llvm/test/CodeGen/X86/midpoint-int-vec-512.ll

; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f | FileCheck %s --check-prefixes=ALL,AVX512F
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512vl | FileCheck %s --check-prefixes=ALL,AVX512VL-FALLBACK
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512bw | FileCheck %s --check-prefixes=ALL,AVX512BW
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512vl,+avx512bw | FileCheck %s --check-prefixes=ALL,AVX512BW

; These test cases are inspired by C++2a std::midpoint().
; See https://bugs.llvm.org/show_bug.cgi?id=40965

; Using 512-bit vector regs.

; ---------------------------------------------------------------------------- ;
; 32-bit width. 512 / 32 = 16 elts.
; ---------------------------------------------------------------------------- ;

; Values come from regs

define <16 x i32> @vec512_i32_signed_reg_reg(<16 x i32> %a1, <16 x i32> %a2) nounwind {
; ALL-LABEL: vec512_i32_signed_reg_reg:
; ALL:       # %bb.0:
; ALL-NEXT:    vpminsd %zmm1, %zmm0, %zmm2
; ALL-NEXT:    vpmaxsd %zmm1, %zmm0, %zmm1
; ALL-NEXT:    vpsubd %zmm2, %zmm1, %zmm1
; ALL-NEXT:    vpsrld $1, %zmm1, %zmm1
; ALL-NEXT:    vpmulld %zmm1, %zmm1, %zmm1
; ALL-NEXT:    vpaddd %zmm0, %zmm1, %zmm0
; ALL-NEXT:    retq
  %t3 = icmp sgt <16 x i32> %a1, %a2 ; signed
  %t4 = select <16 x i1> %t3, <16 x i32> <i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1>, <16 x i32> <i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1>
  %t5 = select <16 x i1> %t3, <16 x i32> %a2, <16 x i32> %a1
  %t6 = select <16 x i1> %t3, <16 x i32> %a1, <16 x i32> %a2
  %t7 = sub <16 x i32> %t6, %t5
  %t16 = lshr <16 x i32> %t7, <i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1>
  %t9 = mul nsw <16 x i32> %t16, %t16 ; signed
  %a10 = add nsw <16 x i32> %t9, %a1 ; signed
  ret <16 x i32> %a10
}

define <16 x i32> @vec512_i32_unsigned_reg_reg(<16 x i32> %a1, <16 x i32> %a2) nounwind {
; ALL-LABEL: vec512_i32_unsigned_reg_reg:
; ALL:       # %bb.0:
; ALL-NEXT:    vpminud %zmm1, %zmm0, %zmm2
; ALL-NEXT:    vpmaxud %zmm1, %zmm0, %zmm1
; ALL-NEXT:    vpsubd %zmm2, %zmm1, %zmm1
; ALL-NEXT:    vpsrld $1, %zmm1, %zmm1
; ALL-NEXT:    vpmulld %zmm1, %zmm1, %zmm1
; ALL-NEXT:    vpaddd %zmm0, %zmm1, %zmm0
; ALL-NEXT:    retq
  %t3 = icmp ugt <16 x i32> %a1, %a2
  %t4 = select <16 x i1> %t3, <16 x i32> <i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1>, <16 x i32> <i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1>
  %t5 = select <16 x i1> %t3, <16 x i32> %a2, <16 x i32> %a1
  %t6 = select <16 x i1> %t3, <16 x i32> %a1, <16 x i32> %a2
  %t7 = sub <16 x i32> %t6, %t5
  %t16 = lshr <16 x i32> %t7, <i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1>
  %t9 = mul <16 x i32> %t16, %t16
  %a10 = add <16 x i32> %t9, %a1
  ret <16 x i32> %a10
}

; Values are loaded. Only check signed case.

define <16 x i32> @vec512_i32_signed_mem_reg(ptr %a1_addr, <16 x i32> %a2) nounwind {
; ALL-LABEL: vec512_i32_signed_mem_reg:
; ALL:       # %bb.0:
; ALL-NEXT:    vmovdqa64 (%rdi), %zmm1
; ALL-NEXT:    vpminsd %zmm0, %zmm1, %zmm2
; ALL-NEXT:    vpmaxsd %zmm0, %zmm1, %zmm0
; ALL-NEXT:    vpsubd %zmm2, %zmm0, %zmm0
; ALL-NEXT:    vpsrld $1, %zmm0, %zmm0
; ALL-NEXT:    vpmulld %zmm0, %zmm0, %zmm0
; ALL-NEXT:    vpaddd %zmm1, %zmm0, %zmm0
; ALL-NEXT:    retq
  %a1 = load <16 x i32>, ptr %a1_addr
  %t3 = icmp sgt <16 x i32> %a1, %a2 ; signed
  %t4 = select <16 x i1> %t3, <16 x i32> <i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1>, <16 x i32> <i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1>
  %t5 = select <16 x i1> %t3, <16 x i32> %a2, <16 x i32> %a1
  %t6 = select <16 x i1> %t3, <16 x i32> %a1, <16 x i32> %a2
  %t7 = sub <16 x i32> %t6, %t5
  %t16 = lshr <16 x i32> %t7, <i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1>
  %t9 = mul nsw <16 x i32> %t16, %t16 ; signed
  %a10 = add nsw <16 x i32> %t9, %a1 ; signed
  ret <16 x i32> %a10
}

