AVX512VL-NEXT: movl 12(%rsi), %ecx
+; X64-AVX512VL-NEXT: andl %eax, %ecx
+; X64-AVX512VL-NEXT: movzwl 12(%rdi), %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm6
+; X64-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; X64-AVX512VL-NEXT: vinserti128 $1, %xmm4, %ymm5, %ymm4
+; X64-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm5 = xmm2[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm5, %ecx
+; X64-AVX512VL-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm5 = xmm3[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm5, %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm5
+; X64-AVX512VL-NEXT: movl 24(%rsi), %ecx
+; X64-AVX512VL-NEXT: andl %eax, %ecx
+; X64-AVX512VL-NEXT: movzwl 24(%rdi), %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm6
+; X64-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; X64-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm6 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm6, %ecx
+; X64-AVX512VL-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm6 = xmm1[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm6, %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm6
+; X64-AVX512VL-NEXT: movl 8(%rsi), %ecx
+; X64-AVX512VL-NEXT: andl %eax, %ecx
+; X64-AVX512VL-NEXT: movzwl 8(%rdi), %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm7
+; X64-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
+; X64-AVX512VL-NEXT: vinserti128 $1, %xmm5, %ymm6, %ymm5
+; X64-AVX512VL-NEXT: vpunpckldq {{.*#+}} ymm4 = ymm5[0],ymm4[0],ymm5[1],ymm4[1],ymm5[4],ymm4[4],ymm5[5],ymm4[5]
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm2, %ecx
+; X64-AVX512VL-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm3, %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm5
+; X64-AVX512VL-NEXT: vpsrld $16, %xmm2, %xmm6
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm6, %ecx
+; X64-AVX512VL-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VL-NEXT: vpsrld $16, %xmm3, %xmm6
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm6, %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm6
+; X64-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm0, %ecx
+; X64-AVX512VL-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm1, %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm6
+; X64-AVX512VL-NEXT: vpsrld $16, %xmm0, %xmm7
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm7, %ecx
+; X64-AVX512VL-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VL-NEXT: vpsrld $16, %xmm1, %xmm7
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm7, %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm7
+; X64-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
+; X64-AVX512VL-NEXT: vinserti128 $1, %xmm5, %ymm6, %ymm5
+; X64-AVX512VL-NEXT: vpsrlq $48, %xmm2, %xmm2
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm2, %ecx
+; X64-AVX512VL-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VL-NEXT: vpsrlq $48, %xmm3, %xmm2
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm2, %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm2
+; X64-AVX512VL-NEXT: movl 20(%rsi), %ecx
+; X64-AVX512VL-NEXT: andl %eax, %ecx
+; X64-AVX512VL-NEXT: movzwl 20(%rdi), %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm3
+; X64-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; X64-AVX512VL-NEXT: vpsrlq $48, %xmm0, %xmm0
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm0, %ecx
+; X64-AVX512VL-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VL-NEXT: vpsrlq $48, %xmm1, %xmm0
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm0, %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm0
+; X64-AVX512VL-NEXT: andl 4(%rsi), %eax
+; X64-AVX512VL-NEXT: movzwl 4(%rdi), %ecx
+; X64-AVX512VL-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %eax, %ecx
+; X64-AVX512VL-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm1
+; X64-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; X64-AVX512VL-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm0
+; X64-AVX512VL-NEXT: vpunpckldq {{.*#+}} ymm0 = ymm5[0],ymm0[0],ymm5[1],ymm0[1],ymm5[4],ymm0[4],ymm5[5],ymm0[5]
+; X64-AVX512VL-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm4[0],ymm0[2],ymm4[2]
+; X64-AVX512VL-NEXT: retq
+;
+; X64-AVX512FP16-LABEL: fcopysign_v16f16:
+; X64-AVX512FP16: # %bb.0:
+; X64-AVX512FP16-NEXT: vmovdqu (%rdi), %ymm1
+; X64-AVX512FP16-NEXT: vpbroadcastq {{.*#+}} ymm0 = [9223231297218904063,9223231297218904063,9223231297218904063,9223231297218904063]
+; X64-AVX512FP16-NEXT: vpternlogq $202, (%rsi), %ymm1, %ymm0
+; X64-AVX512FP16-NEXT: retq
+;
+; X64-AVX512VLDQ-LABEL: fcopysign_v16f16:
+; X64-AVX512VLDQ: # %bb.0:
+; X64-AVX512VLDQ-NEXT: vmovdqa (%rsi), %xmm0
+; X64-AVX512VLDQ-NEXT: vmovdqa 16(%rsi), %xmm2
+; X64-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm1 = xmm2[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm1, %eax
+; X64-AVX512VLDQ-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX512VLDQ-NEXT: vmovdqa (%rdi), %xmm1
+; X64-AVX512VLDQ-NEXT: vmovdqa 16(%rdi), %xmm3
+; X64-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm4 = xmm3[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm4, %ecx
+; X64-AVX512VLDQ-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %eax, %ecx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm4
+; X64-AVX512VLDQ-NEXT: movl $-32768, %eax # imm = 0x8000
+; X64-AVX512VLDQ-NEXT: movl 28(%rsi), %ecx
+; X64-AVX512VLDQ-NEXT: andl %eax, %ecx
+; X64-AVX512VLDQ-NEXT: movzwl 28(%rdi), %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm5
+; X64-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; X64-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm5 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm5, %ecx
+; X64-AVX512VLDQ-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm5 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm5, %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm5
+; X64-AVX512VLDQ-NEXT: movl 12(%rsi), %ecx
+; X64-AVX512VLDQ-NEXT: andl %eax, %ecx
+; X64-AVX512VLDQ-NEXT: movzwl 12(%rdi), %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm6
+; X64-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; X64-AVX512VLDQ-NEXT: vinserti128 $1, %xmm4, %ymm5, %ymm4
+; X64-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm5 = xmm2[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm5, %ecx
+; X64-AVX512VLDQ-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm5 = xmm3[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm5, %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm5
+; X64-AVX512VLDQ-NEXT: movl 24(%rsi), %ecx
+; X64-AVX512VLDQ-NEXT: andl %eax, %ecx
+; X64-AVX512VLDQ-NEXT: movzwl 24(%rdi), %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm6
+; X64-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; X64-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm6 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm6, %ecx
+; X64-AVX512VLDQ-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm6 = xmm1[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm6, %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm6
+; X64-AVX512VLDQ-NEXT: movl 8(%rsi), %ecx
+; X64-AVX512VLDQ-NEXT: andl %eax, %ecx
+; X64-AVX512VLDQ-NEXT: movzwl 8(%rdi), %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm7
+; X64-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
+; X64-AVX512VLDQ-NEXT: vinserti128 $1, %xmm5, %ymm6, %ymm5
+; X64-AVX512VLDQ-NEXT: vpunpckldq {{.*#+}} ymm4 = ymm5[0],ymm4[0],ymm5[1],ymm4[1],ymm5[4],ymm4[4],ymm5[5],ymm4[5]
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm2, %ecx
+; X64-AVX512VLDQ-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm3, %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm5
+; X64-AVX512VLDQ-NEXT: vpsrld $16, %xmm2, %xmm6
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm6, %ecx
+; X64-AVX512VLDQ-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VLDQ-NEXT: vpsrld $16, %xmm3, %xmm6
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm6, %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm6
+; X64-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm0, %ecx
+; X64-AVX512VLDQ-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm1, %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm6
+; X64-AVX512VLDQ-NEXT: vpsrld $16, %xmm0, %xmm7
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm7, %ecx
+; X64-AVX512VLDQ-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VLDQ-NEXT: vpsrld $16, %xmm1, %xmm7
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm7, %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm7
+; X64-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
+; X64-AVX512VLDQ-NEXT: vinserti128 $1, %xmm5, %ymm6, %ymm5
+; X64-AVX512VLDQ-NEXT: vpsrlq $48, %xmm2, %xmm2
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm2, %ecx
+; X64-AVX512VLDQ-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VLDQ-NEXT: vpsrlq $48, %xmm3, %xmm2
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm2, %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm2
+; X64-AVX512VLDQ-NEXT: movl 20(%rsi), %ecx
+; X64-AVX512VLDQ-NEXT: andl %eax, %ecx
+; X64-AVX512VLDQ-NEXT: movzwl 20(%rdi), %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm3
+; X64-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; X64-AVX512VLDQ-NEXT: vpsrlq $48, %xmm0, %xmm0
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm0, %ecx
+; X64-AVX512VLDQ-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VLDQ-NEXT: vpsrlq $48, %xmm1, %xmm0
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm0, %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm0
+; X64-AVX512VLDQ-NEXT: andl 4(%rsi), %eax
+; X64-AVX512VLDQ-NEXT: movzwl 4(%rdi), %ecx
+; X64-AVX512VLDQ-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %eax, %ecx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm1
+; X64-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; X64-AVX512VLDQ-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm0
+; X64-AVX512VLDQ-NEXT: vpunpckldq {{.*#+}} ymm0 = ymm5[0],ymm0[0],ymm5[1],ymm0[1],ymm5[4],ymm0[4],ymm5[5],ymm0[5]
+; X64-AVX512VLDQ-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm4[0],ymm0[2],ymm4[2]
+; X64-AVX512VLDQ-NEXT: retq
+ %a0 = load <16 x half>, ptr %p0, align 16
+ %a1 = load <16 x half>, ptr %p1, align 16
+ %t = call <16 x half> @llvm.copysign.v16f16(<16 x half> %a0, <16 x half> %a1)
+ ret <16 x half> %t
+}
+declare <16 x half> @llvm.copysign.v16f16(<16 x half>, <16 x half>)
+
+;
+; 512-bit Vectors
+;
+
+define <8 x double> @fcopysign_v8f64(<8 x double> %a0, <8 x double> %a1) nounwind {
+; X86-SSE-LABEL: fcopysign_v8f64:
+; X86-SSE: # %bb.0:
+; X86-SSE-NEXT: pushl %ebp
+; X86-SSE-NEXT: movl %esp, %ebp
+; X86-SSE-NEXT: andl $-16, %esp
+; X86-SSE-NEXT: subl $16, %esp
+; X86-SSE-NEXT: movaps {{.*#+}} xmm3 = [NaN,NaN]
+; X86-SSE-NEXT: andps %xmm3, %xmm0
+; X86-SSE-NEXT: movaps %xmm3, %xmm4
+; X86-SSE-NEXT: andnps 24(%ebp), %xmm4
+; X86-SSE-NEXT: orps %xmm4, %xmm0
+; X86-SSE-NEXT: andps %xmm3, %xmm1
+; X86-SSE-NEXT: movaps %xmm3, %xmm4
+; X86-SSE-NEXT: andnps 40(%ebp), %xmm4
+; X86-SSE-NEXT: orps %xmm4, %xmm1
+; X86-SSE-NEXT: andps %xmm3, %xmm2
+; X86-SSE-NEXT: movaps %xmm3, %xmm4
+; X86-SSE-NEXT: andnps 56(%ebp), %xmm4
+; X86-SSE-NEXT: orps %xmm4, %xmm2
+; X86-SSE-NEXT: movaps %xmm3, %xmm4
+; X86-SSE-NEXT: andnps 72(%ebp), %xmm4
+; X86-SSE-NEXT: andps 8(%ebp), %xmm3
+; X86-SSE-NEXT: orps %xmm4, %xmm3
+; X86-SSE-NEXT: movl %ebp, %esp
+; X86-SSE-NEXT: popl %ebp
+; X86-SSE-NEXT: retl
+;
+; X86-AVX1OR2-LABEL: fcopysign_v8f64:
+; X86-AVX1OR2: # %bb.0:
+; X86-AVX1OR2-NEXT: pushl %ebp
+; X86-AVX1OR2-NEXT: movl %esp, %ebp
+; X86-AVX1OR2-NEXT: andl $-32, %esp
+; X86-AVX1OR2-NEXT: subl $32, %esp
+; X86-AVX1OR2-NEXT: vbroadcastsd {{.*#+}} ymm3 = [NaN,NaN,NaN,NaN]
+; X86-AVX1OR2-NEXT: vandnps %ymm2, %ymm3, %ymm2
+; X86-AVX1OR2-NEXT: vandps %ymm3, %ymm0, %ymm0
+; X86-AVX1OR2-NEXT: vorps %ymm2, %ymm0, %ymm0
+; X86-AVX1OR2-NEXT: vandps %ymm3, %ymm1, %ymm1
+; X86-AVX1OR2-NEXT: vandnps 8(%ebp), %ymm3, %ymm2
+; X86-AVX1OR2-NEXT: vorps %ymm2, %ymm1, %ymm1
+; X86-AVX1OR2-NEXT: movl %ebp, %esp
+; X86-AVX1OR2-NEXT: popl %ebp
+; X86-AVX1OR2-NEXT: retl
+;
+; X86-AVX512-LABEL: fcopysign_v8f64:
+; X86-AVX512: # %bb.0:
+; X86-AVX512-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}{1to8}, %zmm1, %zmm0
+; X86-AVX512-NEXT: retl
+;
+; X64-SSE-LABEL: fcopysign_v8f64:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: movaps {{.*#+}} xmm8 = [NaN,NaN]
+; X64-SSE-NEXT: movaps %xmm8, %xmm9
+; X64-SSE-NEXT: andnps %xmm4, %xmm9
+; X64-SSE-NEXT: andps %xmm8, %xmm0
+; X64-SSE-NEXT: orps %xmm9, %xmm0
+; X64-SSE-NEXT: movaps %xmm8, %xmm4
+; X64-SSE-NEXT: andnps %xmm5, %xmm4
+; X64-SSE-NEXT: andps %xmm8, %xmm1
+; X64-SSE-NEXT: orps %xmm4, %xmm1
+; X64-SSE-NEXT: movaps %xmm8, %xmm4
+; X64-SSE-NEXT: andnps %xmm6, %xmm4
+; X64-SSE-NEXT: andps %xmm8, %xmm2
+; X64-SSE-NEXT: orps %xmm4, %xmm2
+; X64-SSE-NEXT: andps %xmm8, %xmm3
+; X64-SSE-NEXT: andnps %xmm7, %xmm8
+; X64-SSE-NEXT: orps %xmm8, %xmm3
+; X64-SSE-NEXT: retq
+;
+; X64-AVX1OR2-LABEL: fcopysign_v8f64:
+; X64-AVX1OR2: # %bb.0:
+; X64-AVX1OR2-NEXT: vbroadcastsd {{.*#+}} ymm4 = [NaN,NaN,NaN,NaN]
+; X64-AVX1OR2-NEXT: vandnps %ymm2, %ymm4, %ymm2
+; X64-AVX1OR2-NEXT: vandps %ymm4, %ymm0, %ymm0
+; X64-AVX1OR2-NEXT: vorps %ymm2, %ymm0, %ymm0
+; X64-AVX1OR2-NEXT: vandnps %ymm3, %ymm4, %ymm2
+; X64-AVX1OR2-NEXT: vandps %ymm4, %ymm1, %ymm1
+; X64-AVX1OR2-NEXT: vorps %ymm2, %ymm1, %ymm1
+; X64-AVX1OR2-NEXT: retq
+;
+; X64-AVX512-LABEL: fcopysign_v8f64:
+; X64-AVX512: # %bb.0:
+; X64-AVX512-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to8}, %zmm1, %zmm0
+; X64-AVX512-NEXT: retq
+ %t = call <8 x double> @llvm.copysign.v8f64(<8 x double> %a0, <8 x double> %a1)
+ ret <8 x double> %t
+}
+declare <8 x double> @llvm.copysign.v8f64(<8 x double>, <8 x double>)
+
+define <16 x float> @fcopysign_v16f32(<16 x float> %a0, <16 x float> %a1) nounwind {
+; X86-SSE-LABEL: fcopysign_v16f32:
+; X86-SSE: # %bb.0:
+; X86-SSE-NEXT: pushl %ebp
+; X86-SSE-NEXT: movl %esp, %ebp
+; X86-SSE-NEXT: andl $-16, %esp
+; X86-SSE-NEXT: subl $16, %esp
+; X86-SSE-NEXT: movaps {{.*#+}} xmm3 = [NaN,NaN,NaN,NaN]
+; X86-SSE-NEXT: andps %xmm3, %xmm0
+; X86-SSE-NEXT: movaps %xmm3, %xmm4
+; X86-SSE-NEXT: andnps 24(%ebp), %xmm4
+; X86-SSE-NEXT: orps %xmm4, %xmm0
+; X86-SSE-NEXT: andps %xmm3, %xmm1
+; X86-SSE-NEXT: movaps %xmm3, %xmm4
+; X86-SSE-NEXT: andnps 40(%ebp), %xmm4
+; X86-SSE-NEXT: orps %xmm4, %xmm1
+; X86-SSE-NEXT: andps %xmm3, %xmm2
+; X86-SSE-NEXT: movaps %xmm3, %xmm4
+; X86-SSE-NEXT: andnps 56(%ebp), %xmm4
+; X86-SSE-NEXT: orps %xmm4, %xmm2
+; X86-SSE-NEXT: movaps %xmm3, %xmm4
+; X86-SSE-NEXT: andnps 72(%ebp), %xmm4
+; X86-SSE-NEXT: andps 8(%ebp), %xmm3
+; X86-SSE-NEXT: orps %xmm4, %xmm3
+; X86-SSE-NEXT: movl %ebp, %esp
+; X86-SSE-NEXT: popl %ebp
+; X86-SSE-NEXT: retl
+;
+; X86-AVX1OR2-LABEL: fcopysign_v16f32:
+; X86-AVX1OR2: # %bb.0:
+; X86-AVX1OR2-NEXT: pushl %ebp
+; X86-AVX1OR2-NEXT: movl %esp, %ebp
+; X86-AVX1OR2-NEXT: andl $-32, %esp
+; X86-AVX1OR2-NEXT: subl $32, %esp
+; X86-AVX1OR2-NEXT: vbroadcastss {{.*#+}} ymm3 = [NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN]
+; X86-AVX1OR2-NEXT: vandnps %ymm2, %ymm3, %ymm2
+; X86-AVX1OR2-NEXT: vandps %ymm3, %ymm0, %ymm0
+; X86-AVX1OR2-NEXT: vorps %ymm2, %ymm0, %ymm0
+; X86-AVX1OR2-NEXT: vandps %ymm3, %ymm1, %ymm1
+; X86-AVX1OR2-NEXT: vandnps 8(%ebp), %ymm3, %ymm2
+; X86-AVX1OR2-NEXT: vorps %ymm2, %ymm1, %ymm1
+; X86-AVX1OR2-NEXT: movl %ebp, %esp
+; X86-AVX1OR2-NEXT: popl %ebp
+; X86-AVX1OR2-NEXT: retl
+;
+; X86-AVX512-LABEL: fcopysign_v16f32:
+; X86-AVX512: # %bb.0:
+; X86-AVX512-NEXT: vpternlogd $228, {{\.?LCPI[0-9]+_[0-9]+}}{1to16}, %zmm1, %zmm0
+; X86-AVX512-NEXT: retl
+;
+; X64-SSE-LABEL: fcopysign_v16f32:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: movaps {{.*#+}} xmm8 = [NaN,NaN,NaN,NaN]
+; X64-SSE-NEXT: movaps %xmm8, %xmm9
+; X64-SSE-NEXT: andnps %xmm4, %xmm9
+; X64-SSE-NEXT: andps %xmm8, %xmm0
+; X64-SSE-NEXT: orps %xmm9, %xmm0
+; X64-SSE-NEXT: movaps %xmm8, %xmm4
+; X64-SSE-NEXT: andnps %xmm5, %xmm4
+; X64-SSE-NEXT: andps %xmm8, %xmm1
+; X64-SSE-NEXT: orps %xmm4, %xmm1
+; X64-SSE-NEXT: movaps %xmm8, %xmm4
+; X64-SSE-NEXT: andnps %xmm6, %xmm4
+; X64-SSE-NEXT: andps %xmm8, %xmm2
+; X64-SSE-NEXT: orps %xmm4, %xmm2
+; X64-SSE-NEXT: andps %xmm8, %xmm3
+; X64-SSE-NEXT: andnps %xmm7, %xmm8
+; X64-SSE-NEXT: orps %xmm8, %xmm3
+; X64-SSE-NEXT: retq
+;
+; X64-AVX1OR2-LABEL: fcopysign_v16f32:
+; X64-AVX1OR2: # %bb.0:
+; X64-AVX1OR2-NEXT: vbroadcastss {{.*#+}} ymm4 = [NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN]
+; X64-AVX1OR2-NEXT: vandnps %ymm2, %ymm4, %ymm2
+; X64-AVX1OR2-NEXT: vandps %ymm4, %ymm0, %ymm0
+; X64-AVX1OR2-NEXT: vorps %ymm2, %ymm0, %ymm0
+; X64-AVX1OR2-NEXT: vandnps %ymm3, %ymm4, %ymm2
+; X64-AVX1OR2-NEXT: vandps %ymm4, %ymm1, %ymm1
+; X64-AVX1OR2-NEXT: vorps %ymm2, %ymm1, %ymm1
+; X64-AVX1OR2-NEXT: retq
+;
+; X64-AVX512-LABEL: fcopysign_v16f32:
+; X64-AVX512: # %bb.0:
+; X64-AVX512-NEXT: vpternlogd $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to16}, %zmm1, %zmm0
+; X64-AVX512-NEXT: retq
+ %t = call <16 x float> @llvm.copysign.v16f32(<16 x float> %a0, <16 x float> %a1)
+ ret <16 x float> %t
+}
+declare <16 x float> @llvm.copysign.v16f32(<16 x float>, <16 x float>)
+
+define <32 x half> @fcopysign_v32f16(ptr %p0, ptr %p1) nounwind {
+; X86-SSE-LABEL: fcopysign_v32f16:
+; X86-SSE: # %bb.0:
+; X86-SSE-NEXT: pushl %esi
+; X86-SSE-NEXT: subl $32, %esp
+; X86-SSE-NEXT: movl {{[0-9]+}}(%esp), %eax
+; X86-SSE-NEXT: movl {{[0-9]+}}(%esp), %ecx
+; X86-SSE-NEXT: movaps 48(%ecx), %xmm0
+; X86-SSE-NEXT: movups %xmm0, {{[-0-9]+}}(%e{{[sb]}}p) # 16-byte Spill
+; X86-SSE-NEXT: movdqa 32(%ecx), %xmm4
+; X86-SSE-NEXT: movdqa (%ecx), %xmm5
+; X86-SSE-NEXT: movdqa 16(%ecx), %xmm3
+; X86-SSE-NEXT: movdqa (%eax), %xmm0
+; X86-SSE-NEXT: movdqa %xmm0, %xmm1
+; X86-SSE-NEXT: psrldq {{.*#+}} xmm1 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-SSE-NEXT: pextrw $0, %xmm1, %ecx
+; X86-SSE-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X86-SSE-NEXT: movdqa %xmm5, %xmm1
+; X86-SSE-NEXT: psrldq {{.*#+}} xmm1 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-SSE-NEXT: pextrw $0, %xmm1, %edx
+; X86-SSE-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %ecx, %edx
+; X86-SSE-NEXT: pinsrw $0, %edx, %xmm6
+; X86-SSE-NEXT: movdqa %xmm0, %xmm1
+; X86-SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[3,3],xmm0[3,3]
+; X86-SSE-NEXT: pextrw $0, %xmm1, %ecx
+; X86-SSE-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X86-SSE-NEXT: movdqa %xmm5, %xmm1
+; X86-SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[3,3],xmm5[3,3]
+; X86-SSE-NEXT: pextrw $0, %xmm1, %edx
+; X86-SSE-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %ecx, %edx
+; X86-SSE-NEXT: pinsrw $0, %edx, %xmm1
+; X86-SSE-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm6[0],xmm1[1],xmm6[1],xmm1[2],xmm6[2],xmm1[3],xmm6[3]
