; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=fast-variable-shuffle,avx512vl,avx512bw,avx512dq,prefer-256-bit | FileCheck %s --check-prefixes=CHECK,CHECK-AVX512
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=fast-variable-shuffle,avx512vl,avx512bw,avx512dq,prefer-256-bit,avx512vbmi | FileCheck %s --check-prefixes=CHECK,CHECK-VBMI
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=skylake-avx512 -mattr=prefer-256-bit | FileCheck %s --check-prefixes=CHECK,CHECK-AVX512
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=skylake-avx512 -mattr=prefer-256-bit,avx512vbmi | FileCheck %s --check-prefixes=CHECK,CHECK-VBMI
; Make sure CPUs default to prefer-256-bit. avx512vnni isn't interesting as it just adds an isel peephole for vpmaddwd+vpaddd
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=skylake-avx512 | FileCheck %s --check-prefixes=CHECK,CHECK-AVX512
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=-avx512vnni -mcpu=cascadelake | FileCheck %s --check-prefixes=CHECK,CHECK-AVX512
define <64 x i1> @v64i1_argument_return(<64 x i1> %x) "min-legal-vector-width"="256" {
; CHECK-LABEL: v64i1_argument_return:
; CHECK: # %bb.0:
+; CHECK-NEXT: movq %rdi, %rax
+; CHECK-NEXT: kmovd %esi, %k0
+; CHECK-NEXT: kshiftlq $63, %k0, %k0
+; CHECK-NEXT: kshiftrq $63, %k0, %k0
+; CHECK-NEXT: kshiftlq $2, %k0, %k1
+; CHECK-NEXT: kmovd %edx, %k2
+; CHECK-NEXT: kshiftlq $1, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $62, %k0, %k0
+; CHECK-NEXT: kshiftrq $62, %k0, %k0
+; CHECK-NEXT: kshiftlq $3, %k0, %k1
+; CHECK-NEXT: kmovd %ecx, %k2
+; CHECK-NEXT: kshiftlq $2, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $61, %k0, %k0
+; CHECK-NEXT: kshiftrq $61, %k0, %k0
+; CHECK-NEXT: kshiftlq $4, %k0, %k1
+; CHECK-NEXT: kmovd %r8d, %k2
+; CHECK-NEXT: kshiftlq $3, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $60, %k0, %k0
+; CHECK-NEXT: kshiftrq $60, %k0, %k0
+; CHECK-NEXT: kshiftlq $5, %k0, %k1
+; CHECK-NEXT: kmovd %r9d, %k2
+; CHECK-NEXT: kshiftlq $4, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $59, %k0, %k0
+; CHECK-NEXT: kshiftrq $59, %k0, %k0
+; CHECK-NEXT: kshiftlq $6, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $5, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $58, %k0, %k0
+; CHECK-NEXT: kshiftrq $58, %k0, %k0
+; CHECK-NEXT: kshiftlq $7, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $6, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $57, %k0, %k0
+; CHECK-NEXT: kshiftrq $57, %k0, %k0
+; CHECK-NEXT: kshiftlq $8, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $7, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $56, %k0, %k0
+; CHECK-NEXT: kshiftrq $56, %k0, %k0
+; CHECK-NEXT: kshiftlq $9, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $8, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $55, %k0, %k0
+; CHECK-NEXT: kshiftrq $55, %k0, %k0
+; CHECK-NEXT: kshiftlq $10, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $9, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $54, %k0, %k0
+; CHECK-NEXT: kshiftrq $54, %k0, %k0
+; CHECK-NEXT: kshiftlq $11, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $10, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $53, %k0, %k0
+; CHECK-NEXT: kshiftrq $53, %k0, %k0
+; CHECK-NEXT: kshiftlq $12, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $11, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $52, %k0, %k0
+; CHECK-NEXT: kshiftrq $52, %k0, %k0
+; CHECK-NEXT: kshiftlq $13, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $12, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $51, %k0, %k0
+; CHECK-NEXT: kshiftrq $51, %k0, %k0
+; CHECK-NEXT: kshiftlq $14, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $13, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $50, %k0, %k0
+; CHECK-NEXT: kshiftrq $50, %k0, %k0