define <16 x i32> @vec512_i32_signed_reg_mem(<16 x i32> %a1, ptr %a2_addr) nounwind {
; ALL-LABEL: vec512_i32_signed_reg_mem:
; ALL:       # %bb.0:
; ALL-NEXT:    vmovdqa64 (%rdi), %zmm1
; ALL-NEXT:    vpminsd %zmm1, %zmm0, %zmm2
; ALL-NEXT:    vpmaxsd %zmm1, %zmm0, %zmm1
; ALL-NEXT:    vpsubd %zmm2, %zmm1, %zmm1
; ALL-NEXT:    vpsrld $1, %zmm1, %zmm1
; ALL-NEXT:    vpmulld %zmm1, %zmm1, %zmm1
; ALL-NEXT:    vpaddd %zmm0, %zmm1, %zmm0
; ALL-NEXT:    retq
  %a2 = load <16 x i32>, ptr %a2_addr
  %t3 = icmp sgt <16 x i32> %a1, %a2 ; signed
  %t4 = select <16 x i1> %t3, <16 x i32> <i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1>, <16 x i32> <i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1>
  %t5 = select <16 x i1> %t3, <16 x i32> %a2, <16 x i32> %a1
  %t6 = select <16 x i1> %t3, <16 x i32> %a1, <16 x i32> %a2
  %t7 = sub <16 x i32> %t6, %t5
  %t16 = lshr <16 x i32> %t7, <i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1>
  %t9 = mul nsw <16 x i32> %t16, %t16 ; signed
  %a10 = add nsw <16 x i32> %t9, %a1 ; signed
  ret <16 x i32> %a10
}

define <16 x i32> @vec512_i32_signed_mem_mem(ptr %a1_addr, ptr %a2_addr) nounwind {
; ALL-LABEL: vec512_i32_signed_mem_mem:
; ALL:       # %bb.0:
; ALL-NEXT:    vmovdqa64 (%rdi), %zmm0
; ALL-NEXT:    vmovdqa64 (%rsi), %zmm1
; ALL-NEXT:    vpminsd %zmm1, %zmm0, %zmm2
; ALL-NEXT:    vpmaxsd %zmm1, %zmm0, %zmm1
; ALL-NEXT:    vpsubd %zmm2, %zmm1, %zmm1
; ALL-NEXT:    vpsrld $1, %zmm1, %zmm1
; ALL-NEXT:    vpmulld %zmm1, %zmm1, %zmm1
; ALL-NEXT:    vpaddd %zmm0, %zmm1, %zmm0
; ALL-NEXT:    retq
  %a1 = load <16 x i32>, ptr %a1_addr
  %a2 = load <16 x i32>, ptr %a2_addr
  %t3 = icmp sgt <16 x i32> %a1, %a2 ; signed
  %t4 = select <16 x i1> %t3, <16 x i32> <i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1>, <16 x i32> <i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1>
  %t5 = select <16 x i1> %t3, <16 x i32> %a2, <16 x i32> %a1
  %t6 = select <16 x i1> %t3, <16 x i32> %a1, <16 x i32> %a2
  %t7 = sub <16 x i32> %t6, %t5
  %t16 = lshr <16 x i32> %t7, <i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1>
  %t9 = mul nsw <16 x i32> %t16, %t16 ; signed
  %a10 = add nsw <16 x i32> %t9, %a1 ; signed
  ret <16 x i32> %a10
}

; ---------------------------------------------------------------------------- ;
; 64-bit width. 512 / 64 = 8 elts.
; ---------------------------------------------------------------------------- ;

; Values come from regs

define <8 x i64> @vec512_i64_signed_reg_reg(<8 x i64> %a1, <8 x i64> %a2) nounwind {
; ALL-LABEL: vec512_i64_signed_reg_reg:
; ALL:       # %bb.0:
; ALL-NEXT:    vpcmpgtq %zmm1, %zmm0, %k1
; ALL-NEXT:    vpminsq %zmm1, %zmm0, %zmm2
; ALL-NEXT:    vpmaxsq %zmm1, %zmm0, %zmm1
; ALL-NEXT:    vpsubq %zmm2, %zmm1, %zmm1
; ALL-NEXT:    vpsrlq $1, %zmm1, %zmm1
; ALL-NEXT:    vpxor %xmm2, %xmm2, %xmm2
; ALL-NEXT:    vpsubq %zmm1, %zmm2, %zmm1 {%k1}
; ALL-NEXT:    vpaddq %zmm0, %zmm1, %zmm0
; ALL-NEXT:    retq
  %t3 = icmp sgt <8 x i64> %a1, %a2 ; signed
  %t4 = select <8 x i1> %t3, <8 x i64> <i64 -1, i64 -1, i64 -1, i64 -1, i64 -1, i64 -1, i64 -1, i64 -1>, <8 x i64> <i64 1, i64 1, i64 1, i64 1, i64 1, i64 1, i64 1, i64 1>
  %t5 = select <8 x i1> %t3, <8 x i64> %a2, <8 x i64> %a1
  %t6 = select <8 x i1> %t3, <8 x i64> %a1, <8 x i64> %a2
  %t7 = sub <8 x i64> %t6, %t5
  %t8 = lshr <8 x i64> %t7, <i64 1, i64 1, i64 1, i64 1, i64 1, i64 1, i64 1, i64 1>
  %t9 = mul nsw <8 x i64> %t8, %t4 ; signed
  %a10 = add nsw <8 x i64> %t9, %a1 ; signed
  ret <8 x i64> %a10
}