+; X86-SSE-NEXT: movdqa %xmm0, %xmm6
+; X86-SSE-NEXT: psrldq {{.*#+}} xmm6 = xmm6[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-SSE-NEXT: pextrw $0, %xmm6, %ecx
+; X86-SSE-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X86-SSE-NEXT: movdqa %xmm5, %xmm6
+; X86-SSE-NEXT: psrldq {{.*#+}} xmm6 = xmm6[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-SSE-NEXT: pextrw $0, %xmm6, %edx
+; X86-SSE-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %ecx, %edx
+; X86-SSE-NEXT: pinsrw $0, %edx, %xmm7
+; X86-SSE-NEXT: movdqa %xmm0, %xmm6
+; X86-SSE-NEXT: punpckhqdq {{.*#+}} xmm6 = xmm6[1],xmm0[1]
+; X86-SSE-NEXT: pextrw $0, %xmm6, %ecx
+; X86-SSE-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X86-SSE-NEXT: movdqa %xmm5, %xmm6
+; X86-SSE-NEXT: punpckhqdq {{.*#+}} xmm6 = xmm6[1],xmm5[1]
+; X86-SSE-NEXT: pextrw $0, %xmm6, %edx
+; X86-SSE-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %ecx, %edx
+; X86-SSE-NEXT: pinsrw $0, %edx, %xmm6
+; X86-SSE-NEXT: punpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
+; X86-SSE-NEXT: punpckldq {{.*#+}} xmm6 = xmm6[0],xmm1[0],xmm6[1],xmm1[1]
+; X86-SSE-NEXT: movdqa %xmm0, %xmm1
+; X86-SSE-NEXT: psrlq $48, %xmm1
+; X86-SSE-NEXT: pextrw $0, %xmm1, %ecx
+; X86-SSE-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X86-SSE-NEXT: movdqa %xmm5, %xmm1
+; X86-SSE-NEXT: psrlq $48, %xmm1
+; X86-SSE-NEXT: pextrw $0, %xmm1, %edx
+; X86-SSE-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %ecx, %edx
+; X86-SSE-NEXT: pinsrw $0, %edx, %xmm1
+; X86-SSE-NEXT: movaps %xmm0, %xmm7
+; X86-SSE-NEXT: shufps {{.*#+}} xmm7 = xmm7[1,1],xmm0[1,1]
+; X86-SSE-NEXT: pextrw $0, %xmm7, %ecx
+; X86-SSE-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X86-SSE-NEXT: movaps %xmm5, %xmm7
+; X86-SSE-NEXT: shufps {{.*#+}} xmm7 = xmm7[1,1],xmm5[1,1]
+; X86-SSE-NEXT: pextrw $0, %xmm7, %edx
+; X86-SSE-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %ecx, %edx
+; X86-SSE-NEXT: pinsrw $0, %edx, %xmm7
+; X86-SSE-NEXT: punpcklwd {{.*#+}} xmm7 = xmm7[0],xmm1[0],xmm7[1],xmm1[1],xmm7[2],xmm1[2],xmm7[3],xmm1[3]
+; X86-SSE-NEXT: movdqa 16(%eax), %xmm2
+; X86-SSE-NEXT: pextrw $0, %xmm0, %edx
+; X86-SSE-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-SSE-NEXT: pextrw $0, %xmm5, %ecx
+; X86-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %edx, %ecx
+; X86-SSE-NEXT: psrld $16, %xmm0
+; X86-SSE-NEXT: pextrw $0, %xmm0, %edx
+; X86-SSE-NEXT: pinsrw $0, %ecx, %xmm1
+; X86-SSE-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-SSE-NEXT: psrld $16, %xmm5
+; X86-SSE-NEXT: pextrw $0, %xmm5, %ecx
+; X86-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %edx, %ecx
+; X86-SSE-NEXT: pinsrw $0, %ecx, %xmm0
+; X86-SSE-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; X86-SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm7[0],xmm1[1],xmm7[1]
+; X86-SSE-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm6[0]
+; X86-SSE-NEXT: movdqu %xmm1, (%esp) # 16-byte Spill
+; X86-SSE-NEXT: movdqa %xmm2, %xmm0
+; X86-SSE-NEXT: psrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-SSE-NEXT: pextrw $0, %xmm0, %ecx
+; X86-SSE-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X86-SSE-NEXT: movdqa %xmm3, %xmm0
+; X86-SSE-NEXT: psrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-SSE-NEXT: pextrw $0, %xmm0, %edx
+; X86-SSE-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %ecx, %edx
+; X86-SSE-NEXT: pinsrw $0, %edx, %xmm0
+; X86-SSE-NEXT: movdqa %xmm2, %xmm5
+; X86-SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[3,3],xmm2[3,3]
+; X86-SSE-NEXT: pextrw $0, %xmm5, %ecx
+; X86-SSE-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X86-SSE-NEXT: movdqa %xmm3, %xmm5
+; X86-SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[3,3],xmm3[3,3]
+; X86-SSE-NEXT: pextrw $0, %xmm5, %edx
+; X86-SSE-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %ecx, %edx
+; X86-SSE-NEXT: pinsrw $0, %edx, %xmm5
+; X86-SSE-NEXT: punpcklwd {{.*#+}} xmm5 = xmm5[0],xmm0[0],xmm5[1],xmm0[1],xmm5[2],xmm0[2],xmm5[3],xmm0[3]
+; X86-SSE-NEXT: movdqa %xmm2, %xmm0
+; X86-SSE-NEXT: psrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-SSE-NEXT: pextrw $0, %xmm0, %ecx
+; X86-SSE-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X86-SSE-NEXT: movdqa %xmm3, %xmm0
+; X86-SSE-NEXT: psrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-SSE-NEXT: pextrw $0, %xmm0, %edx
+; X86-SSE-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %ecx, %edx
+; X86-SSE-NEXT: pinsrw $0, %edx, %xmm0
+; X86-SSE-NEXT: movdqa %xmm2, %xmm6
+; X86-SSE-NEXT: punpckhqdq {{.*#+}} xmm6 = xmm6[1],xmm2[1]
+; X86-SSE-NEXT: pextrw $0, %xmm6, %ecx
+; X86-SSE-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X86-SSE-NEXT: movdqa %xmm3, %xmm6
+; X86-SSE-NEXT: punpckhqdq {{.*#+}} xmm6 = xmm6[1],xmm3[1]
+; X86-SSE-NEXT: pextrw $0, %xmm6, %edx
+; X86-SSE-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %ecx, %edx
+; X86-SSE-NEXT: pinsrw $0, %edx, %xmm6
+; X86-SSE-NEXT: punpcklwd {{.*#+}} xmm6 = xmm6[0],xmm0[0],xmm6[1],xmm0[1],xmm6[2],xmm0[2],xmm6[3],xmm0[3]
+; X86-SSE-NEXT: punpckldq {{.*#+}} xmm6 = xmm6[0],xmm5[0],xmm6[1],xmm5[1]
+; X86-SSE-NEXT: movdqa %xmm2, %xmm0
+; X86-SSE-NEXT: psrlq $48, %xmm0
+; X86-SSE-NEXT: pextrw $0, %xmm0, %ecx
+; X86-SSE-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X86-SSE-NEXT: movdqa %xmm3, %xmm0
+; X86-SSE-NEXT: psrlq $48, %xmm0
+; X86-SSE-NEXT: pextrw $0, %xmm0, %edx
+; X86-SSE-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %ecx, %edx
+; X86-SSE-NEXT: pinsrw $0, %edx, %xmm0
+; X86-SSE-NEXT: movaps %xmm2, %xmm5
+; X86-SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[1,1],xmm2[1,1]
+; X86-SSE-NEXT: pextrw $0, %xmm5, %ecx
+; X86-SSE-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X86-SSE-NEXT: movaps %xmm3, %xmm5
+; X86-SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[1,1],xmm3[1,1]
+; X86-SSE-NEXT: pextrw $0, %xmm5, %edx
+; X86-SSE-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %ecx, %edx
+; X86-SSE-NEXT: pinsrw $0, %edx, %xmm7
+; X86-SSE-NEXT: punpcklwd {{.*#+}} xmm7 = xmm7[0],xmm0[0],xmm7[1],xmm0[1],xmm7[2],xmm0[2],xmm7[3],xmm0[3]
+; X86-SSE-NEXT: movdqa 32(%eax), %xmm5
+; X86-SSE-NEXT: pextrw $0, %xmm2, %edx
+; X86-SSE-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-SSE-NEXT: pextrw $0, %xmm3, %ecx
+; X86-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %edx, %ecx
+; X86-SSE-NEXT: psrld $16, %xmm2
+; X86-SSE-NEXT: pextrw $0, %xmm2, %edx
+; X86-SSE-NEXT: pinsrw $0, %ecx, %xmm1
+; X86-SSE-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-SSE-NEXT: psrld $16, %xmm3
+; X86-SSE-NEXT: pextrw $0, %xmm3, %ecx
+; X86-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %edx, %ecx
+; X86-SSE-NEXT: pinsrw $0, %ecx, %xmm0
+; X86-SSE-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; X86-SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm7[0],xmm1[1],xmm7[1]
+; X86-SSE-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm6[0]
+; X86-SSE-NEXT: movdqa %xmm5, %xmm0
+; X86-SSE-NEXT: psrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-SSE-NEXT: pextrw $0, %xmm0, %ecx
+; X86-SSE-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X86-SSE-NEXT: movdqa %xmm4, %xmm0
+; X86-SSE-NEXT: psrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-SSE-NEXT: pextrw $0, %xmm0, %edx
+; X86-SSE-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %ecx, %edx
+; X86-SSE-NEXT: pinsrw $0, %edx, %xmm0
+; X86-SSE-NEXT: movdqa %xmm5, %xmm2
+; X86-SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[3,3],xmm5[3,3]
+; X86-SSE-NEXT: pextrw $0, %xmm2, %ecx
+; X86-SSE-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X86-SSE-NEXT: movdqa %xmm4, %xmm2
+; X86-SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[3,3],xmm4[3,3]
+; X86-SSE-NEXT: pextrw $0, %xmm2, %edx
+; X86-SSE-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %ecx, %edx
+; X86-SSE-NEXT: pinsrw $0, %edx, %xmm2
+; X86-SSE-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
+; X86-SSE-NEXT: movdqa %xmm5, %xmm0
+; X86-SSE-NEXT: psrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-SSE-NEXT: pextrw $0, %xmm0, %ecx
+; X86-SSE-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X86-SSE-NEXT: movdqa %xmm4, %xmm0
+; X86-SSE-NEXT: psrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-SSE-NEXT: pextrw $0, %xmm0, %edx
+; X86-SSE-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %ecx, %edx
+; X86-SSE-NEXT: pinsrw $0, %edx, %xmm0
+; X86-SSE-NEXT: movdqa %xmm5, %xmm6
+; X86-SSE-NEXT: punpckhqdq {{.*#+}} xmm6 = xmm6[1],xmm5[1]
+; X86-SSE-NEXT: pextrw $0, %xmm6, %ecx
+; X86-SSE-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X86-SSE-NEXT: movdqa %xmm4, %xmm6
+; X86-SSE-NEXT: punpckhqdq {{.*#+}} xmm6 = xmm6[1],xmm4[1]
+; X86-SSE-NEXT: pextrw $0, %xmm6, %edx
+; X86-SSE-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %ecx, %edx
+; X86-SSE-NEXT: pinsrw $0, %edx, %xmm7
+; X86-SSE-NEXT: punpcklwd {{.*#+}} xmm7 = xmm7[0],xmm0[0],xmm7[1],xmm0[1],xmm7[2],xmm0[2],xmm7[3],xmm0[3]
+; X86-SSE-NEXT: movdqa %xmm5, %xmm0
+; X86-SSE-NEXT: psrlq $48, %xmm0
+; X86-SSE-NEXT: pextrw $0, %xmm0, %edx
+; X86-SSE-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-SSE-NEXT: movdqa %xmm4, %xmm0
+; X86-SSE-NEXT: psrlq $48, %xmm0
+; X86-SSE-NEXT: pextrw $0, %xmm0, %ecx
+; X86-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %edx, %ecx
+; X86-SSE-NEXT: movaps %xmm5, %xmm0
+; X86-SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1],xmm5[1,1]
+; X86-SSE-NEXT: pextrw $0, %xmm0, %edx
+; X86-SSE-NEXT: movaps %xmm4, %xmm0
+; X86-SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1],xmm4[1,1]
+; X86-SSE-NEXT: pextrw $0, %xmm0, %esi
+; X86-SSE-NEXT: pinsrw $0, %ecx, %xmm6
+; X86-SSE-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-SSE-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-SSE-NEXT: orl %edx, %esi
+; X86-SSE-NEXT: punpckldq {{.*#+}} xmm7 = xmm7[0],xmm2[0],xmm7[1],xmm2[1]
+; X86-SSE-NEXT: pinsrw $0, %esi, %xmm0
+; X86-SSE-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm6[0],xmm0[1],xmm6[1],xmm0[2],xmm6[2],xmm0[3],xmm6[3]
+; X86-SSE-NEXT: movdqa 48(%eax), %xmm6
+; X86-SSE-NEXT: pextrw $0, %xmm5, %ecx
+; X86-SSE-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X86-SSE-NEXT: pextrw $0, %xmm4, %eax
+; X86-SSE-NEXT: andl $32767, %eax # imm = 0x7FFF
+; X86-SSE-NEXT: orl %ecx, %eax
+; X86-SSE-NEXT: psrld $16, %xmm5
+; X86-SSE-NEXT: pextrw $0, %xmm5, %ecx
+; X86-SSE-NEXT: pinsrw $0, %eax, %xmm2
+; X86-SSE-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X86-SSE-NEXT: psrld $16, %xmm4
+; X86-SSE-NEXT: pextrw $0, %xmm4, %eax
+; X86-SSE-NEXT: andl $32767, %eax # imm = 0x7FFF
+; X86-SSE-NEXT: orl %ecx, %eax
+; X86-SSE-NEXT: pinsrw $0, %eax, %xmm3
+; X86-SSE-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
+; X86-SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]
+; X86-SSE-NEXT: punpcklqdq {{.*#+}} xmm2 = xmm2[0],xmm7[0]
+; X86-SSE-NEXT: movdqa %xmm6, %xmm0
+; X86-SSE-NEXT: psrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-SSE-NEXT: pextrw $0, %xmm0, %eax
+; X86-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-SSE-NEXT: movdqu {{[-0-9]+}}(%e{{[sb]}}p), %xmm4 # 16-byte Reload
+; X86-SSE-NEXT: movdqa %xmm4, %xmm0
+; X86-SSE-NEXT: psrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-SSE-NEXT: pextrw $0, %xmm0, %ecx
+; X86-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %eax, %ecx
+; X86-SSE-NEXT: pinsrw $0, %ecx, %xmm0
+; X86-SSE-NEXT: movdqa %xmm6, %xmm3
+; X86-SSE-NEXT: shufps {{.*#+}} xmm3 = xmm3[3,3],xmm6[3,3]
+; X86-SSE-NEXT: pextrw $0, %xmm3, %eax
+; X86-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-SSE-NEXT: movdqa %xmm4, %xmm3
+; X86-SSE-NEXT: shufps {{.*#+}} xmm3 = xmm3[3,3],xmm4[3,3]
+; X86-SSE-NEXT: pextrw $0, %xmm3, %ecx
+; X86-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %eax, %ecx
+; X86-SSE-NEXT: pinsrw $0, %ecx, %xmm3
+; X86-SSE-NEXT: punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3]
+; X86-SSE-NEXT: movdqa %xmm6, %xmm0
+; X86-SSE-NEXT: psrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-SSE-NEXT: pextrw $0, %xmm0, %eax
+; X86-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-SSE-NEXT: movdqa %xmm4, %xmm0
+; X86-SSE-NEXT: psrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-SSE-NEXT: pextrw $0, %xmm0, %ecx
+; X86-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %eax, %ecx
+; X86-SSE-NEXT: pinsrw $0, %ecx, %xmm0
+; X86-SSE-NEXT: movdqa %xmm6, %xmm5
+; X86-SSE-NEXT: punpckhqdq {{.*#+}} xmm5 = xmm5[1],xmm6[1]
+; X86-SSE-NEXT: pextrw $0, %xmm5, %eax
+; X86-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-SSE-NEXT: movdqa %xmm4, %xmm5
+; X86-SSE-NEXT: punpckhqdq {{.*#+}} xmm5 = xmm5[1],xmm4[1]
+; X86-SSE-NEXT: pextrw $0, %xmm5, %ecx
+; X86-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %eax, %ecx
+; X86-SSE-NEXT: pinsrw $0, %ecx, %xmm5
+; X86-SSE-NEXT: punpcklwd {{.*#+}} xmm5 = xmm5[0],xmm0[0],xmm5[1],xmm0[1],xmm5[2],xmm0[2],xmm5[3],xmm0[3]
+; X86-SSE-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm3[0],xmm5[1],xmm3[1]
+; X86-SSE-NEXT: movdqa %xmm6, %xmm0
+; X86-SSE-NEXT: psrlq $48, %xmm0
+; X86-SSE-NEXT: pextrw $0, %xmm0, %eax
+; X86-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-SSE-NEXT: movdqa %xmm4, %xmm0
+; X86-SSE-NEXT: psrlq $48, %xmm0
+; X86-SSE-NEXT: pextrw $0, %xmm0, %ecx
+; X86-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %eax, %ecx
+; X86-SSE-NEXT: pinsrw $0, %ecx, %xmm3
+; X86-SSE-NEXT: movaps %xmm6, %xmm0
+; X86-SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1],xmm6[1,1]
+; X86-SSE-NEXT: pextrw $0, %xmm0, %eax
+; X86-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-SSE-NEXT: movaps %xmm4, %xmm0
+; X86-SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1],xmm4[1,1]
+; X86-SSE-NEXT: pextrw $0, %xmm0, %ecx
+; X86-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-SSE-NEXT: orl %eax, %ecx
+; X86-SSE-NEXT: pinsrw $0, %ecx, %xmm0
+; X86-SSE-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
+; X86-SSE-NEXT: pextrw $0, %xmm6, %ecx
+; X86-SSE-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X86-SSE-NEXT: pextrw $0, %xmm4, %eax
+; X86-SSE-NEXT: andl $32767, %eax # imm = 0x7FFF
+; X86-SSE-NEXT: orl %ecx, %eax
+; X86-SSE-NEXT: psrld $16, %xmm6
+; X86-SSE-NEXT: pextrw $0, %xmm6, %ecx
+; X86-SSE-NEXT: pinsrw $0, %eax, %xmm3
+; X86-SSE-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X86-SSE-NEXT: psrld $16, %xmm4
+; X86-SSE-NEXT: pextrw $0, %xmm4, %eax
+; X86-SSE-NEXT: andl $32767, %eax # imm = 0x7FFF
+; X86-SSE-NEXT: orl %ecx, %eax
+; X86-SSE-NEXT: pinsrw $0, %eax, %xmm4
+; X86-SSE-NEXT: punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
+; X86-SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1]
+; X86-SSE-NEXT: punpcklqdq {{.*#+}} xmm3 = xmm3[0],xmm5[0]
+; X86-SSE-NEXT: movups (%esp), %xmm0 # 16-byte Reload
+; X86-SSE-NEXT: addl $32, %esp
+; X86-SSE-NEXT: popl %esi
+; X86-SSE-NEXT: retl
+;
+; X86-AVX1-LABEL: fcopysign_v32f16:
+; X86-AVX1: # %bb.0:
+; X86-AVX1-NEXT: pushl %esi
+; X86-AVX1-NEXT: movl {{[0-9]+}}(%esp), %eax
+; X86-AVX1-NEXT: movl {{[0-9]+}}(%esp), %ecx
+; X86-AVX1-NEXT: vbroadcastss 28(%ecx), %xmm0
+; X86-AVX1-NEXT: vpextrw $0, %xmm0, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vbroadcastss 28(%eax), %xmm0
+; X86-AVX1-NEXT: vpextrw $0, %xmm0, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm4
+; X86-AVX1-NEXT: vmovdqa (%ecx), %xmm0
+; X86-AVX1-NEXT: vmovdqa 16(%ecx), %xmm2
+; X86-AVX1-NEXT: vpsrldq {{.*#+}} xmm1 = xmm2[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX1-NEXT: vpextrw $0, %xmm1, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vmovdqa (%eax), %xmm1
+; X86-AVX1-NEXT: vmovdqa 16(%eax), %xmm3
+; X86-AVX1-NEXT: vpsrldq {{.*#+}} xmm5 = xmm3[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX1-NEXT: vpextrw $0, %xmm5, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm5
+; X86-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
+; X86-AVX1-NEXT: vbroadcastss 24(%ecx), %xmm5
+; X86-AVX1-NEXT: vpextrw $0, %xmm5, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vbroadcastss 24(%eax), %xmm5
+; X86-AVX1-NEXT: vpextrw $0, %xmm5, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm5
+; X86-AVX1-NEXT: vpsrldq {{.*#+}} xmm6 = xmm2[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX1-NEXT: vpextrw $0, %xmm6, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vpsrldq {{.*#+}} xmm6 = xmm3[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX1-NEXT: vpextrw $0, %xmm6, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm6
+; X86-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
+; X86-AVX1-NEXT: vpunpckldq {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]
+; X86-AVX1-NEXT: vbroadcastss 20(%ecx), %xmm5
+; X86-AVX1-NEXT: vpextrw $0, %xmm5, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vbroadcastss 20(%eax), %xmm5
+; X86-AVX1-NEXT: vpextrw $0, %xmm5, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm5
+; X86-AVX1-NEXT: vpsrlq $48, %xmm2, %xmm6
+; X86-AVX1-NEXT: vpextrw $0, %xmm6, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vpsrlq $48, %xmm3, %xmm6
+; X86-AVX1-NEXT: vpextrw $0, %xmm6, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm6
+; X86-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
+; X86-AVX1-NEXT: vpextrw $0, %xmm2, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vpextrw $0, %xmm3, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm6
+; X86-AVX1-NEXT: vpsrld $16, %xmm2, %xmm2
+; X86-AVX1-NEXT: vpextrw $0, %xmm2, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vpsrld $16, %xmm3, %xmm2
+; X86-AVX1-NEXT: vpextrw $0, %xmm2, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm2
+; X86-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm6[0],xmm2[0],xmm6[1],xmm2[1],xmm6[2],xmm2[2],xmm6[3],xmm2[3]
+; X86-AVX1-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm2[0],xmm5[0],xmm2[1],xmm5[1]
+; X86-AVX1-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm2[0],xmm4[0]
+; X86-AVX1-NEXT: vbroadcastss 12(%ecx), %xmm3
+; X86-AVX1-NEXT: vpextrw $0, %xmm3, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vbroadcastss 12(%eax), %xmm3
+; X86-AVX1-NEXT: vpextrw $0, %xmm3, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm3
+; X86-AVX1-NEXT: vpsrldq {{.*#+}} xmm4 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX1-NEXT: vpextrw $0, %xmm4, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vpsrldq {{.*#+}} xmm4 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX1-NEXT: vpextrw $0, %xmm4, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm4
+; X86-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
+; X86-AVX1-NEXT: vbroadcastss 8(%ecx), %xmm4
+; X86-AVX1-NEXT: vpextrw $0, %xmm4, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vbroadcastss 8(%eax), %xmm4
+; X86-AVX1-NEXT: vpextrw $0, %xmm4, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm4