+; CHECK-NEXT: kshiftlq $15, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $14, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $49, %k0, %k0
+; CHECK-NEXT: kshiftrq $49, %k0, %k0
+; CHECK-NEXT: kshiftlq $16, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $15, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $48, %k0, %k0
+; CHECK-NEXT: kshiftrq $48, %k0, %k0
+; CHECK-NEXT: kshiftlq $17, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $16, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $47, %k0, %k0
+; CHECK-NEXT: kshiftrq $47, %k0, %k0
+; CHECK-NEXT: kshiftlq $18, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $17, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $46, %k0, %k0
+; CHECK-NEXT: kshiftrq $46, %k0, %k0
+; CHECK-NEXT: kshiftlq $19, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $18, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $45, %k0, %k0
+; CHECK-NEXT: kshiftrq $45, %k0, %k0
+; CHECK-NEXT: kshiftlq $20, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $19, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $44, %k0, %k0
+; CHECK-NEXT: kshiftrq $44, %k0, %k0
+; CHECK-NEXT: kshiftlq $21, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $20, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $43, %k0, %k0
+; CHECK-NEXT: kshiftrq $43, %k0, %k0
+; CHECK-NEXT: kshiftlq $22, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $21, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $42, %k0, %k0
+; CHECK-NEXT: kshiftrq $42, %k0, %k0
+; CHECK-NEXT: kshiftlq $23, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $22, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $41, %k0, %k0
+; CHECK-NEXT: kshiftrq $41, %k0, %k0
+; CHECK-NEXT: kshiftlq $24, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $23, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $40, %k0, %k0
+; CHECK-NEXT: kshiftrq $40, %k0, %k0
+; CHECK-NEXT: kshiftlq $25, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $24, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $39, %k0, %k0
+; CHECK-NEXT: kshiftrq $39, %k0, %k0
+; CHECK-NEXT: kshiftlq $26, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $25, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $38, %k0, %k0
+; CHECK-NEXT: kshiftrq $38, %k0, %k0
+; CHECK-NEXT: kshiftlq $27, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $26, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $37, %k0, %k0
+; CHECK-NEXT: kshiftrq $37, %k0, %k0
+; CHECK-NEXT: kshiftlq $28, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $27, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $36, %k0, %k0
+; CHECK-NEXT: kshiftrq $36, %k0, %k0
+; CHECK-NEXT: kshiftlq $29, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $28, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $35, %k0, %k0
+; CHECK-NEXT: kshiftrq $35, %k0, %k0
+; CHECK-NEXT: kshiftlq $30, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $29, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $34, %k0, %k0
+; CHECK-NEXT: kshiftrq $34, %k0, %k0
+; CHECK-NEXT: kshiftlq $31, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $30, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $33, %k0, %k0
+; CHECK-NEXT: kshiftrq $33, %k0, %k0
+; CHECK-NEXT: kshiftlq $32, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $31, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $32, %k0, %k0
+; CHECK-NEXT: kshiftrq $32, %k0, %k0
+; CHECK-NEXT: kshiftlq $33, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $32, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $31, %k0, %k0
+; CHECK-NEXT: kshiftrq $31, %k0, %k0
+; CHECK-NEXT: kshiftlq $34, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $33, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $30, %k0, %k0
+; CHECK-NEXT: kshiftrq $30, %k0, %k0