define <8 x i64> @vec512_i64_unsigned_reg_reg(<8 x i64> %a1, <8 x i64> %a2) nounwind {
; ALL-LABEL: vec512_i64_unsigned_reg_reg:
; ALL:       # %bb.0:
; ALL-NEXT:    vpcmpnleuq %zmm1, %zmm0, %k1
; ALL-NEXT:    vpminuq %zmm1, %zmm0, %zmm2
; ALL-NEXT:    vpmaxuq %zmm1, %zmm0, %zmm1
; ALL-NEXT:    vpsubq %zmm2, %zmm1, %zmm1
; ALL-NEXT:    vpsrlq $1, %zmm1, %zmm1
; ALL-NEXT:    vpxor %xmm2, %xmm2, %xmm2
; ALL-NEXT:    vpsubq %zmm1, %zmm2, %zmm1 {%k1}
; ALL-NEXT:    vpaddq %zmm0, %zmm1, %zmm0
; ALL-NEXT:    retq
  %t3 = icmp ugt <8 x i64> %a1, %a2
  %t4 = select <8 x i1> %t3, <8 x i64> <i64 -1, i64 -1, i64 -1, i64 -1, i64 -1, i64 -1, i64 -1, i64 -1>, <8 x i64> <i64 1, i64 1, i64 1, i64 1, i64 1, i64 1, i64 1, i64 1>
  %t5 = select <8 x i1> %t3, <8 x i64> %a2, <8 x i64> %a1
  %t6 = select <8 x i1> %t3, <8 x i64> %a1, <8 x i64> %a2
  %t7 = sub <8 x i64> %t6, %t5
  %t8 = lshr <8 x i64> %t7, <i64 1, i64 1, i64 1, i64 1, i64 1, i64 1, i64 1, i64 1>
  %t9 = mul <8 x i64> %t8, %t4
  %a10 = add <8 x i64> %t9, %a1
  ret <8 x i64> %a10
}

; Values are loaded. Only check signed case.

define <8 x i64> @vec512_i64_signed_mem_reg(ptr %a1_addr, <8 x i64> %a2) nounwind {
; ALL-LABEL: vec512_i64_signed_mem_reg:
; ALL:       # %bb.0:
; ALL-NEXT:    vmovdqa64 (%rdi), %zmm1
; ALL-NEXT:    vpcmpgtq %zmm0, %zmm1, %k1
; ALL-NEXT:    vpminsq %zmm0, %zmm1, %zmm2
; ALL-NEXT:    vpmaxsq %zmm0, %zmm1, %zmm0
; ALL-NEXT:    vpsubq %zmm2, %zmm0, %zmm0
; ALL-NEXT:    vpsrlq $1, %zmm0, %zmm0
; ALL-NEXT:    vpxor %xmm2, %xmm2, %xmm2
; ALL-NEXT:    vpsubq %zmm0, %zmm2, %zmm0 {%k1}
; ALL-NEXT:    vpaddq %zmm1, %zmm0, %zmm0
; ALL-NEXT:    retq
  %a1 = load <8 x i64>, ptr %a1_addr
  %t3 = icmp sgt <8 x i64> %a1, %a2 ; signed
  %t4 = select <8 x i1> %t3, <8 x i64> <i64 -1, i64 -1, i64 -1, i64 -1, i64 -1, i64 -1, i64 -1, i64 -1>, <8 x i64> <i64 1, i64 1, i64 1, i64 1, i64 1, i64 1, i64 1, i64 1>
  %t5 = select <8 x i1> %t3, <8 x i64> %a2, <8 x i64> %a1
  %t6 = select <8 x i1> %t3, <8 x i64> %a1, <8 x i64> %a2
  %t7 = sub <8 x i64> %t6, %t5
  %t8 = lshr <8 x i64> %t7, <i64 1, i64 1, i64 1, i64 1, i64 1, i64 1, i64 1, i64 1>
  %t9 = mul nsw <8 x i64> %t8, %t4 ; signed
  %a10 = add nsw <8 x i64> %t9, %a1 ; signed
  ret <8 x i64> %a10
}

define <8 x i64> @vec512_i64_signed_reg_mem(<8 x i64> %a1, ptr %a2_addr) nounwind {
; ALL-LABEL: vec512_i64_signed_reg_mem:
; ALL:       # %bb.0:
; ALL-NEXT:    vmovdqa64 (%rdi), %zmm1
; ALL-NEXT:    vpcmpgtq %zmm1, %zmm0, %k1
; ALL-NEXT:    vpminsq %zmm1, %zmm0, %zmm2
; ALL-NEXT:    vpmaxsq %zmm1, %zmm0, %zmm1
; ALL-NEXT:    vpsubq %zmm2, %zmm1, %zmm1
; ALL-NEXT:    vpsrlq $1, %zmm1, %zmm1
; ALL-NEXT:    vpxor %xmm2, %xmm2, %xmm2
; ALL-NEXT:    vpsubq %zmm1, %zmm2, %zmm1 {%k1}
; ALL-NEXT:    vpaddq %zmm0, %zmm1, %zmm0
; ALL-NEXT:    retq
  %a2 = load <8 x i64>, ptr %a2_addr
  %t3 = icmp sgt <8 x i64> %a1, %a2 ; signed
  %t4 = select <8 x i1> %t3, <8 x i64> <i64 -1, i64 -1, i64 -1, i64 -1, i64 -1, i64 -1, i64 -1, i64 -1>, <8 x i64> <i64 1, i64 1, i64 1, i64 1, i64 1, i64 1, i64 1, i64 1>
  %t5 = select <8 x i1> %t3, <8 x i64> %a2, <8 x i64> %a1
  %t6 = select <8 x i1> %t3, <8 x i64> %a1, <8 x i64> %a2
  %t7 = sub <8 x i64> %t6, %t5
  %t8 = lshr <8 x i64> %t7, <i64 1, i64 1, i64 1, i64 1, i64 1, i64 1, i64 1, i64 1>
  %t9 = mul nsw <8 x i64> %t8, %t4 ; signed
  %a10 = add nsw <8 x i64> %t9, %a1 ; signed
  ret <8 x i64> %a10
}