+; X86-AVX1-NEXT: vpsrldq {{.*#+}} xmm5 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX1-NEXT: vpextrw $0, %xmm5, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vpsrldq {{.*#+}} xmm5 = xmm1[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX1-NEXT: vpextrw $0, %xmm5, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm5
+; X86-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
+; X86-AVX1-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]
+; X86-AVX1-NEXT: vbroadcastss 4(%ecx), %xmm4
+; X86-AVX1-NEXT: vpextrw $0, %xmm4, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vbroadcastss 4(%eax), %xmm4
+; X86-AVX1-NEXT: vpextrw $0, %xmm4, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm4
+; X86-AVX1-NEXT: vpsrlq $48, %xmm0, %xmm5
+; X86-AVX1-NEXT: vpextrw $0, %xmm5, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vpsrlq $48, %xmm1, %xmm5
+; X86-AVX1-NEXT: vpextrw $0, %xmm5, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm5
+; X86-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
+; X86-AVX1-NEXT: vpextrw $0, %xmm0, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vpextrw $0, %xmm1, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm5
+; X86-AVX1-NEXT: vpsrld $16, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpextrw $0, %xmm0, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vpsrld $16, %xmm1, %xmm0
+; X86-AVX1-NEXT: vpextrw $0, %xmm0, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm5[0],xmm0[0],xmm5[1],xmm0[1],xmm5[2],xmm0[2],xmm5[3],xmm0[3]
+; X86-AVX1-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1]
+; X86-AVX1-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm3[0]
+; X86-AVX1-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm0
+; X86-AVX1-NEXT: vbroadcastss 60(%ecx), %xmm1
+; X86-AVX1-NEXT: vpextrw $0, %xmm1, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vbroadcastss 60(%eax), %xmm1
+; X86-AVX1-NEXT: vpextrw $0, %xmm1, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm3
+; X86-AVX1-NEXT: vmovdqa 48(%ecx), %xmm1
+; X86-AVX1-NEXT: vpsrldq {{.*#+}} xmm2 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX1-NEXT: vpextrw $0, %xmm2, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vmovdqa 48(%eax), %xmm2
+; X86-AVX1-NEXT: vpsrldq {{.*#+}} xmm4 = xmm2[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX1-NEXT: vpextrw $0, %xmm4, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm4
+; X86-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
+; X86-AVX1-NEXT: vbroadcastss 56(%ecx), %xmm4
+; X86-AVX1-NEXT: vpextrw $0, %xmm4, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vbroadcastss 56(%eax), %xmm4
+; X86-AVX1-NEXT: vpextrw $0, %xmm4, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm4
+; X86-AVX1-NEXT: vpsrldq {{.*#+}} xmm5 = xmm1[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX1-NEXT: vpextrw $0, %xmm5, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vpsrldq {{.*#+}} xmm5 = xmm2[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX1-NEXT: vpextrw $0, %xmm5, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm5
+; X86-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
+; X86-AVX1-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]
+; X86-AVX1-NEXT: vbroadcastss 52(%ecx), %xmm4
+; X86-AVX1-NEXT: vpextrw $0, %xmm4, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vbroadcastss 52(%eax), %xmm4
+; X86-AVX1-NEXT: vpextrw $0, %xmm4, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm4
+; X86-AVX1-NEXT: vpsrlq $48, %xmm1, %xmm5
+; X86-AVX1-NEXT: vpextrw $0, %xmm5, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vpsrlq $48, %xmm2, %xmm5
+; X86-AVX1-NEXT: vpextrw $0, %xmm5, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm5
+; X86-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
+; X86-AVX1-NEXT: vpextrw $0, %xmm1, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vpextrw $0, %xmm2, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm5
+; X86-AVX1-NEXT: vpsrld $16, %xmm1, %xmm1
+; X86-AVX1-NEXT: vpextrw $0, %xmm1, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vpsrld $16, %xmm2, %xmm1
+; X86-AVX1-NEXT: vpextrw $0, %xmm1, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm1
+; X86-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm5[0],xmm1[0],xmm5[1],xmm1[1],xmm5[2],xmm1[2],xmm5[3],xmm1[3]
+; X86-AVX1-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1]
+; X86-AVX1-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm3[0]
+; X86-AVX1-NEXT: vbroadcastss 44(%ecx), %xmm2
+; X86-AVX1-NEXT: vpextrw $0, %xmm2, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vbroadcastss 44(%eax), %xmm2
+; X86-AVX1-NEXT: vpextrw $0, %xmm2, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm4
+; X86-AVX1-NEXT: vmovdqa 32(%ecx), %xmm2
+; X86-AVX1-NEXT: vpsrldq {{.*#+}} xmm3 = xmm2[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX1-NEXT: vpextrw $0, %xmm3, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vmovdqa 32(%eax), %xmm3
+; X86-AVX1-NEXT: vpsrldq {{.*#+}} xmm5 = xmm3[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX1-NEXT: vpextrw $0, %xmm5, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm5
+; X86-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
+; X86-AVX1-NEXT: vbroadcastss 40(%ecx), %xmm5
+; X86-AVX1-NEXT: vpextrw $0, %xmm5, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vbroadcastss 40(%eax), %xmm5
+; X86-AVX1-NEXT: vpextrw $0, %xmm5, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm5
+; X86-AVX1-NEXT: vpsrldq {{.*#+}} xmm6 = xmm2[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX1-NEXT: vpextrw $0, %xmm6, %edx
+; X86-AVX1-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX1-NEXT: vpsrldq {{.*#+}} xmm6 = xmm3[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX1-NEXT: vpextrw $0, %xmm6, %esi
+; X86-AVX1-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %edx, %esi
+; X86-AVX1-NEXT: vpinsrw $0, %esi, %xmm0, %xmm6
+; X86-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
+; X86-AVX1-NEXT: vpunpckldq {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]
+; X86-AVX1-NEXT: vbroadcastss 36(%ecx), %xmm5
+; X86-AVX1-NEXT: vpextrw $0, %xmm5, %ecx
+; X86-AVX1-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X86-AVX1-NEXT: vbroadcastss 36(%eax), %xmm5
+; X86-AVX1-NEXT: vpextrw $0, %xmm5, %eax
+; X86-AVX1-NEXT: andl $32767, %eax # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %ecx, %eax
+; X86-AVX1-NEXT: vpinsrw $0, %eax, %xmm0, %xmm5
+; X86-AVX1-NEXT: vpsrlq $48, %xmm2, %xmm6
+; X86-AVX1-NEXT: vpextrw $0, %xmm6, %eax
+; X86-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-AVX1-NEXT: vpsrlq $48, %xmm3, %xmm6
+; X86-AVX1-NEXT: vpextrw $0, %xmm6, %ecx
+; X86-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %eax, %ecx
+; X86-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm6
+; X86-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
+; X86-AVX1-NEXT: vpextrw $0, %xmm2, %eax
+; X86-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-AVX1-NEXT: vpextrw $0, %xmm3, %ecx
+; X86-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %eax, %ecx
+; X86-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm6
+; X86-AVX1-NEXT: vpsrld $16, %xmm2, %xmm2
+; X86-AVX1-NEXT: vpextrw $0, %xmm2, %eax
+; X86-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X86-AVX1-NEXT: vpsrld $16, %xmm3, %xmm2
+; X86-AVX1-NEXT: vpextrw $0, %xmm2, %ecx
+; X86-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X86-AVX1-NEXT: orl %eax, %ecx
+; X86-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm2
+; X86-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm6[0],xmm2[0],xmm6[1],xmm2[1],xmm6[2],xmm2[2],xmm6[3],xmm2[3]
+; X86-AVX1-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm2[0],xmm5[0],xmm2[1],xmm5[1]
+; X86-AVX1-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm2[0],xmm4[0]
+; X86-AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm2, %ymm1
+; X86-AVX1-NEXT: popl %esi
+; X86-AVX1-NEXT: retl
+;
+; X86-AVX2-LABEL: fcopysign_v32f16:
+; X86-AVX2: # %bb.0:
+; X86-AVX2-NEXT: pushl %ebx
+; X86-AVX2-NEXT: pushl %edi
+; X86-AVX2-NEXT: pushl %esi
+; X86-AVX2-NEXT: subl $32, %esp
+; X86-AVX2-NEXT: movl {{[0-9]+}}(%esp), %eax
+; X86-AVX2-NEXT: movl {{[0-9]+}}(%esp), %ecx
+; X86-AVX2-NEXT: vmovdqa (%ecx), %xmm0
+; X86-AVX2-NEXT: vmovdqa 16(%ecx), %xmm2
+; X86-AVX2-NEXT: vpsrldq {{.*#+}} xmm1 = xmm2[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX2-NEXT: vpextrw $0, %xmm1, %edx
+; X86-AVX2-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX2-NEXT: vmovdqa (%eax), %xmm1
+; X86-AVX2-NEXT: vmovdqa 16(%eax), %xmm3
+; X86-AVX2-NEXT: vpsrldq {{.*#+}} xmm4 = xmm3[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX2-NEXT: vpextrw $0, %xmm4, %esi
+; X86-AVX2-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %edx, %esi
+; X86-AVX2-NEXT: vpinsrw $0, %esi, %xmm0, %xmm4
+; X86-AVX2-NEXT: movl $-32768, %edx # imm = 0x8000
+; X86-AVX2-NEXT: movl 28(%ecx), %esi
+; X86-AVX2-NEXT: andl %edx, %esi
+; X86-AVX2-NEXT: movzwl 28(%eax), %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm5
+; X86-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; X86-AVX2-NEXT: vpsrldq {{.*#+}} xmm5 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX2-NEXT: vpextrw $0, %xmm5, %esi
+; X86-AVX2-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX2-NEXT: vpsrldq {{.*#+}} xmm5 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX2-NEXT: vpextrw $0, %xmm5, %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm5
+; X86-AVX2-NEXT: movl 12(%ecx), %esi
+; X86-AVX2-NEXT: andl %edx, %esi
+; X86-AVX2-NEXT: movzwl 12(%eax), %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm6
+; X86-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; X86-AVX2-NEXT: vinserti128 $1, %xmm4, %ymm5, %ymm4
+; X86-AVX2-NEXT: vpsrldq {{.*#+}} xmm5 = xmm2[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX2-NEXT: vpextrw $0, %xmm5, %esi
+; X86-AVX2-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX2-NEXT: vpsrldq {{.*#+}} xmm5 = xmm3[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX2-NEXT: vpextrw $0, %xmm5, %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm5
+; X86-AVX2-NEXT: movl 24(%ecx), %esi
+; X86-AVX2-NEXT: andl %edx, %esi
+; X86-AVX2-NEXT: movzwl 24(%eax), %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm6
+; X86-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; X86-AVX2-NEXT: vpsrldq {{.*#+}} xmm6 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX2-NEXT: vpextrw $0, %xmm6, %esi
+; X86-AVX2-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX2-NEXT: vpsrldq {{.*#+}} xmm6 = xmm1[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX2-NEXT: vpextrw $0, %xmm6, %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm6
+; X86-AVX2-NEXT: movl 8(%ecx), %esi
+; X86-AVX2-NEXT: andl %edx, %esi
+; X86-AVX2-NEXT: movzwl 8(%eax), %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm7
+; X86-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
+; X86-AVX2-NEXT: vinserti128 $1, %xmm5, %ymm6, %ymm5
+; X86-AVX2-NEXT: vpunpckldq {{.*#+}} ymm4 = ymm5[0],ymm4[0],ymm5[1],ymm4[1],ymm5[4],ymm4[4],ymm5[5],ymm4[5]
+; X86-AVX2-NEXT: vpextrw $0, %xmm2, %esi
+; X86-AVX2-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX2-NEXT: vpextrw $0, %xmm3, %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm5
+; X86-AVX2-NEXT: vpsrld $16, %xmm2, %xmm6
+; X86-AVX2-NEXT: vpextrw $0, %xmm6, %esi
+; X86-AVX2-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX2-NEXT: vpsrld $16, %xmm3, %xmm6
+; X86-AVX2-NEXT: vpextrw $0, %xmm6, %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm6
+; X86-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
+; X86-AVX2-NEXT: vpextrw $0, %xmm0, %esi
+; X86-AVX2-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX2-NEXT: vpextrw $0, %xmm1, %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm6
+; X86-AVX2-NEXT: vpsrld $16, %xmm0, %xmm7
+; X86-AVX2-NEXT: vpextrw $0, %xmm7, %esi
+; X86-AVX2-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX2-NEXT: vpsrld $16, %xmm1, %xmm7
+; X86-AVX2-NEXT: vpextrw $0, %xmm7, %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm7
+; X86-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
+; X86-AVX2-NEXT: vinserti128 $1, %xmm5, %ymm6, %ymm5
+; X86-AVX2-NEXT: vpsrlq $48, %xmm2, %xmm2
+; X86-AVX2-NEXT: vpextrw $0, %xmm2, %esi
+; X86-AVX2-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX2-NEXT: vpsrlq $48, %xmm3, %xmm2
+; X86-AVX2-NEXT: vpextrw $0, %xmm2, %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm2
+; X86-AVX2-NEXT: movl 20(%ecx), %esi
+; X86-AVX2-NEXT: andl %edx, %esi
+; X86-AVX2-NEXT: movzwl 20(%eax), %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm3
+; X86-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; X86-AVX2-NEXT: vpsrlq $48, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpextrw $0, %xmm0, %esi
+; X86-AVX2-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX2-NEXT: vpsrlq $48, %xmm1, %xmm0
+; X86-AVX2-NEXT: vpextrw $0, %xmm0, %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm0
+; X86-AVX2-NEXT: movl 4(%ecx), %esi
+; X86-AVX2-NEXT: andl %edx, %esi
+; X86-AVX2-NEXT: movzwl 4(%eax), %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm1
+; X86-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; X86-AVX2-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm0
+; X86-AVX2-NEXT: vpunpckldq {{.*#+}} ymm0 = ymm5[0],ymm0[0],ymm5[1],ymm0[1],ymm5[4],ymm0[4],ymm5[5],ymm0[5]
+; X86-AVX2-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm4[0],ymm0[2],ymm4[2]
+; X86-AVX2-NEXT: vmovdqu %ymm0, (%esp) # 32-byte Spill
+; X86-AVX2-NEXT: vmovdqa 48(%ecx), %xmm1
+; X86-AVX2-NEXT: vpsrldq {{.*#+}} xmm2 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX2-NEXT: vpextrw $0, %xmm2, %esi
+; X86-AVX2-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX2-NEXT: vmovdqa 48(%eax), %xmm2
+; X86-AVX2-NEXT: vpsrldq {{.*#+}} xmm3 = xmm2[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX2-NEXT: vpextrw $0, %xmm3, %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm3
+; X86-AVX2-NEXT: movl 60(%ecx), %esi
+; X86-AVX2-NEXT: andl %edx, %esi
+; X86-AVX2-NEXT: movzwl 60(%eax), %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm4
+; X86-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; X86-AVX2-NEXT: vmovdqa 32(%ecx), %xmm3
+; X86-AVX2-NEXT: vpsrldq {{.*#+}} xmm4 = xmm3[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX2-NEXT: vpextrw $0, %xmm4, %esi
+; X86-AVX2-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX2-NEXT: vmovdqa 32(%eax), %xmm4
+; X86-AVX2-NEXT: vpsrldq {{.*#+}} xmm6 = xmm4[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX2-NEXT: vpextrw $0, %xmm6, %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm6
+; X86-AVX2-NEXT: movl 44(%ecx), %esi
+; X86-AVX2-NEXT: andl %edx, %esi
+; X86-AVX2-NEXT: movzwl 44(%eax), %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm7
+; X86-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
+; X86-AVX2-NEXT: vinserti128 $1, %xmm5, %ymm6, %ymm0
+; X86-AVX2-NEXT: vpsrldq {{.*#+}} xmm6 = xmm1[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX2-NEXT: vpextrw $0, %xmm6, %esi
+; X86-AVX2-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX2-NEXT: vpsrldq {{.*#+}} xmm6 = xmm2[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX2-NEXT: vpextrw $0, %xmm6, %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm6
+; X86-AVX2-NEXT: movl 56(%ecx), %esi
+; X86-AVX2-NEXT: andl %edx, %esi
+; X86-AVX2-NEXT: movzwl 56(%eax), %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm7
+; X86-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
+; X86-AVX2-NEXT: vpsrldq {{.*#+}} xmm7 = xmm3[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX2-NEXT: vpextrw $0, %xmm7, %esi
+; X86-AVX2-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX2-NEXT: vpsrldq {{.*#+}} xmm7 = xmm4[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX2-NEXT: vpextrw $0, %xmm7, %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm7
+; X86-AVX2-NEXT: movl 40(%ecx), %esi
+; X86-AVX2-NEXT: andl %edx, %esi
+; X86-AVX2-NEXT: movzwl 40(%eax), %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm5
+; X86-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm7[0],xmm5[1],xmm7[1],xmm5[2],xmm7[2],xmm5[3],xmm7[3]
+; X86-AVX2-NEXT: vinserti128 $1, %xmm6, %ymm5, %ymm5
+; X86-AVX2-NEXT: vpunpckldq {{.*#+}} ymm5 = ymm5[0],ymm0[0],ymm5[1],ymm0[1],ymm5[4],ymm0[4],ymm5[5],ymm0[5]
+; X86-AVX2-NEXT: vpextrw $0, %xmm1, %esi
+; X86-AVX2-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX2-NEXT: vpextrw $0, %xmm2, %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm6
+; X86-AVX2-NEXT: vpsrld $16, %xmm1, %xmm0
+; X86-AVX2-NEXT: vpextrw $0, %xmm0, %esi
+; X86-AVX2-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX2-NEXT: vpsrld $16, %xmm2, %xmm0
+; X86-AVX2-NEXT: vpextrw $0, %xmm0, %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm0[0],xmm6[1],xmm0[1],xmm6[2],xmm0[2],xmm6[3],xmm0[3]
+; X86-AVX2-NEXT: vpextrw $0, %xmm3, %edi
+; X86-AVX2-NEXT: andl $-32768, %edi # imm = 0x8000
+; X86-AVX2-NEXT: vpextrw $0, %xmm4, %esi
+; X86-AVX2-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %edi, %esi
+; X86-AVX2-NEXT: vpsrld $16, %xmm3, %xmm0
+; X86-AVX2-NEXT: vpextrw $0, %xmm0, %edi
+; X86-AVX2-NEXT: vpsrld $16, %xmm4, %xmm0
+; X86-AVX2-NEXT: vpextrw $0, %xmm0, %ebx
+; X86-AVX2-NEXT: vpinsrw $0, %esi, %xmm0, %xmm0
+; X86-AVX2-NEXT: andl $-32768, %edi # imm = 0x8000
+; X86-AVX2-NEXT: andl $32767, %ebx # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %edi, %ebx
+; X86-AVX2-NEXT: vpinsrw $0, %ebx, %xmm0, %xmm7
+; X86-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm7[0],xmm0[1],xmm7[1],xmm0[2],xmm7[2],xmm0[3],xmm7[3]
+; X86-AVX2-NEXT: vinserti128 $1, %xmm6, %ymm0, %ymm6
+; X86-AVX2-NEXT: vpsrlq $48, %xmm1, %xmm0
+; X86-AVX2-NEXT: vpextrw $0, %xmm0, %esi
+; X86-AVX2-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX2-NEXT: vpsrlq $48, %xmm2, %xmm0
+; X86-AVX2-NEXT: vpextrw $0, %xmm0, %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm0
+; X86-AVX2-NEXT: movl 52(%ecx), %esi
+; X86-AVX2-NEXT: andl %edx, %esi
+; X86-AVX2-NEXT: movzwl 52(%eax), %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm1
+; X86-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; X86-AVX2-NEXT: vpsrlq $48, %xmm3, %xmm0
+; X86-AVX2-NEXT: vpextrw $0, %xmm0, %esi
+; X86-AVX2-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX2-NEXT: vpsrlq $48, %xmm4, %xmm0
+; X86-AVX2-NEXT: vpextrw $0, %xmm0, %edi
+; X86-AVX2-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %esi, %edi
+; X86-AVX2-NEXT: vpinsrw $0, %edi, %xmm0, %xmm0
+; X86-AVX2-NEXT: andl 36(%ecx), %edx
+; X86-AVX2-NEXT: movzwl 36(%eax), %eax
+; X86-AVX2-NEXT: andl $32767, %eax # imm = 0x7FFF
+; X86-AVX2-NEXT: orl %edx, %eax
+; X86-AVX2-NEXT: vpinsrw $0, %eax, %xmm0, %xmm2
+; X86-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
+; X86-AVX2-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0
+; X86-AVX2-NEXT: vpunpckldq {{.*#+}} ymm0 = ymm6[0],ymm0[0],ymm6[1],ymm0[1],ymm6[4],ymm0[4],ymm6[5],ymm0[5]
+; X86-AVX2-NEXT: vpunpcklqdq {{.*#+}} ymm1 = ymm0[0],ymm5[0],ymm0[2],ymm5[2]