+; CHECK-NEXT: kshiftlq $35, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $34, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $29, %k0, %k0
+; CHECK-NEXT: kshiftrq $29, %k0, %k0
+; CHECK-NEXT: kshiftlq $36, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $35, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $28, %k0, %k0
+; CHECK-NEXT: kshiftrq $28, %k0, %k0
+; CHECK-NEXT: kshiftlq $37, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $36, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $27, %k0, %k0
+; CHECK-NEXT: kshiftrq $27, %k0, %k0
+; CHECK-NEXT: kshiftlq $38, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $37, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $26, %k0, %k0
+; CHECK-NEXT: kshiftrq $26, %k0, %k0
+; CHECK-NEXT: kshiftlq $39, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $38, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $25, %k0, %k0
+; CHECK-NEXT: kshiftrq $25, %k0, %k0
+; CHECK-NEXT: kshiftlq $40, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $39, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $24, %k0, %k0
+; CHECK-NEXT: kshiftrq $24, %k0, %k0
+; CHECK-NEXT: kshiftlq $41, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $40, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $23, %k0, %k0
+; CHECK-NEXT: kshiftrq $23, %k0, %k0
+; CHECK-NEXT: kshiftlq $42, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $41, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $22, %k0, %k0
+; CHECK-NEXT: kshiftrq $22, %k0, %k0
+; CHECK-NEXT: kshiftlq $43, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $42, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $21, %k0, %k0
+; CHECK-NEXT: kshiftrq $21, %k0, %k0
+; CHECK-NEXT: kshiftlq $44, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $43, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $20, %k0, %k0
+; CHECK-NEXT: kshiftrq $20, %k0, %k0
+; CHECK-NEXT: kshiftlq $45, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $44, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $19, %k0, %k0
+; CHECK-NEXT: kshiftrq $19, %k0, %k0
+; CHECK-NEXT: kshiftlq $46, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $45, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $18, %k0, %k0
+; CHECK-NEXT: kshiftrq $18, %k0, %k0
+; CHECK-NEXT: kshiftlq $47, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $46, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $17, %k0, %k0
+; CHECK-NEXT: kshiftrq $17, %k0, %k0
+; CHECK-NEXT: kshiftlq $48, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $47, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $16, %k0, %k0
+; CHECK-NEXT: kshiftrq $16, %k0, %k0
+; CHECK-NEXT: kshiftlq $49, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $48, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $15, %k0, %k0
+; CHECK-NEXT: kshiftrq $15, %k0, %k0
+; CHECK-NEXT: kshiftlq $50, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $49, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $14, %k0, %k0
+; CHECK-NEXT: kshiftrq $14, %k0, %k0
+; CHECK-NEXT: kshiftlq $51, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $50, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $13, %k0, %k0
+; CHECK-NEXT: kshiftrq $13, %k0, %k0
+; CHECK-NEXT: kshiftlq $52, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $51, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $12, %k0, %k0
+; CHECK-NEXT: kshiftrq $12, %k0, %k0
+; CHECK-NEXT: kshiftlq $53, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $52, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $11, %k0, %k0
+; CHECK-NEXT: kshiftrq $11, %k0, %k0
+; CHECK-NEXT: kshiftlq $54, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $53, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $10, %k0, %k0
+; CHECK-NEXT: kshiftrq $10, %k0, %k0