define <8 x i64> @vec512_i64_signed_mem_mem(ptr %a1_addr, ptr %a2_addr) nounwind {
; ALL-LABEL: vec512_i64_signed_mem_mem:
; ALL:       # %bb.0:
; ALL-NEXT:    vmovdqa64 (%rdi), %zmm0
; ALL-NEXT:    vmovdqa64 (%rsi), %zmm1
; ALL-NEXT:    vpcmpgtq %zmm1, %zmm0, %k1
; ALL-NEXT:    vpminsq %zmm1, %zmm0, %zmm2
; ALL-NEXT:    vpmaxsq %zmm1, %zmm0, %zmm1
; ALL-NEXT:    vpsubq %zmm2, %zmm1, %zmm1
; ALL-NEXT:    vpsrlq $1, %zmm1, %zmm1
; ALL-NEXT:    vpxor %xmm2, %xmm2, %xmm2
; ALL-NEXT:    vpsubq %zmm1, %zmm2, %zmm1 {%k1}
; ALL-NEXT:    vpaddq %zmm0, %zmm1, %zmm0
; ALL-NEXT:    retq
  %a1 = load <8 x i64>, ptr %a1_addr
  %a2 = load <8 x i64>, ptr %a2_addr
  %t3 = icmp sgt <8 x i64> %a1, %a2 ; signed
  %t4 = select <8 x i1> %t3, <8 x i64> <i64 -1, i64 -1, i64 -1, i64 -1, i64 -1, i64 -1, i64 -1, i64 -1>, <8 x i64> <i64 1, i64 1, i64 1, i64 1, i64 1, i64 1, i64 1, i64 1>
  %t5 = select <8 x i1> %t3, <8 x i64> %a2, <8 x i64> %a1
  %t6 = select <8 x i1> %t3, <8 x i64> %a1, <8 x i64> %a2
  %t7 = sub <8 x i64> %t6, %t5
  %t8 = lshr <8 x i64> %t7, <i64 1, i64 1, i64 1, i64 1, i64 1, i64 1, i64 1, i64 1>
  %t9 = mul nsw <8 x i64> %t8, %t4 ; signed
  %a10 = add nsw <8 x i64> %t9, %a1 ; signed
  ret <8 x i64> %a10
}

; ---------------------------------------------------------------------------- ;
; 16-bit width. 512 / 16 = 32 elts.
; ---------------------------------------------------------------------------- ;

; Values come from regs

define <32 x i16> @vec512_i16_signed_reg_reg(<32 x i16> %a1, <32 x i16> %a2) nounwind {
; AVX512F-LABEL: vec512_i16_signed_reg_reg:
; AVX512F:       # %bb.0:
; AVX512F-NEXT:    vextracti64x4 $1, %zmm1, %ymm2
; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm3
; AVX512F-NEXT:    vpcmpgtw %ymm2, %ymm3, %ymm4
; AVX512F-NEXT:    vpcmpgtw %ymm1, %ymm0, %ymm5
; AVX512F-NEXT:    vinserti64x4 $1, %ymm4, %zmm5, %zmm4
; AVX512F-NEXT:    vpminsw %ymm2, %ymm3, %ymm5
; AVX512F-NEXT:    vpmaxsw %ymm2, %ymm3, %ymm2
; AVX512F-NEXT:    vpsubw %ymm5, %ymm2, %ymm2
; AVX512F-NEXT:    vpminsw %ymm1, %ymm0, %ymm5
; AVX512F-NEXT:    vpmaxsw %ymm1, %ymm0, %ymm1
; AVX512F-NEXT:    vpsubw %ymm5, %ymm1, %ymm1
; AVX512F-NEXT:    vpsrlw $1, %ymm1, %ymm1
; AVX512F-NEXT:    vpsrlw $1, %ymm2, %ymm2
; AVX512F-NEXT:    vinserti64x4 $1, %ymm2, %zmm1, %zmm5
; AVX512F-NEXT:    vpxor %xmm6, %xmm6, %xmm6
; AVX512F-NEXT:    vpsubw %ymm2, %ymm6, %ymm2
; AVX512F-NEXT:    vpsubw %ymm1, %ymm6, %ymm1
; AVX512F-NEXT:    vinserti64x4 $1, %ymm2, %zmm1, %zmm1
; AVX512F-NEXT:    vpternlogq $226, %zmm5, %zmm4, %zmm1
; AVX512F-NEXT:    vextracti64x4 $1, %zmm1, %ymm2
; AVX512F-NEXT:    vpaddw %ymm3, %ymm2, %ymm2
; AVX512F-NEXT:    vpaddw %ymm0, %ymm1, %ymm0
; AVX512F-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm0
; AVX512F-NEXT:    retq
;
; AVX512VL-FALLBACK-LABEL: vec512_i16_signed_reg_reg:
; AVX512VL-FALLBACK:       # %bb.0:
; AVX512VL-FALLBACK-NEXT:    vextracti64x4 $1, %zmm1, %ymm2
; AVX512VL-FALLBACK-NEXT:    vextracti64x4 $1, %zmm0, %ymm3
; AVX512VL-FALLBACK-NEXT:    vpcmpgtw %ymm2, %ymm3, %ymm4
; AVX512VL-FALLBACK-NEXT:    vpcmpgtw %ymm1, %ymm0, %ymm5
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm4, %zmm5, %zmm4
; AVX512VL-FALLBACK-NEXT:    vpminsw %ymm2, %ymm3, %ymm5
; AVX512VL-FALLBACK-NEXT:    vpmaxsw %ymm2, %ymm3, %ymm2
; AVX512VL-FALLBACK-NEXT:    vpsubw %ymm5, %ymm2, %ymm2
; AVX512VL-FALLBACK-NEXT:    vpminsw %ymm1, %ymm0, %ymm5
; AVX512VL-FALLBACK-NEXT:    vpmaxsw %ymm1, %ymm0, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpsubw %ymm5, %ymm1, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpsrlw $1, %ymm1, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpsrlw $1, %ymm2, %ymm2
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm2, %zmm1, %zmm5
; AVX512VL-FALLBACK-NEXT:    vpxor %xmm6, %xmm6, %xmm6
; AVX512VL-FALLBACK-NEXT:    vpsubw %ymm2, %ymm6, %ymm2
; AVX512VL-FALLBACK-NEXT:    vpsubw %ymm1, %ymm6, %ymm1
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm2, %zmm1, %zmm1
; AVX512VL-FALLBACK-NEXT:    vpternlogq $226, %zmm5, %zmm4, %zmm1
; AVX512VL-FALLBACK-NEXT:    vextracti64x4 $1, %zmm1, %ymm2
; AVX512VL-FALLBACK-NEXT:    vpaddw %ymm3, %ymm2, %ymm2
; AVX512VL-FALLBACK-NEXT:    vpaddw %ymm0, %ymm1, %ymm0
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm0
; AVX512VL-FALLBACK-NEXT:    retq
;
; AVX512BW-LABEL: vec512_i16_signed_reg_reg:
; AVX512BW:       # %bb.0:
; AVX512BW-NEXT:    vpcmpgtw %zmm1, %zmm0, %k1
; AVX512BW-NEXT:    vpminsw %zmm1, %zmm0, %zmm2
; AVX512BW-NEXT:    vpmaxsw %zmm1, %zmm0, %zmm1
; AVX512BW-NEXT:    vpsubw %zmm2, %zmm1, %zmm1
; AVX512BW-NEXT:    vpsrlw $1, %zmm1, %zmm1
; AVX512BW-NEXT:    vpxor %xmm2, %xmm2, %xmm2
; AVX512BW-NEXT:    vpsubw %zmm1, %zmm2, %zmm1 {%k1}
; AVX512BW-NEXT:    vpaddw %zmm0, %zmm1, %zmm0
; AVX512BW-NEXT:    retq
  %t3 = icmp sgt <32 x i16> %a1, %a2 ; signed
  %t4 = select <32 x i1> %t3, <32 x i16> <i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1>, <32 x i16> <i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1>
  %t5 = select <32 x i1> %t3, <32 x i16> %a2, <32 x i16> %a1
  %t6 = select <32 x i1> %t3, <32 x i16> %a1, <32 x i16> %a2
  %t7 = sub <32 x i16> %t6, %t5
  %t16 = lshr <32 x i16> %t7, <i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1>
  %t9 = mul nsw <32 x i16> %t16, %t4 ; signed
  %a10 = add nsw <32 x i16> %t9, %a1 ; signed
  ret <32 x i16> %a10
}