+; X86-AVX2-NEXT: vmovups (%esp), %ymm0 # 32-byte Reload
+; X86-AVX2-NEXT: addl $32, %esp
+; X86-AVX2-NEXT: popl %esi
+; X86-AVX2-NEXT: popl %edi
+; X86-AVX2-NEXT: popl %ebx
+; X86-AVX2-NEXT: retl
+;
+; X86-AVX512VL-LABEL: fcopysign_v32f16:
+; X86-AVX512VL: # %bb.0:
+; X86-AVX512VL-NEXT: pushl %ebp
+; X86-AVX512VL-NEXT: pushl %ebx
+; X86-AVX512VL-NEXT: pushl %edi
+; X86-AVX512VL-NEXT: pushl %esi
+; X86-AVX512VL-NEXT: subl $400, %esp # imm = 0x190
+; X86-AVX512VL-NEXT: movl {{[0-9]+}}(%esp), %eax
+; X86-AVX512VL-NEXT: movl {{[0-9]+}}(%esp), %ecx
+; X86-AVX512VL-NEXT: vmovdqa 48(%ecx), %xmm0
+; X86-AVX512VL-NEXT: vmovdqu %xmm0, {{[-0-9]+}}(%e{{[sb]}}p) # 16-byte Spill
+; X86-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm0, %edx
+; X86-AVX512VL-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX512VL-NEXT: vmovdqa 48(%eax), %xmm0
+; X86-AVX512VL-NEXT: vmovdqu %xmm0, {{[-0-9]+}}(%e{{[sb]}}p) # 16-byte Spill
+; X86-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm0, %esi
+; X86-AVX512VL-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %edx, %esi
+; X86-AVX512VL-NEXT: vpinsrw $0, %esi, %xmm0, %xmm0
+; X86-AVX512VL-NEXT: movl $-32768, %edx # imm = 0x8000
+; X86-AVX512VL-NEXT: movl 60(%ecx), %esi
+; X86-AVX512VL-NEXT: andl %edx, %esi
+; X86-AVX512VL-NEXT: movzwl 60(%eax), %edi
+; X86-AVX512VL-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %esi, %edi
+; X86-AVX512VL-NEXT: vpinsrw $0, %edi, %xmm0, %xmm2
+; X86-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
+; X86-AVX512VL-NEXT: vmovdqa 32(%ecx), %xmm1
+; X86-AVX512VL-NEXT: vmovdqu %xmm1, (%esp) # 16-byte Spill
+; X86-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm2 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm2, %esi
+; X86-AVX512VL-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX512VL-NEXT: vmovdqa 32(%eax), %xmm1
+; X86-AVX512VL-NEXT: vmovdqu %xmm1, {{[-0-9]+}}(%e{{[sb]}}p) # 16-byte Spill
+; X86-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm2 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm2, %edi
+; X86-AVX512VL-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %esi, %edi
+; X86-AVX512VL-NEXT: vpinsrw $0, %edi, %xmm0, %xmm2
+; X86-AVX512VL-NEXT: movl 44(%ecx), %esi
+; X86-AVX512VL-NEXT: andl %edx, %esi
+; X86-AVX512VL-NEXT: movzwl 44(%eax), %edi
+; X86-AVX512VL-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %esi, %edi
+; X86-AVX512VL-NEXT: vpinsrw $0, %edi, %xmm0, %xmm6
+; X86-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm6[0],xmm2[0],xmm6[1],xmm2[1],xmm6[2],xmm2[2],xmm6[3],xmm2[3]
+; X86-AVX512VL-NEXT: vinserti128 $1, %xmm0, %ymm2, %ymm0
+; X86-AVX512VL-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%e{{[sb]}}p) # 32-byte Spill
+; X86-AVX512VL-NEXT: vmovdqa 16(%ecx), %xmm5
+; X86-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm0 = xmm5[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm0, %esi
+; X86-AVX512VL-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX512VL-NEXT: vmovdqa 16(%eax), %xmm0
+; X86-AVX512VL-NEXT: vmovdqu %xmm0, {{[-0-9]+}}(%e{{[sb]}}p) # 16-byte Spill
+; X86-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm0, %edi
+; X86-AVX512VL-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %esi, %edi
+; X86-AVX512VL-NEXT: vpinsrw $0, %edi, %xmm0, %xmm0
+; X86-AVX512VL-NEXT: movl 28(%ecx), %esi
+; X86-AVX512VL-NEXT: andl %edx, %esi
+; X86-AVX512VL-NEXT: movzwl 28(%eax), %edi
+; X86-AVX512VL-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %esi, %edi
+; X86-AVX512VL-NEXT: vpinsrw $0, %edi, %xmm0, %xmm7
+; X86-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm7[0],xmm0[0],xmm7[1],xmm0[1],xmm7[2],xmm0[2],xmm7[3],xmm0[3]
+; X86-AVX512VL-NEXT: vmovdqa (%ecx), %xmm4
+; X86-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm1 = xmm4[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VL-NEXT: vmovdqu %xmm4, {{[-0-9]+}}(%e{{[sb]}}p) # 16-byte Spill
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm1, %edi
+; X86-AVX512VL-NEXT: vmovdqa (%eax), %xmm2
+; X86-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm1 = xmm2[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VL-NEXT: vmovdqu %xmm2, {{[-0-9]+}}(%e{{[sb]}}p) # 16-byte Spill
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm1, %esi
+; X86-AVX512VL-NEXT: andl $-32768, %edi # imm = 0x8000
+; X86-AVX512VL-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %edi, %esi
+; X86-AVX512VL-NEXT: movl 12(%ecx), %edi
+; X86-AVX512VL-NEXT: andl %edx, %edi
+; X86-AVX512VL-NEXT: movzwl 12(%eax), %ebx
+; X86-AVX512VL-NEXT: andl $32767, %ebx # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %edi, %ebx
+; X86-AVX512VL-NEXT: vpinsrw $0, %esi, %xmm0, %xmm1
+; X86-AVX512VL-NEXT: vpinsrw $0, %ebx, %xmm0, %xmm3
+; X86-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
+; X86-AVX512VL-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
+; X86-AVX512VL-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%e{{[sb]}}p) # 64-byte Spill
+; X86-AVX512VL-NEXT: vmovdqu {{[-0-9]+}}(%e{{[sb]}}p), %xmm6 # 16-byte Reload
+; X86-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm0 = xmm6[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm0, %edi
+; X86-AVX512VL-NEXT: vmovdqu {{[-0-9]+}}(%e{{[sb]}}p), %xmm7 # 16-byte Reload
+; X86-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm0 = xmm7[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm0, %esi
+; X86-AVX512VL-NEXT: andl $-32768, %edi # imm = 0x8000
+; X86-AVX512VL-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %edi, %esi
+; X86-AVX512VL-NEXT: movl 56(%ecx), %edi
+; X86-AVX512VL-NEXT: andl %edx, %edi
+; X86-AVX512VL-NEXT: movzwl 56(%eax), %ebx
+; X86-AVX512VL-NEXT: andl $32767, %ebx # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %edi, %ebx
+; X86-AVX512VL-NEXT: vpinsrw $0, %esi, %xmm0, %xmm0
+; X86-AVX512VL-NEXT: vpinsrw $0, %ebx, %xmm0, %xmm1
+; X86-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; X86-AVX512VL-NEXT: vmovdqu (%esp), %xmm1 # 16-byte Reload
+; X86-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm1 = xmm1[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm1, %edi
+; X86-AVX512VL-NEXT: vmovdqu {{[-0-9]+}}(%e{{[sb]}}p), %xmm1 # 16-byte Reload
+; X86-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm1 = xmm1[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm1, %esi
+; X86-AVX512VL-NEXT: andl $-32768, %edi # imm = 0x8000
+; X86-AVX512VL-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %edi, %esi
+; X86-AVX512VL-NEXT: movl 40(%ecx), %edi
+; X86-AVX512VL-NEXT: andl %edx, %edi
+; X86-AVX512VL-NEXT: movzwl 40(%eax), %ebx
+; X86-AVX512VL-NEXT: andl $32767, %ebx # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %edi, %ebx
+; X86-AVX512VL-NEXT: vpinsrw $0, %esi, %xmm0, %xmm1
+; X86-AVX512VL-NEXT: vpinsrw $0, %ebx, %xmm0, %xmm3
+; X86-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
+; X86-AVX512VL-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
+; X86-AVX512VL-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%e{{[sb]}}p) # 32-byte Spill
+; X86-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm0 = xmm5[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm0, %edi
+; X86-AVX512VL-NEXT: vmovdqu {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 # 16-byte Reload
+; X86-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm0, %esi
+; X86-AVX512VL-NEXT: andl $-32768, %edi # imm = 0x8000
+; X86-AVX512VL-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %edi, %esi
+; X86-AVX512VL-NEXT: movl 24(%ecx), %edi
+; X86-AVX512VL-NEXT: andl %edx, %edi
+; X86-AVX512VL-NEXT: movzwl 24(%eax), %ebx
+; X86-AVX512VL-NEXT: andl $32767, %ebx # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %edi, %ebx
+; X86-AVX512VL-NEXT: vpinsrw $0, %esi, %xmm0, %xmm0
+; X86-AVX512VL-NEXT: vpinsrw $0, %ebx, %xmm0, %xmm1
+; X86-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; X86-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm1 = xmm4[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm1, %edi
+; X86-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm1 = xmm2[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm1, %esi
+; X86-AVX512VL-NEXT: andl $-32768, %edi # imm = 0x8000
+; X86-AVX512VL-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %edi, %esi
+; X86-AVX512VL-NEXT: movl 8(%ecx), %edi
+; X86-AVX512VL-NEXT: andl %edx, %edi
+; X86-AVX512VL-NEXT: movzwl 8(%eax), %ebx
+; X86-AVX512VL-NEXT: andl $32767, %ebx # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %edi, %ebx
+; X86-AVX512VL-NEXT: vpinsrw $0, %esi, %xmm0, %xmm1
+; X86-AVX512VL-NEXT: vpinsrw $0, %ebx, %xmm0, %xmm3
+; X86-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
+; X86-AVX512VL-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
+; X86-AVX512VL-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%e{{[sb]}}p) # 64-byte Spill
+; X86-AVX512VL-NEXT: vpsrld $16, %xmm6, %xmm0
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm0, %esi
+; X86-AVX512VL-NEXT: vpsrld $16, %xmm7, %xmm0
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm0, %edi
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm6, %ebx
+; X86-AVX512VL-NEXT: andl $-32768, %ebx # imm = 0x8000
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm7, %ebp
+; X86-AVX512VL-NEXT: andl $32767, %ebp # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %ebx, %ebp
+; X86-AVX512VL-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX512VL-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %esi, %edi
+; X86-AVX512VL-NEXT: vpinsrw $0, %ebp, %xmm0, %xmm0
+; X86-AVX512VL-NEXT: vpinsrw $0, %edi, %xmm0, %xmm1
+; X86-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; X86-AVX512VL-NEXT: vmovdqu (%esp), %xmm3 # 16-byte Reload
+; X86-AVX512VL-NEXT: vpsrld $16, %xmm3, %xmm1
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm1, %esi
+; X86-AVX512VL-NEXT: vmovdqu {{[-0-9]+}}(%e{{[sb]}}p), %xmm2 # 16-byte Reload
+; X86-AVX512VL-NEXT: vpsrld $16, %xmm2, %xmm1
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm1, %edi
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm3, %ebx
+; X86-AVX512VL-NEXT: andl $-32768, %ebx # imm = 0x8000
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm2, %ebp
+; X86-AVX512VL-NEXT: andl $32767, %ebp # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %ebx, %ebp
+; X86-AVX512VL-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX512VL-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %esi, %edi
+; X86-AVX512VL-NEXT: vpinsrw $0, %ebp, %xmm0, %xmm1
+; X86-AVX512VL-NEXT: vpinsrw $0, %edi, %xmm0, %xmm3
+; X86-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
+; X86-AVX512VL-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
+; X86-AVX512VL-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%e{{[sb]}}p) # 32-byte Spill
+; X86-AVX512VL-NEXT: vpsrld $16, %xmm5, %xmm0
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm0, %esi
+; X86-AVX512VL-NEXT: vmovdqu {{[-0-9]+}}(%e{{[sb]}}p), %xmm6 # 16-byte Reload
+; X86-AVX512VL-NEXT: vpsrld $16, %xmm6, %xmm0
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm0, %edi
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm5, %ebx
+; X86-AVX512VL-NEXT: andl $-32768, %ebx # imm = 0x8000
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm6, %ebp
+; X86-AVX512VL-NEXT: andl $32767, %ebp # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %ebx, %ebp
+; X86-AVX512VL-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX512VL-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %esi, %edi
+; X86-AVX512VL-NEXT: vpinsrw $0, %ebp, %xmm0, %xmm0
+; X86-AVX512VL-NEXT: vpinsrw $0, %edi, %xmm0, %xmm1
+; X86-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; X86-AVX512VL-NEXT: vmovdqu {{[-0-9]+}}(%e{{[sb]}}p), %xmm7 # 16-byte Reload
+; X86-AVX512VL-NEXT: vpsrld $16, %xmm7, %xmm1
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm1, %esi
+; X86-AVX512VL-NEXT: vmovdqu {{[-0-9]+}}(%e{{[sb]}}p), %xmm4 # 16-byte Reload
+; X86-AVX512VL-NEXT: vpsrld $16, %xmm4, %xmm1
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm1, %edi
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm7, %ebx
+; X86-AVX512VL-NEXT: andl $-32768, %ebx # imm = 0x8000
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm4, %ebp
+; X86-AVX512VL-NEXT: andl $32767, %ebp # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %ebx, %ebp
+; X86-AVX512VL-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX512VL-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %esi, %edi
+; X86-AVX512VL-NEXT: vpinsrw $0, %ebp, %xmm0, %xmm1
+; X86-AVX512VL-NEXT: vpinsrw $0, %edi, %xmm0, %xmm3
+; X86-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
+; X86-AVX512VL-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
+; X86-AVX512VL-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%e{{[sb]}}p) # 64-byte Spill
+; X86-AVX512VL-NEXT: vpsrlq $48, {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm0, %edi
+; X86-AVX512VL-NEXT: vpsrlq $48, {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm0, %esi
+; X86-AVX512VL-NEXT: andl $-32768, %edi # imm = 0x8000
+; X86-AVX512VL-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %edi, %esi
+; X86-AVX512VL-NEXT: movl 52(%ecx), %edi
+; X86-AVX512VL-NEXT: andl %edx, %edi
+; X86-AVX512VL-NEXT: movzwl 52(%eax), %ebx
+; X86-AVX512VL-NEXT: andl $32767, %ebx # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %edi, %ebx
+; X86-AVX512VL-NEXT: vpinsrw $0, %esi, %xmm0, %xmm0
+; X86-AVX512VL-NEXT: vpinsrw $0, %ebx, %xmm0, %xmm1
+; X86-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; X86-AVX512VL-NEXT: vpsrlq $48, (%esp), %xmm1 # 16-byte Folded Reload
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm1, %edi
+; X86-AVX512VL-NEXT: vpsrlq $48, %xmm2, %xmm1
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm1, %esi
+; X86-AVX512VL-NEXT: andl $-32768, %edi # imm = 0x8000
+; X86-AVX512VL-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %edi, %esi
+; X86-AVX512VL-NEXT: movl 36(%ecx), %edi
+; X86-AVX512VL-NEXT: andl %edx, %edi
+; X86-AVX512VL-NEXT: movzwl 36(%eax), %ebx
+; X86-AVX512VL-NEXT: andl $32767, %ebx # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %edi, %ebx
+; X86-AVX512VL-NEXT: vpinsrw $0, %esi, %xmm0, %xmm1
+; X86-AVX512VL-NEXT: vpinsrw $0, %ebx, %xmm0, %xmm3
+; X86-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
+; X86-AVX512VL-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm1
+; X86-AVX512VL-NEXT: vpsrlq $48, %xmm5, %xmm0
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm0, %edi
+; X86-AVX512VL-NEXT: vpsrlq $48, %xmm6, %xmm0
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm0, %esi
+; X86-AVX512VL-NEXT: andl $-32768, %edi # imm = 0x8000
+; X86-AVX512VL-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %edi, %esi
+; X86-AVX512VL-NEXT: movl 20(%ecx), %edi
+; X86-AVX512VL-NEXT: andl %edx, %edi
+; X86-AVX512VL-NEXT: movzwl 20(%eax), %ebx
+; X86-AVX512VL-NEXT: andl $32767, %ebx # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %edi, %ebx
+; X86-AVX512VL-NEXT: vpinsrw $0, %esi, %xmm0, %xmm0
+; X86-AVX512VL-NEXT: vpinsrw $0, %ebx, %xmm0, %xmm2
+; X86-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
+; X86-AVX512VL-NEXT: vpsrlq $48, %xmm7, %xmm2
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm2, %edi
+; X86-AVX512VL-NEXT: vpsrlq $48, %xmm4, %xmm2
+; X86-AVX512VL-NEXT: vpextrw $0, %xmm2, %esi
+; X86-AVX512VL-NEXT: andl $-32768, %edi # imm = 0x8000
+; X86-AVX512VL-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %edi, %esi
+; X86-AVX512VL-NEXT: andl 4(%ecx), %edx
+; X86-AVX512VL-NEXT: movzwl 4(%eax), %eax
+; X86-AVX512VL-NEXT: andl $32767, %eax # imm = 0x7FFF
+; X86-AVX512VL-NEXT: orl %edx, %eax
+; X86-AVX512VL-NEXT: vpinsrw $0, %esi, %xmm0, %xmm2
+; X86-AVX512VL-NEXT: vpinsrw $0, %eax, %xmm0, %xmm3
+; X86-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; X86-AVX512VL-NEXT: vinserti128 $1, %xmm0, %ymm2, %ymm0
+; X86-AVX512VL-NEXT: vmovdqu64 {{[-0-9]+}}(%e{{[sb]}}p), %zmm2 # 64-byte Reload
+; X86-AVX512VL-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%e{{[sb]}}p), %zmm2, %zmm2 # 32-byte Folded Reload
+; X86-AVX512VL-NEXT: vmovdqu64 {{[-0-9]+}}(%e{{[sb]}}p), %zmm3 # 64-byte Reload
+; X86-AVX512VL-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%e{{[sb]}}p), %zmm3, %zmm3 # 32-byte Folded Reload
+; X86-AVX512VL-NEXT: vpunpckldq {{.*#+}} zmm2 = zmm3[0],zmm2[0],zmm3[1],zmm2[1],zmm3[4],zmm2[4],zmm3[5],zmm2[5],zmm3[8],zmm2[8],zmm3[9],zmm2[9],zmm3[12],zmm2[12],zmm3[13],zmm2[13]
+; X86-AVX512VL-NEXT: vmovdqu64 {{[-0-9]+}}(%e{{[sb]}}p), %zmm3 # 64-byte Reload
+; X86-AVX512VL-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%e{{[sb]}}p), %zmm3, %zmm3 # 32-byte Folded Reload
+; X86-AVX512VL-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
+; X86-AVX512VL-NEXT: vpunpckldq {{.*#+}} zmm0 = zmm3[0],zmm0[0],zmm3[1],zmm0[1],zmm3[4],zmm0[4],zmm3[5],zmm0[5],zmm3[8],zmm0[8],zmm3[9],zmm0[9],zmm3[12],zmm0[12],zmm3[13],zmm0[13]
+; X86-AVX512VL-NEXT: vpunpcklqdq {{.*#+}} zmm0 = zmm0[0],zmm2[0],zmm0[2],zmm2[2],zmm0[4],zmm2[4],zmm0[6],zmm2[6]
+; X86-AVX512VL-NEXT: addl $400, %esp # imm = 0x190
+; X86-AVX512VL-NEXT: popl %esi
+; X86-AVX512VL-NEXT: popl %edi
+; X86-AVX512VL-NEXT: popl %ebx
+; X86-AVX512VL-NEXT: popl %ebp
+; X86-AVX512VL-NEXT: retl
+;
+; X86-AVX512FP16-LABEL: fcopysign_v32f16:
+; X86-AVX512FP16: # %bb.0:
+; X86-AVX512FP16-NEXT: movl {{[0-9]+}}(%esp), %eax
+; X86-AVX512FP16-NEXT: movl {{[0-9]+}}(%esp), %ecx
+; X86-AVX512FP16-NEXT: vmovdqu64 (%ecx), %zmm1
+; X86-AVX512FP16-NEXT: vpbroadcastw {{.*#+}} zmm0 = [NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN]
+; X86-AVX512FP16-NEXT: vpternlogq $202, (%eax), %zmm1, %zmm0
+; X86-AVX512FP16-NEXT: retl
+;
+; X86-AVX512VLDQ-LABEL: fcopysign_v32f16:
+; X86-AVX512VLDQ: # %bb.0:
+; X86-AVX512VLDQ-NEXT: pushl %ebp
+; X86-AVX512VLDQ-NEXT: pushl %ebx
+; X86-AVX512VLDQ-NEXT: pushl %edi
+; X86-AVX512VLDQ-NEXT: pushl %esi
+; X86-AVX512VLDQ-NEXT: subl $400, %esp # imm = 0x190
+; X86-AVX512VLDQ-NEXT: movl {{[0-9]+}}(%esp), %eax
+; X86-AVX512VLDQ-NEXT: movl {{[0-9]+}}(%esp), %ecx
+; X86-AVX512VLDQ-NEXT: vmovdqa 48(%ecx), %xmm0
+; X86-AVX512VLDQ-NEXT: vmovdqu %xmm0, {{[-0-9]+}}(%e{{[sb]}}p) # 16-byte Spill
+; X86-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm0, %edx
+; X86-AVX512VLDQ-NEXT: andl $-32768, %edx # imm = 0x8000
+; X86-AVX512VLDQ-NEXT: vmovdqa 48(%eax), %xmm0
+; X86-AVX512VLDQ-NEXT: vmovdqu %xmm0, {{[-0-9]+}}(%e{{[sb]}}p) # 16-byte Spill
+; X86-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm0, %esi
+; X86-AVX512VLDQ-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %edx, %esi
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %esi, %xmm0, %xmm0
+; X86-AVX512VLDQ-NEXT: movl $-32768, %edx # imm = 0x8000
+; X86-AVX512VLDQ-NEXT: movl 60(%ecx), %esi
+; X86-AVX512VLDQ-NEXT: andl %edx, %esi
+; X86-AVX512VLDQ-NEXT: movzwl 60(%eax), %edi
+; X86-AVX512VLDQ-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %esi, %edi
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %edi, %xmm0, %xmm2