+; CHECK-NEXT: kshiftlq $55, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $54, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $9, %k0, %k0
+; CHECK-NEXT: kshiftrq $9, %k0, %k0
+; CHECK-NEXT: kshiftlq $56, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $55, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $8, %k0, %k0
+; CHECK-NEXT: kshiftrq $8, %k0, %k0
+; CHECK-NEXT: kshiftlq $57, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $56, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $7, %k0, %k0
+; CHECK-NEXT: kshiftrq $7, %k0, %k0
+; CHECK-NEXT: kshiftlq $58, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $57, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $6, %k0, %k0
+; CHECK-NEXT: kshiftrq $6, %k0, %k0
+; CHECK-NEXT: kshiftlq $59, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $58, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $5, %k0, %k0
+; CHECK-NEXT: kshiftrq $5, %k0, %k0
+; CHECK-NEXT: kshiftlq $60, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $59, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $4, %k0, %k0
+; CHECK-NEXT: kshiftrq $4, %k0, %k0
+; CHECK-NEXT: kshiftlq $61, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $60, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $3, %k0, %k0
+; CHECK-NEXT: kshiftrq $3, %k0, %k0
+; CHECK-NEXT: kshiftlq $62, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $61, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kshiftlq $2, %k0, %k0
+; CHECK-NEXT: kshiftrq $2, %k0, %k0
+; CHECK-NEXT: kshiftlq $63, %k0, %k1
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k2
+; CHECK-NEXT: kshiftlq $62, %k2, %k2
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kmovb {{[0-9]+}}(%rsp), %k1
+; CHECK-NEXT: kshiftlq $1, %k0, %k0
+; CHECK-NEXT: kshiftrq $1, %k0, %k0
+; CHECK-NEXT: kshiftlq $63, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k0
+; CHECK-NEXT: kmovq %k0, (%rdi)
; CHECK-NEXT: retq
ret <64 x i1> %x
}
define void @v64i1_shuffle(<64 x i8>* %x, <64 x i8>* %y) "min-legal-vector-width"="256" {
; CHECK-LABEL: v64i1_shuffle:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: vmovdqa (%rdi), %ymm0
-; CHECK-NEXT: vmovdqa 32(%rdi), %ymm1
+; CHECK-NEXT: vmovdqa (%rdi), %ymm1
+; CHECK-NEXT: vmovdqa 32(%rdi), %ymm0
; CHECK-NEXT: vptestnmb %ymm1, %ymm1, %k0
+; CHECK-NEXT: kshiftrd $3, %k0, %k1
+; CHECK-NEXT: kshiftlq $2, %k0, %k2
+; CHECK-NEXT: kshiftlq $1, %k0, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $1, %k0, %k3
+; CHECK-NEXT: kshiftlq $63, %k3, %k3
+; CHECK-NEXT: kshiftrq $63, %k3, %k3
+; CHECK-NEXT: korq %k2, %k3, %k2
+; CHECK-NEXT: kshiftlq $3, %k0, %k3
+; CHECK-NEXT: kshiftlq $2, %k1, %k1
+; CHECK-NEXT: korq %k1, %k3, %k1
+; CHECK-NEXT: kshiftrd $2, %k0, %k3
+; CHECK-NEXT: kshiftlq $62, %k2, %k2
+; CHECK-NEXT: kshiftrq $62, %k2, %k2
+; CHECK-NEXT: korq %k1, %k2, %k1
+; CHECK-NEXT: kshiftlq $4, %k0, %k2
+; CHECK-NEXT: kshiftlq $3, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $5, %k0, %k3
+; CHECK-NEXT: kshiftlq $61, %k1, %k1
+; CHECK-NEXT: kshiftrq $61, %k1, %k1
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: kshiftlq $5, %k0, %k2
+; CHECK-NEXT: kshiftlq $4, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $4, %k0, %k3
+; CHECK-NEXT: kshiftlq $60, %k1, %k1
+; CHECK-NEXT: kshiftrq $60, %k1, %k1
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: kshiftlq $6, %k0, %k2
+; CHECK-NEXT: kshiftlq $5, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $7, %k0, %k3
+; CHECK-NEXT: kshiftlq $59, %k1, %k1
+; CHECK-NEXT: kshiftrq $59, %k1, %k1
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: kshiftlq $7, %k0, %k2
+; CHECK-NEXT: kshiftlq $6, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $6, %k0, %k3
+; CHECK-NEXT: kshiftlq $58, %k1, %k1
+; CHECK-NEXT: kshiftrq $58, %k1, %k1