define <32 x i16> @vec512_i16_unsigned_reg_reg(<32 x i16> %a1, <32 x i16> %a2) nounwind {
; AVX512F-LABEL: vec512_i16_unsigned_reg_reg:
; AVX512F:       # %bb.0:
; AVX512F-NEXT:    vextracti64x4 $1, %zmm1, %ymm2
; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm3
; AVX512F-NEXT:    vpminuw %ymm2, %ymm3, %ymm4
; AVX512F-NEXT:    vpcmpeqw %ymm4, %ymm3, %ymm5
; AVX512F-NEXT:    vpminuw %ymm1, %ymm0, %ymm6
; AVX512F-NEXT:    vpcmpeqw %ymm6, %ymm0, %ymm7
; AVX512F-NEXT:    vinserti64x4 $1, %ymm5, %zmm7, %zmm5
; AVX512F-NEXT:    vpmaxuw %ymm2, %ymm3, %ymm2
; AVX512F-NEXT:    vpsubw %ymm4, %ymm2, %ymm2
; AVX512F-NEXT:    vpmaxuw %ymm1, %ymm0, %ymm1
; AVX512F-NEXT:    vpsubw %ymm6, %ymm1, %ymm1
; AVX512F-NEXT:    vpsrlw $1, %ymm1, %ymm1
; AVX512F-NEXT:    vpsrlw $1, %ymm2, %ymm2
; AVX512F-NEXT:    vinserti64x4 $1, %ymm2, %zmm1, %zmm4
; AVX512F-NEXT:    vpxor %xmm6, %xmm6, %xmm6
; AVX512F-NEXT:    vpsubw %ymm2, %ymm6, %ymm2
; AVX512F-NEXT:    vpsubw %ymm1, %ymm6, %ymm1
; AVX512F-NEXT:    vinserti64x4 $1, %ymm2, %zmm1, %zmm1
; AVX512F-NEXT:    vpternlogq $184, %zmm4, %zmm5, %zmm1
; AVX512F-NEXT:    vextracti64x4 $1, %zmm1, %ymm2
; AVX512F-NEXT:    vpaddw %ymm3, %ymm2, %ymm2
; AVX512F-NEXT:    vpaddw %ymm0, %ymm1, %ymm0
; AVX512F-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm0
; AVX512F-NEXT:    retq
;
; AVX512VL-FALLBACK-LABEL: vec512_i16_unsigned_reg_reg:
; AVX512VL-FALLBACK:       # %bb.0:
; AVX512VL-FALLBACK-NEXT:    vextracti64x4 $1, %zmm1, %ymm2
; AVX512VL-FALLBACK-NEXT:    vextracti64x4 $1, %zmm0, %ymm3
; AVX512VL-FALLBACK-NEXT:    vpminuw %ymm2, %ymm3, %ymm4
; AVX512VL-FALLBACK-NEXT:    vpcmpeqw %ymm4, %ymm3, %ymm5
; AVX512VL-FALLBACK-NEXT:    vpminuw %ymm1, %ymm0, %ymm6
; AVX512VL-FALLBACK-NEXT:    vpcmpeqw %ymm6, %ymm0, %ymm7
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm5, %zmm7, %zmm5
; AVX512VL-FALLBACK-NEXT:    vpmaxuw %ymm2, %ymm3, %ymm2
; AVX512VL-FALLBACK-NEXT:    vpsubw %ymm4, %ymm2, %ymm2
; AVX512VL-FALLBACK-NEXT:    vpmaxuw %ymm1, %ymm0, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpsubw %ymm6, %ymm1, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpsrlw $1, %ymm1, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpsrlw $1, %ymm2, %ymm2
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm2, %zmm1, %zmm4
; AVX512VL-FALLBACK-NEXT:    vpxor %xmm6, %xmm6, %xmm6
; AVX512VL-FALLBACK-NEXT:    vpsubw %ymm2, %ymm6, %ymm2
; AVX512VL-FALLBACK-NEXT:    vpsubw %ymm1, %ymm6, %ymm1
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm2, %zmm1, %zmm1
; AVX512VL-FALLBACK-NEXT:    vpternlogq $184, %zmm4, %zmm5, %zmm1
; AVX512VL-FALLBACK-NEXT:    vextracti64x4 $1, %zmm1, %ymm2
; AVX512VL-FALLBACK-NEXT:    vpaddw %ymm3, %ymm2, %ymm2
; AVX512VL-FALLBACK-NEXT:    vpaddw %ymm0, %ymm1, %ymm0
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm0
; AVX512VL-FALLBACK-NEXT:    retq
;
; AVX512BW-LABEL: vec512_i16_unsigned_reg_reg:
; AVX512BW:       # %bb.0:
; AVX512BW-NEXT:    vpcmpnleuw %zmm1, %zmm0, %k1
; AVX512BW-NEXT:    vpminuw %zmm1, %zmm0, %zmm2
; AVX512BW-NEXT:    vpmaxuw %zmm1, %zmm0, %zmm1
; AVX512BW-NEXT:    vpsubw %zmm2, %zmm1, %zmm1
; AVX512BW-NEXT:    vpsrlw $1, %zmm1, %zmm1
; AVX512BW-NEXT:    vpxor %xmm2, %xmm2, %xmm2
; AVX512BW-NEXT:    vpsubw %zmm1, %zmm2, %zmm1 {%k1}
; AVX512BW-NEXT:    vpaddw %zmm0, %zmm1, %zmm0
; AVX512BW-NEXT:    retq
  %t3 = icmp ugt <32 x i16> %a1, %a2
  %t4 = select <32 x i1> %t3, <32 x i16> <i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1>, <32 x i16> <i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1>
  %t5 = select <32 x i1> %t3, <32 x i16> %a2, <32 x i16> %a1
  %t6 = select <32 x i1> %t3, <32 x i16> %a1, <32 x i16> %a2
  %t7 = sub <32 x i16> %t6, %t5
  %t16 = lshr <32 x i16> %t7, <i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1>
  %t9 = mul <32 x i16> %t16, %t4
  %a10 = add <32 x i16> %t9, %a1
  ret <32 x i16> %a10
}