+; X86-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
+; X86-AVX512VLDQ-NEXT: vmovdqa 32(%ecx), %xmm1
+; X86-AVX512VLDQ-NEXT: vmovdqu %xmm1, (%esp) # 16-byte Spill
+; X86-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm2 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm2, %esi
+; X86-AVX512VLDQ-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX512VLDQ-NEXT: vmovdqa 32(%eax), %xmm1
+; X86-AVX512VLDQ-NEXT: vmovdqu %xmm1, {{[-0-9]+}}(%e{{[sb]}}p) # 16-byte Spill
+; X86-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm2 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm2, %edi
+; X86-AVX512VLDQ-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %esi, %edi
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %edi, %xmm0, %xmm2
+; X86-AVX512VLDQ-NEXT: movl 44(%ecx), %esi
+; X86-AVX512VLDQ-NEXT: andl %edx, %esi
+; X86-AVX512VLDQ-NEXT: movzwl 44(%eax), %edi
+; X86-AVX512VLDQ-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %esi, %edi
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %edi, %xmm0, %xmm6
+; X86-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm6[0],xmm2[0],xmm6[1],xmm2[1],xmm6[2],xmm2[2],xmm6[3],xmm2[3]
+; X86-AVX512VLDQ-NEXT: vinserti128 $1, %xmm0, %ymm2, %ymm0
+; X86-AVX512VLDQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%e{{[sb]}}p) # 32-byte Spill
+; X86-AVX512VLDQ-NEXT: vmovdqa 16(%ecx), %xmm5
+; X86-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm0 = xmm5[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm0, %esi
+; X86-AVX512VLDQ-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX512VLDQ-NEXT: vmovdqa 16(%eax), %xmm0
+; X86-AVX512VLDQ-NEXT: vmovdqu %xmm0, {{[-0-9]+}}(%e{{[sb]}}p) # 16-byte Spill
+; X86-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm0, %edi
+; X86-AVX512VLDQ-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %esi, %edi
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %edi, %xmm0, %xmm0
+; X86-AVX512VLDQ-NEXT: movl 28(%ecx), %esi
+; X86-AVX512VLDQ-NEXT: andl %edx, %esi
+; X86-AVX512VLDQ-NEXT: movzwl 28(%eax), %edi
+; X86-AVX512VLDQ-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %esi, %edi
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %edi, %xmm0, %xmm7
+; X86-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm7[0],xmm0[0],xmm7[1],xmm0[1],xmm7[2],xmm0[2],xmm7[3],xmm0[3]
+; X86-AVX512VLDQ-NEXT: vmovdqa (%ecx), %xmm4
+; X86-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm1 = xmm4[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VLDQ-NEXT: vmovdqu %xmm4, {{[-0-9]+}}(%e{{[sb]}}p) # 16-byte Spill
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm1, %edi
+; X86-AVX512VLDQ-NEXT: vmovdqa (%eax), %xmm2
+; X86-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm1 = xmm2[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VLDQ-NEXT: vmovdqu %xmm2, {{[-0-9]+}}(%e{{[sb]}}p) # 16-byte Spill
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm1, %esi
+; X86-AVX512VLDQ-NEXT: andl $-32768, %edi # imm = 0x8000
+; X86-AVX512VLDQ-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %edi, %esi
+; X86-AVX512VLDQ-NEXT: movl 12(%ecx), %edi
+; X86-AVX512VLDQ-NEXT: andl %edx, %edi
+; X86-AVX512VLDQ-NEXT: movzwl 12(%eax), %ebx
+; X86-AVX512VLDQ-NEXT: andl $32767, %ebx # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %edi, %ebx
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %esi, %xmm0, %xmm1
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %ebx, %xmm0, %xmm3
+; X86-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
+; X86-AVX512VLDQ-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
+; X86-AVX512VLDQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%e{{[sb]}}p) # 64-byte Spill
+; X86-AVX512VLDQ-NEXT: vmovdqu {{[-0-9]+}}(%e{{[sb]}}p), %xmm6 # 16-byte Reload
+; X86-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm0 = xmm6[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm0, %edi
+; X86-AVX512VLDQ-NEXT: vmovdqu {{[-0-9]+}}(%e{{[sb]}}p), %xmm7 # 16-byte Reload
+; X86-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm0 = xmm7[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm0, %esi
+; X86-AVX512VLDQ-NEXT: andl $-32768, %edi # imm = 0x8000
+; X86-AVX512VLDQ-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %edi, %esi
+; X86-AVX512VLDQ-NEXT: movl 56(%ecx), %edi
+; X86-AVX512VLDQ-NEXT: andl %edx, %edi
+; X86-AVX512VLDQ-NEXT: movzwl 56(%eax), %ebx
+; X86-AVX512VLDQ-NEXT: andl $32767, %ebx # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %edi, %ebx
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %esi, %xmm0, %xmm0
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %ebx, %xmm0, %xmm1
+; X86-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; X86-AVX512VLDQ-NEXT: vmovdqu (%esp), %xmm1 # 16-byte Reload
+; X86-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm1 = xmm1[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm1, %edi
+; X86-AVX512VLDQ-NEXT: vmovdqu {{[-0-9]+}}(%e{{[sb]}}p), %xmm1 # 16-byte Reload
+; X86-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm1 = xmm1[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm1, %esi
+; X86-AVX512VLDQ-NEXT: andl $-32768, %edi # imm = 0x8000
+; X86-AVX512VLDQ-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %edi, %esi
+; X86-AVX512VLDQ-NEXT: movl 40(%ecx), %edi
+; X86-AVX512VLDQ-NEXT: andl %edx, %edi
+; X86-AVX512VLDQ-NEXT: movzwl 40(%eax), %ebx
+; X86-AVX512VLDQ-NEXT: andl $32767, %ebx # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %edi, %ebx
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %esi, %xmm0, %xmm1
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %ebx, %xmm0, %xmm3
+; X86-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
+; X86-AVX512VLDQ-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
+; X86-AVX512VLDQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%e{{[sb]}}p) # 32-byte Spill
+; X86-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm0 = xmm5[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm0, %edi
+; X86-AVX512VLDQ-NEXT: vmovdqu {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 # 16-byte Reload
+; X86-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm0, %esi
+; X86-AVX512VLDQ-NEXT: andl $-32768, %edi # imm = 0x8000
+; X86-AVX512VLDQ-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %edi, %esi
+; X86-AVX512VLDQ-NEXT: movl 24(%ecx), %edi
+; X86-AVX512VLDQ-NEXT: andl %edx, %edi
+; X86-AVX512VLDQ-NEXT: movzwl 24(%eax), %ebx
+; X86-AVX512VLDQ-NEXT: andl $32767, %ebx # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %edi, %ebx
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %esi, %xmm0, %xmm0
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %ebx, %xmm0, %xmm1
+; X86-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; X86-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm1 = xmm4[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm1, %edi
+; X86-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm1 = xmm2[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm1, %esi
+; X86-AVX512VLDQ-NEXT: andl $-32768, %edi # imm = 0x8000
+; X86-AVX512VLDQ-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %edi, %esi
+; X86-AVX512VLDQ-NEXT: movl 8(%ecx), %edi
+; X86-AVX512VLDQ-NEXT: andl %edx, %edi
+; X86-AVX512VLDQ-NEXT: movzwl 8(%eax), %ebx
+; X86-AVX512VLDQ-NEXT: andl $32767, %ebx # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %edi, %ebx
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %esi, %xmm0, %xmm1
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %ebx, %xmm0, %xmm3
+; X86-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
+; X86-AVX512VLDQ-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
+; X86-AVX512VLDQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%e{{[sb]}}p) # 64-byte Spill
+; X86-AVX512VLDQ-NEXT: vpsrld $16, %xmm6, %xmm0
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm0, %esi
+; X86-AVX512VLDQ-NEXT: vpsrld $16, %xmm7, %xmm0
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm0, %edi
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm6, %ebx
+; X86-AVX512VLDQ-NEXT: andl $-32768, %ebx # imm = 0x8000
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm7, %ebp
+; X86-AVX512VLDQ-NEXT: andl $32767, %ebp # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %ebx, %ebp
+; X86-AVX512VLDQ-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX512VLDQ-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %esi, %edi
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %ebp, %xmm0, %xmm0
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %edi, %xmm0, %xmm1
+; X86-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; X86-AVX512VLDQ-NEXT: vmovdqu (%esp), %xmm3 # 16-byte Reload
+; X86-AVX512VLDQ-NEXT: vpsrld $16, %xmm3, %xmm1
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm1, %esi
+; X86-AVX512VLDQ-NEXT: vmovdqu {{[-0-9]+}}(%e{{[sb]}}p), %xmm2 # 16-byte Reload
+; X86-AVX512VLDQ-NEXT: vpsrld $16, %xmm2, %xmm1
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm1, %edi
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm3, %ebx
+; X86-AVX512VLDQ-NEXT: andl $-32768, %ebx # imm = 0x8000
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm2, %ebp
+; X86-AVX512VLDQ-NEXT: andl $32767, %ebp # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %ebx, %ebp
+; X86-AVX512VLDQ-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX512VLDQ-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %esi, %edi
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %ebp, %xmm0, %xmm1
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %edi, %xmm0, %xmm3
+; X86-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
+; X86-AVX512VLDQ-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
+; X86-AVX512VLDQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%e{{[sb]}}p) # 32-byte Spill
+; X86-AVX512VLDQ-NEXT: vpsrld $16, %xmm5, %xmm0
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm0, %esi
+; X86-AVX512VLDQ-NEXT: vmovdqu {{[-0-9]+}}(%e{{[sb]}}p), %xmm6 # 16-byte Reload
+; X86-AVX512VLDQ-NEXT: vpsrld $16, %xmm6, %xmm0
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm0, %edi
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm5, %ebx
+; X86-AVX512VLDQ-NEXT: andl $-32768, %ebx # imm = 0x8000
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm6, %ebp
+; X86-AVX512VLDQ-NEXT: andl $32767, %ebp # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %ebx, %ebp
+; X86-AVX512VLDQ-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX512VLDQ-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %esi, %edi
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %ebp, %xmm0, %xmm0
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %edi, %xmm0, %xmm1
+; X86-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; X86-AVX512VLDQ-NEXT: vmovdqu {{[-0-9]+}}(%e{{[sb]}}p), %xmm7 # 16-byte Reload
+; X86-AVX512VLDQ-NEXT: vpsrld $16, %xmm7, %xmm1
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm1, %esi
+; X86-AVX512VLDQ-NEXT: vmovdqu {{[-0-9]+}}(%e{{[sb]}}p), %xmm4 # 16-byte Reload
+; X86-AVX512VLDQ-NEXT: vpsrld $16, %xmm4, %xmm1
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm1, %edi
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm7, %ebx
+; X86-AVX512VLDQ-NEXT: andl $-32768, %ebx # imm = 0x8000
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm4, %ebp
+; X86-AVX512VLDQ-NEXT: andl $32767, %ebp # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %ebx, %ebp
+; X86-AVX512VLDQ-NEXT: andl $-32768, %esi # imm = 0x8000
+; X86-AVX512VLDQ-NEXT: andl $32767, %edi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %esi, %edi
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %ebp, %xmm0, %xmm1
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %edi, %xmm0, %xmm3
+; X86-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
+; X86-AVX512VLDQ-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0
+; X86-AVX512VLDQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%e{{[sb]}}p) # 64-byte Spill
+; X86-AVX512VLDQ-NEXT: vpsrlq $48, {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm0, %edi
+; X86-AVX512VLDQ-NEXT: vpsrlq $48, {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm0, %esi
+; X86-AVX512VLDQ-NEXT: andl $-32768, %edi # imm = 0x8000
+; X86-AVX512VLDQ-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %edi, %esi
+; X86-AVX512VLDQ-NEXT: movl 52(%ecx), %edi
+; X86-AVX512VLDQ-NEXT: andl %edx, %edi
+; X86-AVX512VLDQ-NEXT: movzwl 52(%eax), %ebx
+; X86-AVX512VLDQ-NEXT: andl $32767, %ebx # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %edi, %ebx
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %esi, %xmm0, %xmm0
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %ebx, %xmm0, %xmm1
+; X86-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; X86-AVX512VLDQ-NEXT: vpsrlq $48, (%esp), %xmm1 # 16-byte Folded Reload
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm1, %edi
+; X86-AVX512VLDQ-NEXT: vpsrlq $48, %xmm2, %xmm1
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm1, %esi
+; X86-AVX512VLDQ-NEXT: andl $-32768, %edi # imm = 0x8000
+; X86-AVX512VLDQ-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %edi, %esi
+; X86-AVX512VLDQ-NEXT: movl 36(%ecx), %edi
+; X86-AVX512VLDQ-NEXT: andl %edx, %edi
+; X86-AVX512VLDQ-NEXT: movzwl 36(%eax), %ebx
+; X86-AVX512VLDQ-NEXT: andl $32767, %ebx # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %edi, %ebx
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %esi, %xmm0, %xmm1
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %ebx, %xmm0, %xmm3
+; X86-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
+; X86-AVX512VLDQ-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm1
+; X86-AVX512VLDQ-NEXT: vpsrlq $48, %xmm5, %xmm0
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm0, %edi
+; X86-AVX512VLDQ-NEXT: vpsrlq $48, %xmm6, %xmm0
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm0, %esi
+; X86-AVX512VLDQ-NEXT: andl $-32768, %edi # imm = 0x8000
+; X86-AVX512VLDQ-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %edi, %esi
+; X86-AVX512VLDQ-NEXT: movl 20(%ecx), %edi
+; X86-AVX512VLDQ-NEXT: andl %edx, %edi
+; X86-AVX512VLDQ-NEXT: movzwl 20(%eax), %ebx
+; X86-AVX512VLDQ-NEXT: andl $32767, %ebx # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %edi, %ebx
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %esi, %xmm0, %xmm0
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %ebx, %xmm0, %xmm2
+; X86-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
+; X86-AVX512VLDQ-NEXT: vpsrlq $48, %xmm7, %xmm2
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm2, %edi
+; X86-AVX512VLDQ-NEXT: vpsrlq $48, %xmm4, %xmm2
+; X86-AVX512VLDQ-NEXT: vpextrw $0, %xmm2, %esi
+; X86-AVX512VLDQ-NEXT: andl $-32768, %edi # imm = 0x8000
+; X86-AVX512VLDQ-NEXT: andl $32767, %esi # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %edi, %esi
+; X86-AVX512VLDQ-NEXT: andl 4(%ecx), %edx
+; X86-AVX512VLDQ-NEXT: movzwl 4(%eax), %eax
+; X86-AVX512VLDQ-NEXT: andl $32767, %eax # imm = 0x7FFF
+; X86-AVX512VLDQ-NEXT: orl %edx, %eax
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %esi, %xmm0, %xmm2
+; X86-AVX512VLDQ-NEXT: vpinsrw $0, %eax, %xmm0, %xmm3
+; X86-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; X86-AVX512VLDQ-NEXT: vinserti128 $1, %xmm0, %ymm2, %ymm0
+; X86-AVX512VLDQ-NEXT: vmovdqu64 {{[-0-9]+}}(%e{{[sb]}}p), %zmm2 # 64-byte Reload
+; X86-AVX512VLDQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%e{{[sb]}}p), %zmm2, %zmm2 # 32-byte Folded Reload
+; X86-AVX512VLDQ-NEXT: vmovdqu64 {{[-0-9]+}}(%e{{[sb]}}p), %zmm3 # 64-byte Reload
+; X86-AVX512VLDQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%e{{[sb]}}p), %zmm3, %zmm3 # 32-byte Folded Reload
+; X86-AVX512VLDQ-NEXT: vpunpckldq {{.*#+}} zmm2 = zmm3[0],zmm2[0],zmm3[1],zmm2[1],zmm3[4],zmm2[4],zmm3[5],zmm2[5],zmm3[8],zmm2[8],zmm3[9],zmm2[9],zmm3[12],zmm2[12],zmm3[13],zmm2[13]
+; X86-AVX512VLDQ-NEXT: vmovdqu64 {{[-0-9]+}}(%e{{[sb]}}p), %zmm3 # 64-byte Reload
+; X86-AVX512VLDQ-NEXT: vinserti64x4 $1, {{[-0-9]+}}(%e{{[sb]}}p), %zmm3, %zmm3 # 32-byte Folded Reload
+; X86-AVX512VLDQ-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
+; X86-AVX512VLDQ-NEXT: vpunpckldq {{.*#+}} zmm0 = zmm3[0],zmm0[0],zmm3[1],zmm0[1],zmm3[4],zmm0[4],zmm3[5],zmm0[5],zmm3[8],zmm0[8],zmm3[9],zmm0[9],zmm3[12],zmm0[12],zmm3[13],zmm0[13]
+; X86-AVX512VLDQ-NEXT: vpunpcklqdq {{.*#+}} zmm0 = zmm0[0],zmm2[0],zmm0[2],zmm2[2],zmm0[4],zmm2[4],zmm0[6],zmm2[6]
+; X86-AVX512VLDQ-NEXT: addl $400, %esp # imm = 0x190
+; X86-AVX512VLDQ-NEXT: popl %esi
+; X86-AVX512VLDQ-NEXT: popl %edi
+; X86-AVX512VLDQ-NEXT: popl %ebx
+; X86-AVX512VLDQ-NEXT: popl %ebp
+; X86-AVX512VLDQ-NEXT: retl
+;
+; X64-SSE-LABEL: fcopysign_v32f16:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: movdqa (%rdi), %xmm1
+; X64-SSE-NEXT: movdqa 16(%rdi), %xmm2
+; X64-SSE-NEXT: movdqa 32(%rdi), %xmm3
+; X64-SSE-NEXT: movdqa 48(%rdi), %xmm4
+; X64-SSE-NEXT: movdqa (%rsi), %xmm8
+; X64-SSE-NEXT: movdqa 16(%rsi), %xmm7
+; X64-SSE-NEXT: movdqa 32(%rsi), %xmm6
+; X64-SSE-NEXT: movdqa 48(%rsi), %xmm5
+; X64-SSE-NEXT: movdqa %xmm8, %xmm0
+; X64-SSE-NEXT: psrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-SSE-NEXT: pextrw $0, %xmm0, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movdqa %xmm1, %xmm0
+; X64-SSE-NEXT: psrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-SSE-NEXT: pextrw $0, %xmm0, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm9
+; X64-SSE-NEXT: movdqa %xmm8, %xmm0
+; X64-SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm8[3,3]
+; X64-SSE-NEXT: pextrw $0, %xmm0, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movdqa %xmm1, %xmm0
+; X64-SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm1[3,3]
+; X64-SSE-NEXT: pextrw $0, %xmm0, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm0
+; X64-SSE-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm9[0],xmm0[1],xmm9[1],xmm0[2],xmm9[2],xmm0[3],xmm9[3]
+; X64-SSE-NEXT: movdqa %xmm8, %xmm9
+; X64-SSE-NEXT: psrldq {{.*#+}} xmm9 = xmm9[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-SSE-NEXT: pextrw $0, %xmm9, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movdqa %xmm1, %xmm9
+; X64-SSE-NEXT: psrldq {{.*#+}} xmm9 = xmm9[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-SSE-NEXT: pextrw $0, %xmm9, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm10
+; X64-SSE-NEXT: movdqa %xmm8, %xmm9
+; X64-SSE-NEXT: punpckhqdq {{.*#+}} xmm9 = xmm9[1],xmm8[1]
+; X64-SSE-NEXT: pextrw $0, %xmm9, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movdqa %xmm1, %xmm9
+; X64-SSE-NEXT: punpckhqdq {{.*#+}} xmm9 = xmm9[1],xmm1[1]
+; X64-SSE-NEXT: pextrw $0, %xmm9, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm9
+; X64-SSE-NEXT: punpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
+; X64-SSE-NEXT: punpckldq {{.*#+}} xmm9 = xmm9[0],xmm0[0],xmm9[1],xmm0[1]
+; X64-SSE-NEXT: movdqa %xmm8, %xmm0
+; X64-SSE-NEXT: psrlq $48, %xmm0