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: kshiftlq $8, %k0, %k2
+; CHECK-NEXT: kshiftlq $7, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $9, %k0, %k3
+; CHECK-NEXT: kshiftlq $57, %k1, %k1
+; CHECK-NEXT: kshiftrq $57, %k1, %k1
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: kshiftlq $9, %k0, %k2
+; CHECK-NEXT: kshiftlq $8, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $8, %k0, %k3
+; CHECK-NEXT: kshiftlq $56, %k1, %k1
+; CHECK-NEXT: kshiftrq $56, %k1, %k1
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: kshiftlq $10, %k0, %k2
+; CHECK-NEXT: kshiftlq $9, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $11, %k0, %k3
+; CHECK-NEXT: kshiftlq $55, %k1, %k1
+; CHECK-NEXT: kshiftrq $55, %k1, %k1
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: kshiftlq $11, %k0, %k2
+; CHECK-NEXT: kshiftlq $10, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $10, %k0, %k3
+; CHECK-NEXT: kshiftlq $54, %k1, %k1
+; CHECK-NEXT: kshiftrq $54, %k1, %k1
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: kshiftlq $12, %k0, %k2
+; CHECK-NEXT: kshiftlq $11, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $13, %k0, %k3
+; CHECK-NEXT: kshiftlq $53, %k1, %k1
+; CHECK-NEXT: kshiftrq $53, %k1, %k1
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: kshiftlq $13, %k0, %k2
+; CHECK-NEXT: kshiftlq $12, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $12, %k0, %k3
+; CHECK-NEXT: kshiftlq $52, %k1, %k1
+; CHECK-NEXT: kshiftrq $52, %k1, %k1
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: kshiftlq $14, %k0, %k2
+; CHECK-NEXT: kshiftlq $13, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $15, %k0, %k3
+; CHECK-NEXT: kshiftlq $51, %k1, %k1
+; CHECK-NEXT: kshiftrq $51, %k1, %k1
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: kshiftlq $15, %k0, %k2
+; CHECK-NEXT: kshiftlq $14, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $14, %k0, %k3
+; CHECK-NEXT: kshiftlq $50, %k1, %k1
+; CHECK-NEXT: kshiftrq $50, %k1, %k1
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: kshiftlq $16, %k0, %k2
+; CHECK-NEXT: kshiftlq $15, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $17, %k0, %k3
+; CHECK-NEXT: kshiftlq $49, %k1, %k1
+; CHECK-NEXT: kshiftrq $49, %k1, %k1
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: kshiftlq $17, %k0, %k2
+; CHECK-NEXT: kshiftlq $16, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $16, %k0, %k3
+; CHECK-NEXT: kshiftlq $48, %k1, %k1
+; CHECK-NEXT: kshiftrq $48, %k1, %k1
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: kshiftlq $18, %k0, %k2
+; CHECK-NEXT: kshiftlq $17, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $19, %k0, %k3
+; CHECK-NEXT: kshiftlq $47, %k1, %k1
+; CHECK-NEXT: kshiftrq $47, %k1, %k1
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: kshiftlq $19, %k0, %k2
+; CHECK-NEXT: kshiftlq $18, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $18, %k0, %k3
+; CHECK-NEXT: kshiftlq $46, %k1, %k1
+; CHECK-NEXT: kshiftrq $46, %k1, %k1
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: kshiftlq $20, %k0, %k2
+; CHECK-NEXT: kshiftlq $19, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $21, %k0, %k3
+; CHECK-NEXT: kshiftlq $45, %k1, %k1
+; CHECK-NEXT: kshiftrq $45, %k1, %k1
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: kshiftlq $21, %k0, %k2
+; CHECK-NEXT: kshiftlq $20, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $20, %k0, %k3
+; CHECK-NEXT: kshiftlq $44, %k1, %k1
+; CHECK-NEXT: kshiftrq $44, %k1, %k1
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: kshiftlq $22, %k0, %k2
+; CHECK-NEXT: kshiftlq $21, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $23, %k0, %k3
+; CHECK-NEXT: kshiftlq $43, %k1, %k1
+; CHECK-NEXT: kshiftrq $43, %k1, %k1
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: kshiftlq $23, %k0, %k2