; Values are loaded. Only check signed case.

define <32 x i16> @vec512_i16_signed_mem_reg(ptr %a1_addr, <32 x i16> %a2) nounwind {
; AVX512F-LABEL: vec512_i16_signed_mem_reg:
; AVX512F:       # %bb.0:
; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
; AVX512F-NEXT:    vmovdqa (%rdi), %ymm2
; AVX512F-NEXT:    vmovdqa 32(%rdi), %ymm3
; AVX512F-NEXT:    vpcmpgtw %ymm1, %ymm3, %ymm4
; AVX512F-NEXT:    vpcmpgtw %ymm0, %ymm2, %ymm5
; AVX512F-NEXT:    vinserti64x4 $1, %ymm4, %zmm5, %zmm4
; AVX512F-NEXT:    vpminsw %ymm1, %ymm3, %ymm5
; AVX512F-NEXT:    vpmaxsw %ymm1, %ymm3, %ymm1
; AVX512F-NEXT:    vpsubw %ymm5, %ymm1, %ymm1
; AVX512F-NEXT:    vpminsw %ymm0, %ymm2, %ymm5
; AVX512F-NEXT:    vpmaxsw %ymm0, %ymm2, %ymm0
; AVX512F-NEXT:    vpsubw %ymm5, %ymm0, %ymm0
; AVX512F-NEXT:    vpsrlw $1, %ymm0, %ymm0
; AVX512F-NEXT:    vpsrlw $1, %ymm1, %ymm1
; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm5
; AVX512F-NEXT:    vpxor %xmm6, %xmm6, %xmm6
; AVX512F-NEXT:    vpsubw %ymm1, %ymm6, %ymm1
; AVX512F-NEXT:    vpsubw %ymm0, %ymm6, %ymm0
; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512F-NEXT:    vpternlogq $226, %zmm5, %zmm4, %zmm0
; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
; AVX512F-NEXT:    vpaddw %ymm3, %ymm1, %ymm1
; AVX512F-NEXT:    vpaddw %ymm2, %ymm0, %ymm0
; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512F-NEXT:    retq
;
; AVX512VL-FALLBACK-LABEL: vec512_i16_signed_mem_reg:
; AVX512VL-FALLBACK:       # %bb.0:
; AVX512VL-FALLBACK-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
; AVX512VL-FALLBACK-NEXT:    vmovdqa (%rdi), %ymm2
; AVX512VL-FALLBACK-NEXT:    vmovdqa 32(%rdi), %ymm3
; AVX512VL-FALLBACK-NEXT:    vpcmpgtw %ymm1, %ymm3, %ymm4
; AVX512VL-FALLBACK-NEXT:    vpcmpgtw %ymm0, %ymm2, %ymm5
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm4, %zmm5, %zmm4
; AVX512VL-FALLBACK-NEXT:    vpminsw %ymm1, %ymm3, %ymm5
; AVX512VL-FALLBACK-NEXT:    vpmaxsw %ymm1, %ymm3, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpsubw %ymm5, %ymm1, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpminsw %ymm0, %ymm2, %ymm5
; AVX512VL-FALLBACK-NEXT:    vpmaxsw %ymm0, %ymm2, %ymm0
; AVX512VL-FALLBACK-NEXT:    vpsubw %ymm5, %ymm0, %ymm0
; AVX512VL-FALLBACK-NEXT:    vpsrlw $1, %ymm0, %ymm0
; AVX512VL-FALLBACK-NEXT:    vpsrlw $1, %ymm1, %ymm1
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm5
; AVX512VL-FALLBACK-NEXT:    vpxor %xmm6, %xmm6, %xmm6
; AVX512VL-FALLBACK-NEXT:    vpsubw %ymm1, %ymm6, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpsubw %ymm0, %ymm6, %ymm0
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512VL-FALLBACK-NEXT:    vpternlogq $226, %zmm5, %zmm4, %zmm0
; AVX512VL-FALLBACK-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpaddw %ymm3, %ymm1, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpaddw %ymm2, %ymm0, %ymm0
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512VL-FALLBACK-NEXT:    retq
;
; AVX512BW-LABEL: vec512_i16_signed_mem_reg:
; AVX512BW:       # %bb.0:
; AVX512BW-NEXT:    vmovdqa64 (%rdi), %zmm1
; AVX512BW-NEXT:    vpcmpgtw %zmm0, %zmm1, %k1
; AVX512BW-NEXT:    vpminsw %zmm0, %zmm1, %zmm2
; AVX512BW-NEXT:    vpmaxsw %zmm0, %zmm1, %zmm0
; AVX512BW-NEXT:    vpsubw %zmm2, %zmm0, %zmm0
; AVX512BW-NEXT:    vpsrlw $1, %zmm0, %zmm0
; AVX512BW-NEXT:    vpxor %xmm2, %xmm2, %xmm2
; AVX512BW-NEXT:    vpsubw %zmm0, %zmm2, %zmm0 {%k1}
; AVX512BW-NEXT:    vpaddw %zmm1, %zmm0, %zmm0
; AVX512BW-NEXT:    retq
  %a1 = load <32 x i16>, ptr %a1_addr
  %t3 = icmp sgt <32 x i16> %a1, %a2 ; signed
  %t4 = select <32 x i1> %t3, <32 x i16> <i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1>, <32 x i16> <i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1>
  %t5 = select <32 x i1> %t3, <32 x i16> %a2, <32 x i16> %a1
  %t6 = select <32 x i1> %t3, <32 x i16> %a1, <32 x i16> %a2
  %t7 = sub <32 x i16> %t6, %t5
  %t16 = lshr <32 x i16> %t7, <i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1>
  %t9 = mul nsw <32 x i16> %t16, %t4 ; signed
  %a10 = add nsw <32 x i16> %t9, %a1 ; signed
  ret <32 x i16> %a10
}