+; X64-SSE-NEXT: pextrw $0, %xmm0, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movdqa %xmm1, %xmm0
+; X64-SSE-NEXT: psrlq $48, %xmm0
+; X64-SSE-NEXT: pextrw $0, %xmm0, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm0
+; X64-SSE-NEXT: movaps %xmm8, %xmm10
+; X64-SSE-NEXT: shufps {{.*#+}} xmm10 = xmm10[1,1],xmm8[1,1]
+; X64-SSE-NEXT: pextrw $0, %xmm10, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movaps %xmm1, %xmm10
+; X64-SSE-NEXT: shufps {{.*#+}} xmm10 = xmm10[1,1],xmm1[1,1]
+; X64-SSE-NEXT: pextrw $0, %xmm10, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm10
+; X64-SSE-NEXT: punpcklwd {{.*#+}} xmm10 = xmm10[0],xmm0[0],xmm10[1],xmm0[1],xmm10[2],xmm0[2],xmm10[3],xmm0[3]
+; X64-SSE-NEXT: pextrw $0, %xmm8, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: pextrw $0, %xmm1, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm0
+; X64-SSE-NEXT: psrld $16, %xmm8
+; X64-SSE-NEXT: pextrw $0, %xmm8, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: psrld $16, %xmm1
+; X64-SSE-NEXT: pextrw $0, %xmm1, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm1
+; X64-SSE-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; X64-SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm10[0],xmm0[1],xmm10[1]
+; X64-SSE-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm9[0]
+; X64-SSE-NEXT: movdqa %xmm7, %xmm1
+; X64-SSE-NEXT: psrldq {{.*#+}} xmm1 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-SSE-NEXT: pextrw $0, %xmm1, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movdqa %xmm2, %xmm1
+; X64-SSE-NEXT: psrldq {{.*#+}} xmm1 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-SSE-NEXT: pextrw $0, %xmm1, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm8
+; X64-SSE-NEXT: movdqa %xmm7, %xmm1
+; X64-SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[3,3],xmm7[3,3]
+; X64-SSE-NEXT: pextrw $0, %xmm1, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movdqa %xmm2, %xmm1
+; X64-SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[3,3],xmm2[3,3]
+; X64-SSE-NEXT: pextrw $0, %xmm1, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm1
+; X64-SSE-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm8[0],xmm1[1],xmm8[1],xmm1[2],xmm8[2],xmm1[3],xmm8[3]
+; X64-SSE-NEXT: movdqa %xmm7, %xmm8
+; X64-SSE-NEXT: psrldq {{.*#+}} xmm8 = xmm8[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-SSE-NEXT: pextrw $0, %xmm8, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movdqa %xmm2, %xmm8
+; X64-SSE-NEXT: psrldq {{.*#+}} xmm8 = xmm8[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-SSE-NEXT: pextrw $0, %xmm8, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm9
+; X64-SSE-NEXT: movdqa %xmm7, %xmm8
+; X64-SSE-NEXT: punpckhqdq {{.*#+}} xmm8 = xmm8[1],xmm7[1]
+; X64-SSE-NEXT: pextrw $0, %xmm8, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movdqa %xmm2, %xmm8
+; X64-SSE-NEXT: punpckhqdq {{.*#+}} xmm8 = xmm8[1],xmm2[1]
+; X64-SSE-NEXT: pextrw $0, %xmm8, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm8
+; X64-SSE-NEXT: punpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
+; X64-SSE-NEXT: punpckldq {{.*#+}} xmm8 = xmm8[0],xmm1[0],xmm8[1],xmm1[1]
+; X64-SSE-NEXT: movdqa %xmm7, %xmm1
+; X64-SSE-NEXT: psrlq $48, %xmm1
+; X64-SSE-NEXT: pextrw $0, %xmm1, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movdqa %xmm2, %xmm1
+; X64-SSE-NEXT: psrlq $48, %xmm1
+; X64-SSE-NEXT: pextrw $0, %xmm1, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm1
+; X64-SSE-NEXT: movaps %xmm7, %xmm9
+; X64-SSE-NEXT: shufps {{.*#+}} xmm9 = xmm9[1,1],xmm7[1,1]
+; X64-SSE-NEXT: pextrw $0, %xmm9, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movaps %xmm2, %xmm9
+; X64-SSE-NEXT: shufps {{.*#+}} xmm9 = xmm9[1,1],xmm2[1,1]
+; X64-SSE-NEXT: pextrw $0, %xmm9, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm9
+; X64-SSE-NEXT: punpcklwd {{.*#+}} xmm9 = xmm9[0],xmm1[0],xmm9[1],xmm1[1],xmm9[2],xmm1[2],xmm9[3],xmm1[3]
+; X64-SSE-NEXT: pextrw $0, %xmm7, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: pextrw $0, %xmm2, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm1
+; X64-SSE-NEXT: psrld $16, %xmm7
+; X64-SSE-NEXT: pextrw $0, %xmm7, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: psrld $16, %xmm2
+; X64-SSE-NEXT: pextrw $0, %xmm2, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm2
+; X64-SSE-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
+; X64-SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm9[0],xmm1[1],xmm9[1]
+; X64-SSE-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm8[0]
+; X64-SSE-NEXT: movdqa %xmm6, %xmm2
+; X64-SSE-NEXT: psrldq {{.*#+}} xmm2 = xmm2[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-SSE-NEXT: pextrw $0, %xmm2, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movdqa %xmm3, %xmm2
+; X64-SSE-NEXT: psrldq {{.*#+}} xmm2 = xmm2[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-SSE-NEXT: pextrw $0, %xmm2, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm7
+; X64-SSE-NEXT: movdqa %xmm6, %xmm2
+; X64-SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[3,3],xmm6[3,3]
+; X64-SSE-NEXT: pextrw $0, %xmm2, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movdqa %xmm3, %xmm2
+; X64-SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[3,3],xmm3[3,3]
+; X64-SSE-NEXT: pextrw $0, %xmm2, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm2
+; X64-SSE-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm7[0],xmm2[1],xmm7[1],xmm2[2],xmm7[2],xmm2[3],xmm7[3]
+; X64-SSE-NEXT: movdqa %xmm6, %xmm7
+; X64-SSE-NEXT: psrldq {{.*#+}} xmm7 = xmm7[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-SSE-NEXT: pextrw $0, %xmm7, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movdqa %xmm3, %xmm7
+; X64-SSE-NEXT: psrldq {{.*#+}} xmm7 = xmm7[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-SSE-NEXT: pextrw $0, %xmm7, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm8
+; X64-SSE-NEXT: movdqa %xmm6, %xmm7
+; X64-SSE-NEXT: punpckhqdq {{.*#+}} xmm7 = xmm7[1],xmm6[1]
+; X64-SSE-NEXT: pextrw $0, %xmm7, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movdqa %xmm3, %xmm7
+; X64-SSE-NEXT: punpckhqdq {{.*#+}} xmm7 = xmm7[1],xmm3[1]
+; X64-SSE-NEXT: pextrw $0, %xmm7, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm7
+; X64-SSE-NEXT: punpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
+; X64-SSE-NEXT: punpckldq {{.*#+}} xmm7 = xmm7[0],xmm2[0],xmm7[1],xmm2[1]
+; X64-SSE-NEXT: movdqa %xmm6, %xmm2
+; X64-SSE-NEXT: psrlq $48, %xmm2
+; X64-SSE-NEXT: pextrw $0, %xmm2, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movdqa %xmm3, %xmm2
+; X64-SSE-NEXT: psrlq $48, %xmm2
+; X64-SSE-NEXT: pextrw $0, %xmm2, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm2
+; X64-SSE-NEXT: movaps %xmm6, %xmm8
+; X64-SSE-NEXT: shufps {{.*#+}} xmm8 = xmm8[1,1],xmm6[1,1]
+; X64-SSE-NEXT: pextrw $0, %xmm8, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movaps %xmm3, %xmm8
+; X64-SSE-NEXT: shufps {{.*#+}} xmm8 = xmm8[1,1],xmm3[1,1]
+; X64-SSE-NEXT: pextrw $0, %xmm8, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm8
+; X64-SSE-NEXT: punpcklwd {{.*#+}} xmm8 = xmm8[0],xmm2[0],xmm8[1],xmm2[1],xmm8[2],xmm2[2],xmm8[3],xmm2[3]
+; X64-SSE-NEXT: pextrw $0, %xmm6, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: pextrw $0, %xmm3, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm2
+; X64-SSE-NEXT: psrld $16, %xmm6
+; X64-SSE-NEXT: pextrw $0, %xmm6, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: psrld $16, %xmm3
+; X64-SSE-NEXT: pextrw $0, %xmm3, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm3
+; X64-SSE-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
+; X64-SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm8[0],xmm2[1],xmm8[1]
+; X64-SSE-NEXT: punpcklqdq {{.*#+}} xmm2 = xmm2[0],xmm7[0]
+; X64-SSE-NEXT: movdqa %xmm5, %xmm3
+; X64-SSE-NEXT: psrldq {{.*#+}} xmm3 = xmm3[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-SSE-NEXT: pextrw $0, %xmm3, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movdqa %xmm4, %xmm3
+; X64-SSE-NEXT: psrldq {{.*#+}} xmm3 = xmm3[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-SSE-NEXT: pextrw $0, %xmm3, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm6
+; X64-SSE-NEXT: movdqa %xmm5, %xmm3
+; X64-SSE-NEXT: shufps {{.*#+}} xmm3 = xmm3[3,3],xmm5[3,3]
+; X64-SSE-NEXT: pextrw $0, %xmm3, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movdqa %xmm4, %xmm3
+; X64-SSE-NEXT: shufps {{.*#+}} xmm3 = xmm3[3,3],xmm4[3,3]
+; X64-SSE-NEXT: pextrw $0, %xmm3, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm3
+; X64-SSE-NEXT: punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm6[0],xmm3[1],xmm6[1],xmm3[2],xmm6[2],xmm3[3],xmm6[3]
+; X64-SSE-NEXT: movdqa %xmm5, %xmm6
+; X64-SSE-NEXT: psrldq {{.*#+}} xmm6 = xmm6[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-SSE-NEXT: pextrw $0, %xmm6, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movdqa %xmm4, %xmm6
+; X64-SSE-NEXT: psrldq {{.*#+}} xmm6 = xmm6[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-SSE-NEXT: pextrw $0, %xmm6, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm7
+; X64-SSE-NEXT: movdqa %xmm5, %xmm6
+; X64-SSE-NEXT: punpckhqdq {{.*#+}} xmm6 = xmm6[1],xmm5[1]
+; X64-SSE-NEXT: pextrw $0, %xmm6, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movdqa %xmm4, %xmm6
+; X64-SSE-NEXT: punpckhqdq {{.*#+}} xmm6 = xmm6[1],xmm4[1]
+; X64-SSE-NEXT: pextrw $0, %xmm6, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm6
+; X64-SSE-NEXT: punpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
+; X64-SSE-NEXT: punpckldq {{.*#+}} xmm6 = xmm6[0],xmm3[0],xmm6[1],xmm3[1]
+; X64-SSE-NEXT: movdqa %xmm5, %xmm3
+; X64-SSE-NEXT: psrlq $48, %xmm3
+; X64-SSE-NEXT: pextrw $0, %xmm3, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movdqa %xmm4, %xmm3
+; X64-SSE-NEXT: psrlq $48, %xmm3
+; X64-SSE-NEXT: pextrw $0, %xmm3, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm3
+; X64-SSE-NEXT: movaps %xmm5, %xmm7
+; X64-SSE-NEXT: shufps {{.*#+}} xmm7 = xmm7[1,1],xmm5[1,1]
+; X64-SSE-NEXT: pextrw $0, %xmm7, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: movaps %xmm4, %xmm7
+; X64-SSE-NEXT: shufps {{.*#+}} xmm7 = xmm7[1,1],xmm4[1,1]
+; X64-SSE-NEXT: pextrw $0, %xmm7, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm7
+; X64-SSE-NEXT: punpcklwd {{.*#+}} xmm7 = xmm7[0],xmm3[0],xmm7[1],xmm3[1],xmm7[2],xmm3[2],xmm7[3],xmm3[3]
+; X64-SSE-NEXT: pextrw $0, %xmm5, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: pextrw $0, %xmm4, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm3
+; X64-SSE-NEXT: psrld $16, %xmm5
+; X64-SSE-NEXT: pextrw $0, %xmm5, %eax
+; X64-SSE-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-SSE-NEXT: psrld $16, %xmm4
+; X64-SSE-NEXT: pextrw $0, %xmm4, %ecx
+; X64-SSE-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-SSE-NEXT: orl %eax, %ecx
+; X64-SSE-NEXT: pinsrw $0, %ecx, %xmm4
+; X64-SSE-NEXT: punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
+; X64-SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm7[0],xmm3[1],xmm7[1]
+; X64-SSE-NEXT: punpcklqdq {{.*#+}} xmm3 = xmm3[0],xmm6[0]
+; X64-SSE-NEXT: retq
+;
+; X64-AVX1-LABEL: fcopysign_v32f16:
+; X64-AVX1: # %bb.0:
+; X64-AVX1-NEXT: vbroadcastss 28(%rsi), %xmm0
+; X64-AVX1-NEXT: vpextrw $0, %xmm0, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vbroadcastss 28(%rdi), %xmm0
+; X64-AVX1-NEXT: vpextrw $0, %xmm0, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm8
+; X64-AVX1-NEXT: vmovdqa (%rsi), %xmm0
+; X64-AVX1-NEXT: vmovdqa 16(%rsi), %xmm6
+; X64-AVX1-NEXT: vmovdqa 32(%rsi), %xmm1
+; X64-AVX1-NEXT: vmovdqa 48(%rsi), %xmm3
+; X64-AVX1-NEXT: vpsrldq {{.*#+}} xmm2 = xmm6[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX1-NEXT: vpextrw $0, %xmm2, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vmovdqa (%rdi), %xmm5
+; X64-AVX1-NEXT: vmovdqa 16(%rdi), %xmm7
+; X64-AVX1-NEXT: vmovdqa 32(%rdi), %xmm2
+; X64-AVX1-NEXT: vmovdqa 48(%rdi), %xmm4
+; X64-AVX1-NEXT: vpsrldq {{.*#+}} xmm9 = xmm7[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX1-NEXT: vpextrw $0, %xmm9, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm9
+; X64-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
+; X64-AVX1-NEXT: vbroadcastss 24(%rsi), %xmm9
+; X64-AVX1-NEXT: vpextrw $0, %xmm9, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vbroadcastss 24(%rdi), %xmm9
+; X64-AVX1-NEXT: vpextrw $0, %xmm9, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm9
+; X64-AVX1-NEXT: vpsrldq {{.*#+}} xmm10 = xmm6[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX1-NEXT: vpextrw $0, %xmm10, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vpsrldq {{.*#+}} xmm10 = xmm7[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX1-NEXT: vpextrw $0, %xmm10, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm10
+; X64-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
+; X64-AVX1-NEXT: vpunpckldq {{.*#+}} xmm8 = xmm9[0],xmm8[0],xmm9[1],xmm8[1]
+; X64-AVX1-NEXT: vbroadcastss 20(%rsi), %xmm9
+; X64-AVX1-NEXT: vpextrw $0, %xmm9, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vbroadcastss 20(%rdi), %xmm9
+; X64-AVX1-NEXT: vpextrw $0, %xmm9, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm9
+; X64-AVX1-NEXT: vpsrlq $48, %xmm6, %xmm10
+; X64-AVX1-NEXT: vpextrw $0, %xmm10, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vpsrlq $48, %xmm7, %xmm10
+; X64-AVX1-NEXT: vpextrw $0, %xmm10, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm10
+; X64-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
+; X64-AVX1-NEXT: vpextrw $0, %xmm6, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vpextrw $0, %xmm7, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm10
+; X64-AVX1-NEXT: vpsrld $16, %xmm6, %xmm6
+; X64-AVX1-NEXT: vpextrw $0, %xmm6, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vpsrld $16, %xmm7, %xmm6
+; X64-AVX1-NEXT: vpextrw $0, %xmm6, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm6
+; X64-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm10[0],xmm6[0],xmm10[1],xmm6[1],xmm10[2],xmm6[2],xmm10[3],xmm6[3]
+; X64-AVX1-NEXT: vpunpckldq {{.*#+}} xmm6 = xmm6[0],xmm9[0],xmm6[1],xmm9[1]
+; X64-AVX1-NEXT: vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],xmm8[0]
+; X64-AVX1-NEXT: vbroadcastss 12(%rsi), %xmm7
+; X64-AVX1-NEXT: vpextrw $0, %xmm7, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vbroadcastss 12(%rdi), %xmm7
+; X64-AVX1-NEXT: vpextrw $0, %xmm7, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm7
+; X64-AVX1-NEXT: vpsrldq {{.*#+}} xmm8 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX1-NEXT: vpextrw $0, %xmm8, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vpsrldq {{.*#+}} xmm8 = xmm5[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX1-NEXT: vpextrw $0, %xmm8, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm8
+; X64-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
+; X64-AVX1-NEXT: vbroadcastss 8(%rsi), %xmm8
+; X64-AVX1-NEXT: vpextrw $0, %xmm8, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vbroadcastss 8(%rdi), %xmm8
+; X64-AVX1-NEXT: vpextrw $0, %xmm8, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm8
+; X64-AVX1-NEXT: vpsrldq {{.*#+}} xmm9 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX1-NEXT: vpextrw $0, %xmm9, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vpsrldq {{.*#+}} xmm9 = xmm5[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX1-NEXT: vpextrw $0, %xmm9, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm9
+; X64-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
+; X64-AVX1-NEXT: vpunpckldq {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]
+; X64-AVX1-NEXT: vbroadcastss 4(%rsi), %xmm8
+; X64-AVX1-NEXT: vpextrw $0, %xmm8, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vbroadcastss 4(%rdi), %xmm8
+; X64-AVX1-NEXT: vpextrw $0, %xmm8, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm8
+; X64-AVX1-NEXT: vpsrlq $48, %xmm0, %xmm9
+; X64-AVX1-NEXT: vpextrw $0, %xmm9, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vpsrlq $48, %xmm5, %xmm9
+; X64-AVX1-NEXT: vpextrw $0, %xmm9, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm9
+; X64-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
+; X64-AVX1-NEXT: vpextrw $0, %xmm0, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vpextrw $0, %xmm5, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm9
+; X64-AVX1-NEXT: vpsrld $16, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpextrw $0, %xmm0, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vpsrld $16, %xmm5, %xmm0
+; X64-AVX1-NEXT: vpextrw $0, %xmm0, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm9[0],xmm0[0],xmm9[1],xmm0[1],xmm9[2],xmm0[2],xmm9[3],xmm0[3]
+; X64-AVX1-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm8[0],xmm0[1],xmm8[1]
+; X64-AVX1-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm7[0]
+; X64-AVX1-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm0
+; X64-AVX1-NEXT: vbroadcastss 60(%rsi), %xmm5
+; X64-AVX1-NEXT: vpextrw $0, %xmm5, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vbroadcastss 60(%rdi), %xmm5
+; X64-AVX1-NEXT: vpextrw $0, %xmm5, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm5
+; X64-AVX1-NEXT: vpsrldq {{.*#+}} xmm6 = xmm3[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX1-NEXT: vpextrw $0, %xmm6, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vpsrldq {{.*#+}} xmm6 = xmm4[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX1-NEXT: vpextrw $0, %xmm6, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm6
+; X64-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
+; X64-AVX1-NEXT: vbroadcastss 56(%rsi), %xmm6
+; X64-AVX1-NEXT: vpextrw $0, %xmm6, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vbroadcastss 56(%rdi), %xmm6
+; X64-AVX1-NEXT: vpextrw $0, %xmm6, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm6
+; X64-AVX1-NEXT: vpsrldq {{.*#+}} xmm7 = xmm3[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX1-NEXT: vpextrw $0, %xmm7, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vpsrldq {{.*#+}} xmm7 = xmm4[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX1-NEXT: vpextrw $0, %xmm7, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm7
+; X64-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
+; X64-AVX1-NEXT: vpunpckldq {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1]
+; X64-AVX1-NEXT: vbroadcastss 52(%rsi), %xmm6
+; X64-AVX1-NEXT: vpextrw $0, %xmm6, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vbroadcastss 52(%rdi), %xmm6
+; X64-AVX1-NEXT: vpextrw $0, %xmm6, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm6
+; X64-AVX1-NEXT: vpsrlq $48, %xmm3, %xmm7
+; X64-AVX1-NEXT: vpextrw $0, %xmm7, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vpsrlq $48, %xmm4, %xmm7
+; X64-AVX1-NEXT: vpextrw $0, %xmm7, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm7
+; X64-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
+; X64-AVX1-NEXT: vpextrw $0, %xmm3, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vpextrw $0, %xmm4, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm7