+; CHECK-NEXT: kshiftlq $22, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $22, %k0, %k3
+; CHECK-NEXT: kshiftlq $42, %k1, %k1
+; CHECK-NEXT: kshiftrq $42, %k1, %k1
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: kshiftlq $24, %k0, %k2
+; CHECK-NEXT: kshiftlq $23, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $25, %k0, %k3
+; CHECK-NEXT: kshiftlq $41, %k1, %k1
+; CHECK-NEXT: kshiftrq $41, %k1, %k1
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: kshiftlq $25, %k0, %k2
+; CHECK-NEXT: kshiftlq $24, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $24, %k0, %k3
+; CHECK-NEXT: kshiftlq $40, %k1, %k1
+; CHECK-NEXT: kshiftrq $40, %k1, %k1
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: kshiftlq $26, %k0, %k2
+; CHECK-NEXT: kshiftlq $25, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $27, %k0, %k3
+; CHECK-NEXT: kshiftlq $39, %k1, %k1
+; CHECK-NEXT: kshiftrq $39, %k1, %k1
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: kshiftlq $27, %k0, %k2
+; CHECK-NEXT: kshiftlq $26, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $26, %k0, %k3
+; CHECK-NEXT: kshiftlq $38, %k1, %k1
+; CHECK-NEXT: kshiftrq $38, %k1, %k1
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: kshiftlq $28, %k0, %k2
+; CHECK-NEXT: kshiftlq $27, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $29, %k0, %k3
+; CHECK-NEXT: kshiftlq $37, %k1, %k1
+; CHECK-NEXT: kshiftrq $37, %k1, %k1
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: kshiftlq $29, %k0, %k2
+; CHECK-NEXT: kshiftlq $28, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $28, %k0, %k3
+; CHECK-NEXT: kshiftlq $36, %k1, %k1
+; CHECK-NEXT: kshiftrq $36, %k1, %k1
+; CHECK-NEXT: korq %k2, %k1, %k1
+; CHECK-NEXT: kshiftlq $30, %k0, %k2
+; CHECK-NEXT: kshiftlq $29, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $31, %k0, %k3
+; CHECK-NEXT: kshiftlq $35, %k1, %k1
+; CHECK-NEXT: kshiftrq $35, %k1, %k1
+; CHECK-NEXT: korq %k2, %k1, %k2
+; CHECK-NEXT: kshiftlq $31, %k0, %k1
+; CHECK-NEXT: kshiftlq $30, %k3, %k3
+; CHECK-NEXT: korq %k3, %k1, %k3
; CHECK-NEXT: vptestnmb %ymm0, %ymm0, %k1
-; CHECK-NEXT: vpmovm2b %k1, %ymm2
-; CHECK-NEXT: vmovdqa {{.*#+}} ymm3 = [1,0,3,2,5,4,7,6,9,8,11,10,13,12,15,14,1,0,3,2,5,4,7,6,9,8,11,10,13,12,15,14]
-; CHECK-NEXT: vpshufb %ymm3, %ymm2, %ymm2
-; CHECK-NEXT: vpmovb2m %ymm2, %k1
-; CHECK-NEXT: vpmovm2b %k0, %ymm2
-; CHECK-NEXT: vpshufb %ymm3, %ymm2, %ymm2
-; CHECK-NEXT: vpmovb2m %ymm2, %k2
-; CHECK-NEXT: vmovdqu8 %ymm1, 32(%rsi) {%k2}
-; CHECK-NEXT: vmovdqu8 %ymm0, (%rsi) {%k1}
+; CHECK-NEXT: kshiftrd $30, %k0, %k0
+; CHECK-NEXT: kshiftlq $34, %k2, %k2
+; CHECK-NEXT: kshiftrq $34, %k2, %k2
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftlq $32, %k0, %k3
+; CHECK-NEXT: kshiftlq $31, %k0, %k0
+; CHECK-NEXT: korq %k0, %k3, %k0
+; CHECK-NEXT: kshiftrd $1, %k1, %k3
+; CHECK-NEXT: kshiftlq $33, %k2, %k2
+; CHECK-NEXT: kshiftrq $33, %k2, %k2
+; CHECK-NEXT: korq %k0, %k2, %k0
+; CHECK-NEXT: kshiftlq $32, %k0, %k0
+; CHECK-NEXT: kshiftrq $32, %k0, %k0
+; CHECK-NEXT: kshiftlq $33, %k0, %k2
+; CHECK-NEXT: kshiftlq $32, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: korq %k2, %k0, %k0
+; CHECK-NEXT: kshiftlq $34, %k0, %k2
+; CHECK-NEXT: kshiftlq $33, %k1, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $3, %k1, %k3
+; CHECK-NEXT: kshiftlq $31, %k0, %k0
+; CHECK-NEXT: kshiftrq $31, %k0, %k0
+; CHECK-NEXT: korq %k2, %k0, %k0
+; CHECK-NEXT: kshiftlq $35, %k0, %k2
+; CHECK-NEXT: kshiftlq $34, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $2, %k1, %k3
+; CHECK-NEXT: kshiftlq $30, %k0, %k0
+; CHECK-NEXT: kshiftrq $30, %k0, %k0
+; CHECK-NEXT: korq %k2, %k0, %k0
+; CHECK-NEXT: kshiftlq $36, %k0, %k2
+; CHECK-NEXT: kshiftlq $35, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $5, %k1, %k3
+; CHECK-NEXT: kshiftlq $29, %k0, %k0
+; CHECK-NEXT: kshiftrq $29, %k0, %k0