define <32 x i16> @vec512_i16_signed_reg_mem(<32 x i16> %a1, ptr %a2_addr) nounwind {
; AVX512F-LABEL: vec512_i16_signed_reg_mem:
; AVX512F:       # %bb.0:
; AVX512F-NEXT:    vmovdqa (%rdi), %ymm1
; AVX512F-NEXT:    vmovdqa 32(%rdi), %ymm2
; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm3
; AVX512F-NEXT:    vpcmpgtw %ymm2, %ymm3, %ymm4
; AVX512F-NEXT:    vpcmpgtw %ymm1, %ymm0, %ymm5
; AVX512F-NEXT:    vinserti64x4 $1, %ymm4, %zmm5, %zmm4
; AVX512F-NEXT:    vpminsw %ymm2, %ymm3, %ymm5
; AVX512F-NEXT:    vpmaxsw %ymm2, %ymm3, %ymm2
; AVX512F-NEXT:    vpsubw %ymm5, %ymm2, %ymm2
; AVX512F-NEXT:    vpminsw %ymm1, %ymm0, %ymm5
; AVX512F-NEXT:    vpmaxsw %ymm1, %ymm0, %ymm1
; AVX512F-NEXT:    vpsubw %ymm5, %ymm1, %ymm1
; AVX512F-NEXT:    vpsrlw $1, %ymm1, %ymm1
; AVX512F-NEXT:    vpsrlw $1, %ymm2, %ymm2
; AVX512F-NEXT:    vinserti64x4 $1, %ymm2, %zmm1, %zmm5
; AVX512F-NEXT:    vpxor %xmm6, %xmm6, %xmm6
; AVX512F-NEXT:    vpsubw %ymm2, %ymm6, %ymm2
; AVX512F-NEXT:    vpsubw %ymm1, %ymm6, %ymm1
; AVX512F-NEXT:    vinserti64x4 $1, %ymm2, %zmm1, %zmm1
; AVX512F-NEXT:    vpternlogq $226, %zmm5, %zmm4, %zmm1
; AVX512F-NEXT:    vextracti64x4 $1, %zmm1, %ymm2
; AVX512F-NEXT:    vpaddw %ymm3, %ymm2, %ymm2
; AVX512F-NEXT:    vpaddw %ymm0, %ymm1, %ymm0
; AVX512F-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm0
; AVX512F-NEXT:    retq
;
; AVX512VL-FALLBACK-LABEL: vec512_i16_signed_reg_mem:
; AVX512VL-FALLBACK:       # %bb.0:
; AVX512VL-FALLBACK-NEXT:    vmovdqa (%rdi), %ymm1
; AVX512VL-FALLBACK-NEXT:    vmovdqa 32(%rdi), %ymm2
; AVX512VL-FALLBACK-NEXT:    vextracti64x4 $1, %zmm0, %ymm3
; AVX512VL-FALLBACK-NEXT:    vpcmpgtw %ymm2, %ymm3, %ymm4
; AVX512VL-FALLBACK-NEXT:    vpcmpgtw %ymm1, %ymm0, %ymm5
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm4, %zmm5, %zmm4
; AVX512VL-FALLBACK-NEXT:    vpminsw %ymm2, %ymm3, %ymm5
; AVX512VL-FALLBACK-NEXT:    vpmaxsw %ymm2, %ymm3, %ymm2
; AVX512VL-FALLBACK-NEXT:    vpsubw %ymm5, %ymm2, %ymm2
; AVX512VL-FALLBACK-NEXT:    vpminsw %ymm1, %ymm0, %ymm5
; AVX512VL-FALLBACK-NEXT:    vpmaxsw %ymm1, %ymm0, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpsubw %ymm5, %ymm1, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpsrlw $1, %ymm1, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpsrlw $1, %ymm2, %ymm2
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm2, %zmm1, %zmm5
; AVX512VL-FALLBACK-NEXT:    vpxor %xmm6, %xmm6, %xmm6
; AVX512VL-FALLBACK-NEXT:    vpsubw %ymm2, %ymm6, %ymm2
; AVX512VL-FALLBACK-NEXT:    vpsubw %ymm1, %ymm6, %ymm1
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm2, %zmm1, %zmm1
; AVX512VL-FALLBACK-NEXT:    vpternlogq $226, %zmm5, %zmm4, %zmm1
; AVX512VL-FALLBACK-NEXT:    vextracti64x4 $1, %zmm1, %ymm2
; AVX512VL-FALLBACK-NEXT:    vpaddw %ymm3, %ymm2, %ymm2
; AVX512VL-FALLBACK-NEXT:    vpaddw %ymm0, %ymm1, %ymm0
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm0
; AVX512VL-FALLBACK-NEXT:    retq
;
; AVX512BW-LABEL: vec512_i16_signed_reg_mem:
; AVX512BW:       # %bb.0:
; AVX512BW-NEXT:    vmovdqa64 (%rdi), %zmm1
; AVX512BW-NEXT:    vpcmpgtw %zmm1, %zmm0, %k1
; AVX512BW-NEXT:    vpminsw %zmm1, %zmm0, %zmm2
; AVX512BW-NEXT:    vpmaxsw %zmm1, %zmm0, %zmm1
; AVX512BW-NEXT:    vpsubw %zmm2, %zmm1, %zmm1
; AVX512BW-NEXT:    vpsrlw $1, %zmm1, %zmm1
; AVX512BW-NEXT:    vpxor %xmm2, %xmm2, %xmm2
; AVX512BW-NEXT:    vpsubw %zmm1, %zmm2, %zmm1 {%k1}
; AVX512BW-NEXT:    vpaddw %zmm0, %zmm1, %zmm0
; AVX512BW-NEXT:    retq
  %a2 = load <32 x i16>, ptr %a2_addr
  %t3 = icmp sgt <32 x i16> %a1, %a2 ; signed
  %t4 = select <32 x i1> %t3, <32 x i16> <i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1>, <32 x i16> <i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1>
  %t5 = select <32 x i1> %t3, <32 x i16> %a2, <32 x i16> %a1
  %t6 = select <32 x i1> %t3, <32 x i16> %a1, <32 x i16> %a2
  %t7 = sub <32 x i16> %t6, %t5
  %t16 = lshr <32 x i16> %t7, <i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1>
  %t9 = mul nsw <32 x i16> %t16, %t4 ; signed
  %a10 = add nsw <32 x i16> %t9, %a1 ; signed
  ret <32 x i16> %a10
}