+; X64-AVX1-NEXT: vpsrld $16, %xmm3, %xmm3
+; X64-AVX1-NEXT: vpextrw $0, %xmm3, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vpsrld $16, %xmm4, %xmm3
+; X64-AVX1-NEXT: vpextrw $0, %xmm3, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm3
+; X64-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm7[0],xmm3[0],xmm7[1],xmm3[1],xmm7[2],xmm3[2],xmm7[3],xmm3[3]
+; X64-AVX1-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm3[0],xmm6[0],xmm3[1],xmm6[1]
+; X64-AVX1-NEXT: vpunpcklqdq {{.*#+}} xmm3 = xmm3[0],xmm5[0]
+; X64-AVX1-NEXT: vbroadcastss 44(%rsi), %xmm4
+; X64-AVX1-NEXT: vpextrw $0, %xmm4, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vbroadcastss 44(%rdi), %xmm4
+; X64-AVX1-NEXT: vpextrw $0, %xmm4, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm4
+; X64-AVX1-NEXT: vpsrldq {{.*#+}} xmm5 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX1-NEXT: vpextrw $0, %xmm5, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vpsrldq {{.*#+}} xmm5 = xmm2[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX1-NEXT: vpextrw $0, %xmm5, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm5
+; X64-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
+; X64-AVX1-NEXT: vbroadcastss 40(%rsi), %xmm5
+; X64-AVX1-NEXT: vpextrw $0, %xmm5, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vbroadcastss 40(%rdi), %xmm5
+; X64-AVX1-NEXT: vpextrw $0, %xmm5, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm5
+; X64-AVX1-NEXT: vpsrldq {{.*#+}} xmm6 = xmm1[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX1-NEXT: vpextrw $0, %xmm6, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vpsrldq {{.*#+}} xmm6 = xmm2[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX1-NEXT: vpextrw $0, %xmm6, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm6
+; X64-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
+; X64-AVX1-NEXT: vpunpckldq {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]
+; X64-AVX1-NEXT: vbroadcastss 36(%rsi), %xmm5
+; X64-AVX1-NEXT: vpextrw $0, %xmm5, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vbroadcastss 36(%rdi), %xmm5
+; X64-AVX1-NEXT: vpextrw $0, %xmm5, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm5
+; X64-AVX1-NEXT: vpsrlq $48, %xmm1, %xmm6
+; X64-AVX1-NEXT: vpextrw $0, %xmm6, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vpsrlq $48, %xmm2, %xmm6
+; X64-AVX1-NEXT: vpextrw $0, %xmm6, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm6
+; X64-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
+; X64-AVX1-NEXT: vpextrw $0, %xmm1, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vpextrw $0, %xmm2, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm6
+; X64-AVX1-NEXT: vpsrld $16, %xmm1, %xmm1
+; X64-AVX1-NEXT: vpextrw $0, %xmm1, %eax
+; X64-AVX1-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX1-NEXT: vpsrld $16, %xmm2, %xmm1
+; X64-AVX1-NEXT: vpextrw $0, %xmm1, %ecx
+; X64-AVX1-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX1-NEXT: orl %eax, %ecx
+; X64-AVX1-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm1
+; X64-AVX1-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm6[0],xmm1[0],xmm6[1],xmm1[1],xmm6[2],xmm1[2],xmm6[3],xmm1[3]
+; X64-AVX1-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],xmm5[0],xmm1[1],xmm5[1]
+; X64-AVX1-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm4[0]
+; X64-AVX1-NEXT: vinsertf128 $1, %xmm3, %ymm1, %ymm1
+; X64-AVX1-NEXT: retq
+;
+; X64-AVX2-LABEL: fcopysign_v32f16:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: vmovdqa (%rsi), %xmm0
+; X64-AVX2-NEXT: vmovdqa 16(%rsi), %xmm6
+; X64-AVX2-NEXT: vmovdqa 32(%rsi), %xmm1
+; X64-AVX2-NEXT: vmovdqa 48(%rsi), %xmm3
+; X64-AVX2-NEXT: vpsrldq {{.*#+}} xmm2 = xmm6[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX2-NEXT: vpextrw $0, %xmm2, %eax
+; X64-AVX2-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX2-NEXT: vmovdqa (%rdi), %xmm5
+; X64-AVX2-NEXT: vmovdqa 16(%rdi), %xmm7
+; X64-AVX2-NEXT: vmovdqa 32(%rdi), %xmm2
+; X64-AVX2-NEXT: vmovdqa 48(%rdi), %xmm4
+; X64-AVX2-NEXT: vpsrldq {{.*#+}} xmm8 = xmm7[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX2-NEXT: vpextrw $0, %xmm8, %ecx
+; X64-AVX2-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %eax, %ecx
+; X64-AVX2-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm8
+; X64-AVX2-NEXT: movl $-32768, %eax # imm = 0x8000
+; X64-AVX2-NEXT: movl 28(%rsi), %ecx
+; X64-AVX2-NEXT: andl %eax, %ecx
+; X64-AVX2-NEXT: movzwl 28(%rdi), %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm9
+; X64-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
+; X64-AVX2-NEXT: vpsrldq {{.*#+}} xmm9 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX2-NEXT: vpextrw $0, %xmm9, %ecx
+; X64-AVX2-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX2-NEXT: vpsrldq {{.*#+}} xmm9 = xmm5[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX2-NEXT: vpextrw $0, %xmm9, %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm9
+; X64-AVX2-NEXT: movl 12(%rsi), %ecx
+; X64-AVX2-NEXT: andl %eax, %ecx
+; X64-AVX2-NEXT: movzwl 12(%rdi), %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm10
+; X64-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
+; X64-AVX2-NEXT: vinserti128 $1, %xmm8, %ymm9, %ymm8
+; X64-AVX2-NEXT: vpsrldq {{.*#+}} xmm9 = xmm6[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX2-NEXT: vpextrw $0, %xmm9, %ecx
+; X64-AVX2-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX2-NEXT: vpsrldq {{.*#+}} xmm9 = xmm7[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX2-NEXT: vpextrw $0, %xmm9, %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm9
+; X64-AVX2-NEXT: movl 24(%rsi), %ecx
+; X64-AVX2-NEXT: andl %eax, %ecx
+; X64-AVX2-NEXT: movzwl 24(%rdi), %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm10
+; X64-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
+; X64-AVX2-NEXT: vpsrldq {{.*#+}} xmm10 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX2-NEXT: vpextrw $0, %xmm10, %ecx
+; X64-AVX2-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX2-NEXT: vpsrldq {{.*#+}} xmm10 = xmm5[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX2-NEXT: vpextrw $0, %xmm10, %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm10
+; X64-AVX2-NEXT: movl 8(%rsi), %ecx
+; X64-AVX2-NEXT: andl %eax, %ecx
+; X64-AVX2-NEXT: movzwl 8(%rdi), %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm11
+; X64-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm10 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3]
+; X64-AVX2-NEXT: vinserti128 $1, %xmm9, %ymm10, %ymm9
+; X64-AVX2-NEXT: vpunpckldq {{.*#+}} ymm8 = ymm9[0],ymm8[0],ymm9[1],ymm8[1],ymm9[4],ymm8[4],ymm9[5],ymm8[5]
+; X64-AVX2-NEXT: vpextrw $0, %xmm6, %ecx
+; X64-AVX2-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX2-NEXT: vpextrw $0, %xmm7, %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm9
+; X64-AVX2-NEXT: vpsrld $16, %xmm6, %xmm10
+; X64-AVX2-NEXT: vpextrw $0, %xmm10, %ecx
+; X64-AVX2-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX2-NEXT: vpsrld $16, %xmm7, %xmm10
+; X64-AVX2-NEXT: vpextrw $0, %xmm10, %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm10
+; X64-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
+; X64-AVX2-NEXT: vpextrw $0, %xmm0, %ecx
+; X64-AVX2-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX2-NEXT: vpextrw $0, %xmm5, %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm10
+; X64-AVX2-NEXT: vpsrld $16, %xmm0, %xmm11
+; X64-AVX2-NEXT: vpextrw $0, %xmm11, %ecx
+; X64-AVX2-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX2-NEXT: vpsrld $16, %xmm5, %xmm11
+; X64-AVX2-NEXT: vpextrw $0, %xmm11, %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm11
+; X64-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
+; X64-AVX2-NEXT: vinserti128 $1, %xmm9, %ymm10, %ymm9
+; X64-AVX2-NEXT: vpsrlq $48, %xmm6, %xmm6
+; X64-AVX2-NEXT: vpextrw $0, %xmm6, %ecx
+; X64-AVX2-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX2-NEXT: vpsrlq $48, %xmm7, %xmm6
+; X64-AVX2-NEXT: vpextrw $0, %xmm6, %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm6
+; X64-AVX2-NEXT: movl 20(%rsi), %ecx
+; X64-AVX2-NEXT: andl %eax, %ecx
+; X64-AVX2-NEXT: movzwl 20(%rdi), %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm7
+; X64-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
+; X64-AVX2-NEXT: vpsrlq $48, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpextrw $0, %xmm0, %ecx
+; X64-AVX2-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX2-NEXT: vpsrlq $48, %xmm5, %xmm0
+; X64-AVX2-NEXT: vpextrw $0, %xmm0, %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm0
+; X64-AVX2-NEXT: movl 4(%rsi), %ecx
+; X64-AVX2-NEXT: andl %eax, %ecx
+; X64-AVX2-NEXT: movzwl 4(%rdi), %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm5
+; X64-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm5[0],xmm0[0],xmm5[1],xmm0[1],xmm5[2],xmm0[2],xmm5[3],xmm0[3]
+; X64-AVX2-NEXT: vinserti128 $1, %xmm6, %ymm0, %ymm0
+; X64-AVX2-NEXT: vpunpckldq {{.*#+}} ymm0 = ymm9[0],ymm0[0],ymm9[1],ymm0[1],ymm9[4],ymm0[4],ymm9[5],ymm0[5]
+; X64-AVX2-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm8[0],ymm0[2],ymm8[2]
+; X64-AVX2-NEXT: vpsrldq {{.*#+}} xmm5 = xmm3[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX2-NEXT: vpextrw $0, %xmm5, %ecx
+; X64-AVX2-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX2-NEXT: vpsrldq {{.*#+}} xmm5 = xmm4[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX2-NEXT: vpextrw $0, %xmm5, %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm5
+; X64-AVX2-NEXT: movl 60(%rsi), %ecx
+; X64-AVX2-NEXT: andl %eax, %ecx
+; X64-AVX2-NEXT: movzwl 60(%rdi), %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm6
+; X64-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
+; X64-AVX2-NEXT: vpsrldq {{.*#+}} xmm6 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX2-NEXT: vpextrw $0, %xmm6, %ecx
+; X64-AVX2-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX2-NEXT: vpsrldq {{.*#+}} xmm6 = xmm2[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX2-NEXT: vpextrw $0, %xmm6, %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm6
+; X64-AVX2-NEXT: movl 44(%rsi), %ecx
+; X64-AVX2-NEXT: andl %eax, %ecx
+; X64-AVX2-NEXT: movzwl 44(%rdi), %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm7
+; X64-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
+; X64-AVX2-NEXT: vinserti128 $1, %xmm5, %ymm6, %ymm5
+; X64-AVX2-NEXT: vpsrldq {{.*#+}} xmm6 = xmm3[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX2-NEXT: vpextrw $0, %xmm6, %ecx
+; X64-AVX2-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX2-NEXT: vpsrldq {{.*#+}} xmm6 = xmm4[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX2-NEXT: vpextrw $0, %xmm6, %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm6
+; X64-AVX2-NEXT: movl 56(%rsi), %ecx
+; X64-AVX2-NEXT: andl %eax, %ecx
+; X64-AVX2-NEXT: movzwl 56(%rdi), %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm7
+; X64-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
+; X64-AVX2-NEXT: vpsrldq {{.*#+}} xmm7 = xmm1[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX2-NEXT: vpextrw $0, %xmm7, %ecx
+; X64-AVX2-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX2-NEXT: vpsrldq {{.*#+}} xmm7 = xmm2[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX2-NEXT: vpextrw $0, %xmm7, %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm7
+; X64-AVX2-NEXT: movl 40(%rsi), %ecx
+; X64-AVX2-NEXT: andl %eax, %ecx
+; X64-AVX2-NEXT: movzwl 40(%rdi), %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm8
+; X64-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
+; X64-AVX2-NEXT: vinserti128 $1, %xmm6, %ymm7, %ymm6
+; X64-AVX2-NEXT: vpunpckldq {{.*#+}} ymm5 = ymm6[0],ymm5[0],ymm6[1],ymm5[1],ymm6[4],ymm5[4],ymm6[5],ymm5[5]
+; X64-AVX2-NEXT: vpextrw $0, %xmm3, %ecx
+; X64-AVX2-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX2-NEXT: vpextrw $0, %xmm4, %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm6
+; X64-AVX2-NEXT: vpsrld $16, %xmm3, %xmm7
+; X64-AVX2-NEXT: vpextrw $0, %xmm7, %ecx
+; X64-AVX2-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX2-NEXT: vpsrld $16, %xmm4, %xmm7
+; X64-AVX2-NEXT: vpextrw $0, %xmm7, %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm7
+; X64-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
+; X64-AVX2-NEXT: vpextrw $0, %xmm1, %ecx
+; X64-AVX2-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX2-NEXT: vpextrw $0, %xmm2, %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm7
+; X64-AVX2-NEXT: vpsrld $16, %xmm1, %xmm8
+; X64-AVX2-NEXT: vpextrw $0, %xmm8, %ecx
+; X64-AVX2-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX2-NEXT: vpsrld $16, %xmm2, %xmm8
+; X64-AVX2-NEXT: vpextrw $0, %xmm8, %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm8
+; X64-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
+; X64-AVX2-NEXT: vinserti128 $1, %xmm6, %ymm7, %ymm6
+; X64-AVX2-NEXT: vpsrlq $48, %xmm3, %xmm3
+; X64-AVX2-NEXT: vpextrw $0, %xmm3, %ecx
+; X64-AVX2-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX2-NEXT: vpsrlq $48, %xmm4, %xmm3
+; X64-AVX2-NEXT: vpextrw $0, %xmm3, %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm3
+; X64-AVX2-NEXT: movl 52(%rsi), %ecx
+; X64-AVX2-NEXT: andl %eax, %ecx
+; X64-AVX2-NEXT: movzwl 52(%rdi), %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm4
+; X64-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
+; X64-AVX2-NEXT: vpsrlq $48, %xmm1, %xmm1
+; X64-AVX2-NEXT: vpextrw $0, %xmm1, %ecx
+; X64-AVX2-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX2-NEXT: vpsrlq $48, %xmm2, %xmm1
+; X64-AVX2-NEXT: vpextrw $0, %xmm1, %edx
+; X64-AVX2-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %ecx, %edx
+; X64-AVX2-NEXT: vpinsrw $0, %edx, %xmm0, %xmm1
+; X64-AVX2-NEXT: andl 36(%rsi), %eax
+; X64-AVX2-NEXT: movzwl 36(%rdi), %ecx
+; X64-AVX2-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX2-NEXT: orl %eax, %ecx
+; X64-AVX2-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm2
+; X64-AVX2-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
+; X64-AVX2-NEXT: vinserti128 $1, %xmm3, %ymm1, %ymm1
+; X64-AVX2-NEXT: vpunpckldq {{.*#+}} ymm1 = ymm6[0],ymm1[0],ymm6[1],ymm1[1],ymm6[4],ymm1[4],ymm6[5],ymm1[5]
+; X64-AVX2-NEXT: vpunpcklqdq {{.*#+}} ymm1 = ymm1[0],ymm5[0],ymm1[2],ymm5[2]
+; X64-AVX2-NEXT: retq
+;
+; X64-AVX512VL-LABEL: fcopysign_v32f16:
+; X64-AVX512VL: # %bb.0:
+; X64-AVX512VL-NEXT: vmovdqa (%rsi), %xmm0
+; X64-AVX512VL-NEXT: vmovdqa 16(%rsi), %xmm2
+; X64-AVX512VL-NEXT: vmovdqa 32(%rsi), %xmm4
+; X64-AVX512VL-NEXT: vmovdqa 48(%rsi), %xmm6
+; X64-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm1 = xmm6[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm1, %eax
+; X64-AVX512VL-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX512VL-NEXT: vmovdqa (%rdi), %xmm1
+; X64-AVX512VL-NEXT: vmovdqa 16(%rdi), %xmm3
+; X64-AVX512VL-NEXT: vmovdqa 32(%rdi), %xmm5
+; X64-AVX512VL-NEXT: vmovdqa 48(%rdi), %xmm7
+; X64-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm8 = xmm7[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm8, %ecx
+; X64-AVX512VL-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %eax, %ecx
+; X64-AVX512VL-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm8
+; X64-AVX512VL-NEXT: movl $-32768, %eax # imm = 0x8000
+; X64-AVX512VL-NEXT: movl 60(%rsi), %ecx
+; X64-AVX512VL-NEXT: andl %eax, %ecx
+; X64-AVX512VL-NEXT: movzwl 60(%rdi), %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm9
+; X64-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
+; X64-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm9 = xmm4[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm9, %ecx
+; X64-AVX512VL-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm9 = xmm5[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm9, %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm9
+; X64-AVX512VL-NEXT: movl 44(%rsi), %ecx
+; X64-AVX512VL-NEXT: andl %eax, %ecx
+; X64-AVX512VL-NEXT: movzwl 44(%rdi), %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm10
+; X64-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
+; X64-AVX512VL-NEXT: vinserti128 $1, %xmm8, %ymm9, %ymm8
+; X64-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm9 = xmm2[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm9, %ecx
+; X64-AVX512VL-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm9 = xmm3[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm9, %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm9
+; X64-AVX512VL-NEXT: movl 28(%rsi), %ecx
+; X64-AVX512VL-NEXT: andl %eax, %ecx
+; X64-AVX512VL-NEXT: movzwl 28(%rdi), %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm10
+; X64-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
+; X64-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm10 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm10, %ecx
+; X64-AVX512VL-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm10 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm10, %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm10
+; X64-AVX512VL-NEXT: movl 12(%rsi), %ecx
+; X64-AVX512VL-NEXT: andl %eax, %ecx
+; X64-AVX512VL-NEXT: movzwl 12(%rdi), %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm11
+; X64-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm10 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3]
+; X64-AVX512VL-NEXT: vinserti128 $1, %xmm9, %ymm10, %ymm9
+; X64-AVX512VL-NEXT: vinserti64x4 $1, %ymm8, %zmm9, %zmm8
+; X64-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm9 = xmm6[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm9, %ecx
+; X64-AVX512VL-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm9 = xmm7[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm9, %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm9
+; X64-AVX512VL-NEXT: movl 56(%rsi), %ecx
+; X64-AVX512VL-NEXT: andl %eax, %ecx
+; X64-AVX512VL-NEXT: movzwl 56(%rdi), %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm10
+; X64-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
+; X64-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm10 = xmm4[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm10, %ecx
+; X64-AVX512VL-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm10 = xmm5[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm10, %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm10
+; X64-AVX512VL-NEXT: movl 40(%rsi), %ecx
+; X64-AVX512VL-NEXT: andl %eax, %ecx
+; X64-AVX512VL-NEXT: movzwl 40(%rdi), %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm11
+; X64-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm10 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3]
+; X64-AVX512VL-NEXT: vinserti128 $1, %xmm9, %ymm10, %ymm9
+; X64-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm10 = xmm2[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm10, %ecx
+; X64-AVX512VL-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm10 = xmm3[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm10, %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm10