+; CHECK-NEXT: korq %k2, %k0, %k0
+; CHECK-NEXT: kshiftlq $37, %k0, %k2
+; CHECK-NEXT: kshiftlq $36, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $4, %k1, %k3
+; CHECK-NEXT: kshiftlq $28, %k0, %k0
+; CHECK-NEXT: kshiftrq $28, %k0, %k0
+; CHECK-NEXT: korq %k2, %k0, %k0
+; CHECK-NEXT: kshiftlq $38, %k0, %k2
+; CHECK-NEXT: kshiftlq $37, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $7, %k1, %k3
+; CHECK-NEXT: kshiftlq $27, %k0, %k0
+; CHECK-NEXT: kshiftrq $27, %k0, %k0
+; CHECK-NEXT: korq %k2, %k0, %k0
+; CHECK-NEXT: kshiftlq $39, %k0, %k2
+; CHECK-NEXT: kshiftlq $38, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $6, %k1, %k3
+; CHECK-NEXT: kshiftlq $26, %k0, %k0
+; CHECK-NEXT: kshiftrq $26, %k0, %k0
+; CHECK-NEXT: korq %k2, %k0, %k0
+; CHECK-NEXT: kshiftlq $40, %k0, %k2
+; CHECK-NEXT: kshiftlq $39, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $9, %k1, %k3
+; CHECK-NEXT: kshiftlq $25, %k0, %k0
+; CHECK-NEXT: kshiftrq $25, %k0, %k0
+; CHECK-NEXT: korq %k2, %k0, %k0
+; CHECK-NEXT: kshiftlq $41, %k0, %k2
+; CHECK-NEXT: kshiftlq $40, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $8, %k1, %k3
+; CHECK-NEXT: kshiftlq $24, %k0, %k0
+; CHECK-NEXT: kshiftrq $24, %k0, %k0
+; CHECK-NEXT: korq %k2, %k0, %k0
+; CHECK-NEXT: kshiftlq $42, %k0, %k2
+; CHECK-NEXT: kshiftlq $41, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $11, %k1, %k3
+; CHECK-NEXT: kshiftlq $23, %k0, %k0
+; CHECK-NEXT: kshiftrq $23, %k0, %k0
+; CHECK-NEXT: korq %k2, %k0, %k0
+; CHECK-NEXT: kshiftlq $43, %k0, %k2
+; CHECK-NEXT: kshiftlq $42, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $10, %k1, %k3
+; CHECK-NEXT: kshiftlq $22, %k0, %k0
+; CHECK-NEXT: kshiftrq $22, %k0, %k0
+; CHECK-NEXT: korq %k2, %k0, %k0
+; CHECK-NEXT: kshiftlq $44, %k0, %k2
+; CHECK-NEXT: kshiftlq $43, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $13, %k1, %k3
+; CHECK-NEXT: kshiftlq $21, %k0, %k0
+; CHECK-NEXT: kshiftrq $21, %k0, %k0
+; CHECK-NEXT: korq %k2, %k0, %k0
+; CHECK-NEXT: kshiftlq $45, %k0, %k2
+; CHECK-NEXT: kshiftlq $44, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $12, %k1, %k3
+; CHECK-NEXT: kshiftlq $20, %k0, %k0
+; CHECK-NEXT: kshiftrq $20, %k0, %k0
+; CHECK-NEXT: korq %k2, %k0, %k0
+; CHECK-NEXT: kshiftlq $46, %k0, %k2
+; CHECK-NEXT: kshiftlq $45, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $15, %k1, %k3
+; CHECK-NEXT: kshiftlq $19, %k0, %k0
+; CHECK-NEXT: kshiftrq $19, %k0, %k0
+; CHECK-NEXT: korq %k2, %k0, %k0
+; CHECK-NEXT: kshiftlq $47, %k0, %k2
+; CHECK-NEXT: kshiftlq $46, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $14, %k1, %k3
+; CHECK-NEXT: kshiftlq $18, %k0, %k0
+; CHECK-NEXT: kshiftrq $18, %k0, %k0
+; CHECK-NEXT: korq %k2, %k0, %k0
+; CHECK-NEXT: kshiftlq $48, %k0, %k2
+; CHECK-NEXT: kshiftlq $47, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $17, %k1, %k3
+; CHECK-NEXT: kshiftlq $17, %k0, %k0
+; CHECK-NEXT: kshiftrq $17, %k0, %k0
+; CHECK-NEXT: korq %k2, %k0, %k0
+; CHECK-NEXT: kshiftlq $49, %k0, %k2
+; CHECK-NEXT: kshiftlq $48, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $16, %k1, %k3
+; CHECK-NEXT: kshiftlq $16, %k0, %k0
+; CHECK-NEXT: kshiftrq $16, %k0, %k0
+; CHECK-NEXT: korq %k2, %k0, %k0
+; CHECK-NEXT: kshiftlq $50, %k0, %k2
+; CHECK-NEXT: kshiftlq $49, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $19, %k1, %k3
+; CHECK-NEXT: kshiftlq $15, %k0, %k0
+; CHECK-NEXT: kshiftrq $15, %k0, %k0
+; CHECK-NEXT: korq %k2, %k0, %k0
+; CHECK-NEXT: kshiftlq $51, %k0, %k2
+; CHECK-NEXT: kshiftlq $50, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $18, %k1, %k3
+; CHECK-NEXT: kshiftlq $14, %k0, %k0
+; CHECK-NEXT: kshiftrq $14, %k0, %k0
+; CHECK-NEXT: korq %k2, %k0, %k0
+; CHECK-NEXT: kshiftlq $52, %k0, %k2
+; CHECK-NEXT: kshiftlq $51, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $21, %k1, %k3
+; CHECK-NEXT: kshiftlq $13, %k0, %k0
+; CHECK-NEXT: kshiftrq $13, %k0, %k0
+; CHECK-NEXT: korq %k2, %k0, %k0
+; CHECK-NEXT: kshiftlq $53, %k0, %k2
+; CHECK-NEXT: kshiftlq $52, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $20, %k1, %k3
+; CHECK-NEXT: kshiftlq $12, %k0, %k0
+; CHECK-NEXT: kshiftrq $12, %k0, %k0
+; CHECK-NEXT: korq %k2, %k0, %k0
+; CHECK-NEXT: kshiftlq $54, %k0, %k2
+; CHECK-NEXT: kshiftlq $53, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $23, %k1, %k3
+; CHECK-NEXT: kshiftlq $11, %k0, %k0
+; CHECK-NEXT: kshiftrq $11, %k0, %k0
+; CHECK-NEXT: korq %k2, %k0, %k0
+; CHECK-NEXT: kshiftlq $55, %k0, %k2
+; CHECK-NEXT: kshiftlq $54, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $22, %k1, %k3
+; CHECK-NEXT: kshiftlq $10, %k0, %k0
+; CHECK-NEXT: kshiftrq $10, %k0, %k0
+; CHECK-NEXT: korq %k2, %k0, %k0
+; CHECK-NEXT: kshiftlq $56, %k0, %k2
+; CHECK-NEXT: kshiftlq $55, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $25, %k1, %k3
+; CHECK-NEXT: kshiftlq $9, %k0, %k0
+; CHECK-NEXT: kshiftrq $9, %k0, %k0
+; CHECK-NEXT: korq %k2, %k0, %k0
+; CHECK-NEXT: kshiftlq $57, %k0, %k2
+; CHECK-NEXT: kshiftlq $56, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $24, %k1, %k3
+; CHECK-NEXT: kshiftlq $8, %k0, %k0
+; CHECK-NEXT: kshiftrq $8, %k0, %k0
+; CHECK-NEXT: korq %k2, %k0, %k0
+; CHECK-NEXT: kshiftlq $58, %k0, %k2
+; CHECK-NEXT: kshiftlq $57, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $27, %k1, %k3
+; CHECK-NEXT: kshiftlq $7, %k0, %k0
+; CHECK-NEXT: kshiftrq $7, %k0, %k0
+; CHECK-NEXT: korq %k2, %k0, %k0
+; CHECK-NEXT: kshiftlq $59, %k0, %k2
+; CHECK-NEXT: kshiftlq $58, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $26, %k1, %k3
+; CHECK-NEXT: kshiftlq $6, %k0, %k0
+; CHECK-NEXT: kshiftrq $6, %k0, %k0
+; CHECK-NEXT: korq %k2, %k0, %k0
+; CHECK-NEXT: kshiftlq $60, %k0, %k2
+; CHECK-NEXT: kshiftlq $59, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $29, %k1, %k3
+; CHECK-NEXT: kshiftlq $5, %k0, %k0
+; CHECK-NEXT: kshiftrq $5, %k0, %k0
+; CHECK-NEXT: korq %k2, %k0, %k0
+; CHECK-NEXT: kshiftlq $61, %k0, %k2
+; CHECK-NEXT: kshiftlq $60, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $28, %k1, %k3
+; CHECK-NEXT: kshiftlq $4, %k0, %k0
+; CHECK-NEXT: kshiftrq $4, %k0, %k0
+; CHECK-NEXT: korq %k2, %k0, %k0
+; CHECK-NEXT: kshiftlq $62, %k0, %k2
+; CHECK-NEXT: kshiftlq $61, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftrd $31, %k1, %k3
+; CHECK-NEXT: kshiftlq $3, %k0, %k0
+; CHECK-NEXT: kshiftrq $3, %k0, %k0
+; CHECK-NEXT: korq %k2, %k0, %k0
+; CHECK-NEXT: kshiftlq $63, %k0, %k2
+; CHECK-NEXT: kshiftlq $62, %k3, %k3
+; CHECK-NEXT: korq %k3, %k2, %k2
+; CHECK-NEXT: kshiftlq $2, %k0, %k0
+; CHECK-NEXT: kshiftrq $2, %k0, %k0
+; CHECK-NEXT: korq %k2, %k0, %k0
+; CHECK-NEXT: kshiftrd $30, %k1, %k1
+; CHECK-NEXT: kshiftlq $1, %k0, %k0
+; CHECK-NEXT: kshiftrq $1, %k0, %k0
+; CHECK-NEXT: kshiftlq $63, %k1, %k1
+; CHECK-NEXT: korq %k1, %k0, %k1
+; CHECK-NEXT: vmovdqu8 %ymm1, (%rsi) {%k1}
+; CHECK-NEXT: kshiftrq $32, %k1, %k1
+; CHECK-NEXT: vmovdqu8 %ymm0, 32(%rsi) {%k1}
; CHECK-NEXT: vzeroupper
; CHECK-NEXT: retq
entry:
}
declare void @llvm.masked.store.v64i8.p0v64i8(<64 x i8>, <64 x i8>*, i32, <64 x i1>)
+@mem64_dst = global i64 0, align 8
+@mem64_src = global i64 0, align 8
+define i32 @v64i1_inline_asm() "min-legal-vector-width"="256" {
+; CHECK-LABEL: v64i1_inline_asm:
+; CHECK: # %bb.0:
+; CHECK-NEXT: kmovq {{.*}}(%rip), %k0
+; CHECK-NEXT: #APP
+; CHECK-NEXT: #NO_APP
+; CHECK-NEXT: kmovq %k0, {{.*}}(%rip)
+; CHECK-NEXT: movl -{{[0-9]+}}(%rsp), %eax
+; CHECK-NEXT: retq
+ %1 = alloca i32, align 4
+ %2 = load i64, i64* @mem64_src, align 8
+ %3 = call i64 asm "", "=k,k,~{dirflag},~{fpsr},~{flags}"(i64 %2)
+ store i64 %3, i64* @mem64_dst, align 8
+ %4 = load i32, i32* %1, align 4
+ ret i32 %4
+}