+; X64-AVX512VL-NEXT: movl 24(%rsi), %ecx
+; X64-AVX512VL-NEXT: andl %eax, %ecx
+; X64-AVX512VL-NEXT: movzwl 24(%rdi), %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm11
+; X64-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm10 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3]
+; X64-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm11 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm11, %ecx
+; X64-AVX512VL-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VL-NEXT: vpsrldq {{.*#+}} xmm11 = xmm1[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm11, %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm11
+; X64-AVX512VL-NEXT: movl 8(%rsi), %ecx
+; X64-AVX512VL-NEXT: andl %eax, %ecx
+; X64-AVX512VL-NEXT: movzwl 8(%rdi), %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm12
+; X64-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3]
+; X64-AVX512VL-NEXT: vinserti128 $1, %xmm10, %ymm11, %ymm10
+; X64-AVX512VL-NEXT: vinserti64x4 $1, %ymm9, %zmm10, %zmm9
+; X64-AVX512VL-NEXT: vpunpckldq {{.*#+}} zmm8 = zmm9[0],zmm8[0],zmm9[1],zmm8[1],zmm9[4],zmm8[4],zmm9[5],zmm8[5],zmm9[8],zmm8[8],zmm9[9],zmm8[9],zmm9[12],zmm8[12],zmm9[13],zmm8[13]
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm6, %ecx
+; X64-AVX512VL-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm7, %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm9
+; X64-AVX512VL-NEXT: vpsrld $16, %xmm6, %xmm10
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm10, %ecx
+; X64-AVX512VL-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VL-NEXT: vpsrld $16, %xmm7, %xmm10
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm10, %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm10
+; X64-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm4, %ecx
+; X64-AVX512VL-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm5, %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm10
+; X64-AVX512VL-NEXT: vpsrld $16, %xmm4, %xmm11
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm11, %ecx
+; X64-AVX512VL-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VL-NEXT: vpsrld $16, %xmm5, %xmm11
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm11, %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm11
+; X64-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
+; X64-AVX512VL-NEXT: vinserti128 $1, %xmm9, %ymm10, %ymm9
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm2, %ecx
+; X64-AVX512VL-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm3, %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm10
+; X64-AVX512VL-NEXT: vpsrld $16, %xmm2, %xmm11
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm11, %ecx
+; X64-AVX512VL-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VL-NEXT: vpsrld $16, %xmm3, %xmm11
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm11, %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm11
+; X64-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm0, %ecx
+; X64-AVX512VL-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm1, %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm11
+; X64-AVX512VL-NEXT: vpsrld $16, %xmm0, %xmm12
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm12, %ecx
+; X64-AVX512VL-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VL-NEXT: vpsrld $16, %xmm1, %xmm12
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm12, %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm12
+; X64-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
+; X64-AVX512VL-NEXT: vinserti128 $1, %xmm10, %ymm11, %ymm10
+; X64-AVX512VL-NEXT: vinserti64x4 $1, %ymm9, %zmm10, %zmm9
+; X64-AVX512VL-NEXT: vpsrlq $48, %xmm6, %xmm6
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm6, %ecx
+; X64-AVX512VL-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VL-NEXT: vpsrlq $48, %xmm7, %xmm6
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm6, %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm6
+; X64-AVX512VL-NEXT: movl 52(%rsi), %ecx
+; X64-AVX512VL-NEXT: andl %eax, %ecx
+; X64-AVX512VL-NEXT: movzwl 52(%rdi), %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm7
+; X64-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
+; X64-AVX512VL-NEXT: vpsrlq $48, %xmm4, %xmm4
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm4, %ecx
+; X64-AVX512VL-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VL-NEXT: vpsrlq $48, %xmm5, %xmm4
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm4, %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm4
+; X64-AVX512VL-NEXT: movl 36(%rsi), %ecx
+; X64-AVX512VL-NEXT: andl %eax, %ecx
+; X64-AVX512VL-NEXT: movzwl 36(%rdi), %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm5
+; X64-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; X64-AVX512VL-NEXT: vinserti128 $1, %xmm6, %ymm4, %ymm4
+; X64-AVX512VL-NEXT: vpsrlq $48, %xmm2, %xmm2
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm2, %ecx
+; X64-AVX512VL-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VL-NEXT: vpsrlq $48, %xmm3, %xmm2
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm2, %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm2
+; X64-AVX512VL-NEXT: movl 20(%rsi), %ecx
+; X64-AVX512VL-NEXT: andl %eax, %ecx
+; X64-AVX512VL-NEXT: movzwl 20(%rdi), %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm3
+; X64-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; X64-AVX512VL-NEXT: vpsrlq $48, %xmm0, %xmm0
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm0, %ecx
+; X64-AVX512VL-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VL-NEXT: vpsrlq $48, %xmm1, %xmm0
+; X64-AVX512VL-NEXT: vpextrw $0, %xmm0, %edx
+; X64-AVX512VL-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %ecx, %edx
+; X64-AVX512VL-NEXT: vpinsrw $0, %edx, %xmm0, %xmm0
+; X64-AVX512VL-NEXT: andl 4(%rsi), %eax
+; X64-AVX512VL-NEXT: movzwl 4(%rdi), %ecx
+; X64-AVX512VL-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX512VL-NEXT: orl %eax, %ecx
+; X64-AVX512VL-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm1
+; X64-AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; X64-AVX512VL-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm0
+; X64-AVX512VL-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm0
+; X64-AVX512VL-NEXT: vpunpckldq {{.*#+}} zmm0 = zmm9[0],zmm0[0],zmm9[1],zmm0[1],zmm9[4],zmm0[4],zmm9[5],zmm0[5],zmm9[8],zmm0[8],zmm9[9],zmm0[9],zmm9[12],zmm0[12],zmm9[13],zmm0[13]
+; X64-AVX512VL-NEXT: vpunpcklqdq {{.*#+}} zmm0 = zmm0[0],zmm8[0],zmm0[2],zmm8[2],zmm0[4],zmm8[4],zmm0[6],zmm8[6]
+; X64-AVX512VL-NEXT: retq
+;
+; X64-AVX512FP16-LABEL: fcopysign_v32f16:
+; X64-AVX512FP16: # %bb.0:
+; X64-AVX512FP16-NEXT: vmovdqu64 (%rdi), %zmm1
+; X64-AVX512FP16-NEXT: vpbroadcastq {{.*#+}} zmm0 = [9223231297218904063,9223231297218904063,9223231297218904063,9223231297218904063,9223231297218904063,9223231297218904063,9223231297218904063,9223231297218904063]
+; X64-AVX512FP16-NEXT: vpternlogq $202, (%rsi), %zmm1, %zmm0
+; X64-AVX512FP16-NEXT: retq
+;
+; X64-AVX512VLDQ-LABEL: fcopysign_v32f16:
+; X64-AVX512VLDQ: # %bb.0:
+; X64-AVX512VLDQ-NEXT: vmovdqa (%rsi), %xmm0
+; X64-AVX512VLDQ-NEXT: vmovdqa 16(%rsi), %xmm2
+; X64-AVX512VLDQ-NEXT: vmovdqa 32(%rsi), %xmm4
+; X64-AVX512VLDQ-NEXT: vmovdqa 48(%rsi), %xmm6
+; X64-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm1 = xmm6[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm1, %eax
+; X64-AVX512VLDQ-NEXT: andl $-32768, %eax # imm = 0x8000
+; X64-AVX512VLDQ-NEXT: vmovdqa (%rdi), %xmm1
+; X64-AVX512VLDQ-NEXT: vmovdqa 16(%rdi), %xmm3
+; X64-AVX512VLDQ-NEXT: vmovdqa 32(%rdi), %xmm5
+; X64-AVX512VLDQ-NEXT: vmovdqa 48(%rdi), %xmm7
+; X64-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm8 = xmm7[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm8, %ecx
+; X64-AVX512VLDQ-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %eax, %ecx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm8
+; X64-AVX512VLDQ-NEXT: movl $-32768, %eax # imm = 0x8000
+; X64-AVX512VLDQ-NEXT: movl 60(%rsi), %ecx
+; X64-AVX512VLDQ-NEXT: andl %eax, %ecx
+; X64-AVX512VLDQ-NEXT: movzwl 60(%rdi), %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm9
+; X64-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm8 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
+; X64-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm9 = xmm4[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm9, %ecx
+; X64-AVX512VLDQ-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm9 = xmm5[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm9, %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm9
+; X64-AVX512VLDQ-NEXT: movl 44(%rsi), %ecx
+; X64-AVX512VLDQ-NEXT: andl %eax, %ecx
+; X64-AVX512VLDQ-NEXT: movzwl 44(%rdi), %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm10
+; X64-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
+; X64-AVX512VLDQ-NEXT: vinserti128 $1, %xmm8, %ymm9, %ymm8
+; X64-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm9 = xmm2[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm9, %ecx
+; X64-AVX512VLDQ-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm9 = xmm3[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm9, %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm9
+; X64-AVX512VLDQ-NEXT: movl 28(%rsi), %ecx
+; X64-AVX512VLDQ-NEXT: andl %eax, %ecx
+; X64-AVX512VLDQ-NEXT: movzwl 28(%rdi), %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm10
+; X64-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
+; X64-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm10 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm10, %ecx
+; X64-AVX512VLDQ-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm10 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm10, %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm10
+; X64-AVX512VLDQ-NEXT: movl 12(%rsi), %ecx
+; X64-AVX512VLDQ-NEXT: andl %eax, %ecx
+; X64-AVX512VLDQ-NEXT: movzwl 12(%rdi), %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm11
+; X64-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm10 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3]
+; X64-AVX512VLDQ-NEXT: vinserti128 $1, %xmm9, %ymm10, %ymm9
+; X64-AVX512VLDQ-NEXT: vinserti64x4 $1, %ymm8, %zmm9, %zmm8
+; X64-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm9 = xmm6[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm9, %ecx
+; X64-AVX512VLDQ-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm9 = xmm7[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm9, %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm9
+; X64-AVX512VLDQ-NEXT: movl 56(%rsi), %ecx
+; X64-AVX512VLDQ-NEXT: andl %eax, %ecx
+; X64-AVX512VLDQ-NEXT: movzwl 56(%rdi), %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm10
+; X64-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
+; X64-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm10 = xmm4[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm10, %ecx
+; X64-AVX512VLDQ-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm10 = xmm5[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm10, %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm10
+; X64-AVX512VLDQ-NEXT: movl 40(%rsi), %ecx
+; X64-AVX512VLDQ-NEXT: andl %eax, %ecx
+; X64-AVX512VLDQ-NEXT: movzwl 40(%rdi), %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm11
+; X64-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm10 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3]
+; X64-AVX512VLDQ-NEXT: vinserti128 $1, %xmm9, %ymm10, %ymm9
+; X64-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm10 = xmm2[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm10, %ecx
+; X64-AVX512VLDQ-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm10 = xmm3[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm10, %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm10
+; X64-AVX512VLDQ-NEXT: movl 24(%rsi), %ecx
+; X64-AVX512VLDQ-NEXT: andl %eax, %ecx
+; X64-AVX512VLDQ-NEXT: movzwl 24(%rdi), %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm11
+; X64-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm10 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3]
+; X64-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm11 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm11, %ecx
+; X64-AVX512VLDQ-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VLDQ-NEXT: vpsrldq {{.*#+}} xmm11 = xmm1[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm11, %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm11
+; X64-AVX512VLDQ-NEXT: movl 8(%rsi), %ecx
+; X64-AVX512VLDQ-NEXT: andl %eax, %ecx
+; X64-AVX512VLDQ-NEXT: movzwl 8(%rdi), %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm12
+; X64-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3]
+; X64-AVX512VLDQ-NEXT: vinserti128 $1, %xmm10, %ymm11, %ymm10
+; X64-AVX512VLDQ-NEXT: vinserti64x4 $1, %ymm9, %zmm10, %zmm9
+; X64-AVX512VLDQ-NEXT: vpunpckldq {{.*#+}} zmm8 = zmm9[0],zmm8[0],zmm9[1],zmm8[1],zmm9[4],zmm8[4],zmm9[5],zmm8[5],zmm9[8],zmm8[8],zmm9[9],zmm8[9],zmm9[12],zmm8[12],zmm9[13],zmm8[13]
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm6, %ecx
+; X64-AVX512VLDQ-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm7, %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm9
+; X64-AVX512VLDQ-NEXT: vpsrld $16, %xmm6, %xmm10
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm10, %ecx
+; X64-AVX512VLDQ-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VLDQ-NEXT: vpsrld $16, %xmm7, %xmm10
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm10, %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm10
+; X64-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm4, %ecx
+; X64-AVX512VLDQ-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm5, %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm10
+; X64-AVX512VLDQ-NEXT: vpsrld $16, %xmm4, %xmm11
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm11, %ecx
+; X64-AVX512VLDQ-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VLDQ-NEXT: vpsrld $16, %xmm5, %xmm11
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm11, %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm11
+; X64-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
+; X64-AVX512VLDQ-NEXT: vinserti128 $1, %xmm9, %ymm10, %ymm9
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm2, %ecx
+; X64-AVX512VLDQ-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm3, %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm10
+; X64-AVX512VLDQ-NEXT: vpsrld $16, %xmm2, %xmm11
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm11, %ecx
+; X64-AVX512VLDQ-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VLDQ-NEXT: vpsrld $16, %xmm3, %xmm11
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm11, %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm11
+; X64-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm0, %ecx
+; X64-AVX512VLDQ-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm1, %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm11
+; X64-AVX512VLDQ-NEXT: vpsrld $16, %xmm0, %xmm12
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm12, %ecx
+; X64-AVX512VLDQ-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VLDQ-NEXT: vpsrld $16, %xmm1, %xmm12
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm12, %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm12
+; X64-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
+; X64-AVX512VLDQ-NEXT: vinserti128 $1, %xmm10, %ymm11, %ymm10
+; X64-AVX512VLDQ-NEXT: vinserti64x4 $1, %ymm9, %zmm10, %zmm9
+; X64-AVX512VLDQ-NEXT: vpsrlq $48, %xmm6, %xmm6
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm6, %ecx
+; X64-AVX512VLDQ-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VLDQ-NEXT: vpsrlq $48, %xmm7, %xmm6
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm6, %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm6
+; X64-AVX512VLDQ-NEXT: movl 52(%rsi), %ecx
+; X64-AVX512VLDQ-NEXT: andl %eax, %ecx
+; X64-AVX512VLDQ-NEXT: movzwl 52(%rdi), %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm7
+; X64-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
+; X64-AVX512VLDQ-NEXT: vpsrlq $48, %xmm4, %xmm4
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm4, %ecx
+; X64-AVX512VLDQ-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VLDQ-NEXT: vpsrlq $48, %xmm5, %xmm4
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm4, %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm4
+; X64-AVX512VLDQ-NEXT: movl 36(%rsi), %ecx
+; X64-AVX512VLDQ-NEXT: andl %eax, %ecx
+; X64-AVX512VLDQ-NEXT: movzwl 36(%rdi), %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm5
+; X64-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
+; X64-AVX512VLDQ-NEXT: vinserti128 $1, %xmm6, %ymm4, %ymm4
+; X64-AVX512VLDQ-NEXT: vpsrlq $48, %xmm2, %xmm2
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm2, %ecx
+; X64-AVX512VLDQ-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VLDQ-NEXT: vpsrlq $48, %xmm3, %xmm2
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm2, %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm2
+; X64-AVX512VLDQ-NEXT: movl 20(%rsi), %ecx
+; X64-AVX512VLDQ-NEXT: andl %eax, %ecx
+; X64-AVX512VLDQ-NEXT: movzwl 20(%rdi), %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm3
+; X64-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
+; X64-AVX512VLDQ-NEXT: vpsrlq $48, %xmm0, %xmm0
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm0, %ecx
+; X64-AVX512VLDQ-NEXT: andl $-32768, %ecx # imm = 0x8000
+; X64-AVX512VLDQ-NEXT: vpsrlq $48, %xmm1, %xmm0
+; X64-AVX512VLDQ-NEXT: vpextrw $0, %xmm0, %edx
+; X64-AVX512VLDQ-NEXT: andl $32767, %edx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %ecx, %edx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %edx, %xmm0, %xmm0
+; X64-AVX512VLDQ-NEXT: andl 4(%rsi), %eax
+; X64-AVX512VLDQ-NEXT: movzwl 4(%rdi), %ecx
+; X64-AVX512VLDQ-NEXT: andl $32767, %ecx # imm = 0x7FFF
+; X64-AVX512VLDQ-NEXT: orl %eax, %ecx
+; X64-AVX512VLDQ-NEXT: vpinsrw $0, %ecx, %xmm0, %xmm1
+; X64-AVX512VLDQ-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; X64-AVX512VLDQ-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm0
+; X64-AVX512VLDQ-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm0
+; X64-AVX512VLDQ-NEXT: vpunpckldq {{.*#+}} zmm0 = zmm9[0],zmm0[0],zmm9[1],zmm0[1],zmm9[4],zmm0[4],zmm9[5],zmm0[5],zmm9[8],zmm0[8],zmm9[9],zmm0[9],zmm9[12],zmm0[12],zmm9[13],zmm0[13]
+; X64-AVX512VLDQ-NEXT: vpunpcklqdq {{.*#+}} zmm0 = zmm0[0],zmm8[0],zmm0[2],zmm8[2],zmm0[4],zmm8[4],zmm0[6],zmm8[6]
+; X64-AVX512VLDQ-NEXT: retq
+ %a0 = load <32 x half>, ptr %p0, align 16
+ %a1 = load <32 x half>, ptr %p1, align 16
+ %t = call <32 x half> @llvm.copysign.v32f16(<32 x half> %a0, <32 x half> %a1)
+ ret <32 x half> %t
+}
+declare <32 x half> @llvm.copysign.v32f16(<32 x half>, <32 x half>)
+;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+; X64: {{.*}}
+; X64-AVX: {{.*}}
+; X86: {{.*}}
+; X86-AVX: {{.*}}
More information about the llvm-commits
mailing list