define <32 x i16> @vec512_i16_signed_mem_mem(ptr %a1_addr, ptr %a2_addr) nounwind {
; AVX512F-LABEL: vec512_i16_signed_mem_mem:
; AVX512F:       # %bb.0:
; AVX512F-NEXT:    vmovdqa (%rsi), %ymm0
; AVX512F-NEXT:    vmovdqa 32(%rsi), %ymm1
; AVX512F-NEXT:    vmovdqa (%rdi), %ymm2
; AVX512F-NEXT:    vmovdqa 32(%rdi), %ymm3
; AVX512F-NEXT:    vpcmpgtw %ymm1, %ymm3, %ymm4
; AVX512F-NEXT:    vpcmpgtw %ymm0, %ymm2, %ymm5
; AVX512F-NEXT:    vinserti64x4 $1, %ymm4, %zmm5, %zmm4
; AVX512F-NEXT:    vpminsw %ymm1, %ymm3, %ymm5
; AVX512F-NEXT:    vpmaxsw %ymm1, %ymm3, %ymm1
; AVX512F-NEXT:    vpsubw %ymm5, %ymm1, %ymm1
; AVX512F-NEXT:    vpminsw %ymm0, %ymm2, %ymm5
; AVX512F-NEXT:    vpmaxsw %ymm0, %ymm2, %ymm0
; AVX512F-NEXT:    vpsubw %ymm5, %ymm0, %ymm0
; AVX512F-NEXT:    vpsrlw $1, %ymm0, %ymm0
; AVX512F-NEXT:    vpsrlw $1, %ymm1, %ymm1
; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm5
; AVX512F-NEXT:    vpxor %xmm6, %xmm6, %xmm6
; AVX512F-NEXT:    vpsubw %ymm1, %ymm6, %ymm1
; AVX512F-NEXT:    vpsubw %ymm0, %ymm6, %ymm0
; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512F-NEXT:    vpternlogq $226, %zmm5, %zmm4, %zmm0
; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
; AVX512F-NEXT:    vpaddw %ymm3, %ymm1, %ymm1
; AVX512F-NEXT:    vpaddw %ymm2, %ymm0, %ymm0
; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512F-NEXT:    retq
;
; AVX512VL-FALLBACK-LABEL: vec512_i16_signed_mem_mem:
; AVX512VL-FALLBACK:       # %bb.0:
; AVX512VL-FALLBACK-NEXT:    vmovdqa (%rsi), %ymm0
; AVX512VL-FALLBACK-NEXT:    vmovdqa 32(%rsi), %ymm1
; AVX512VL-FALLBACK-NEXT:    vmovdqa (%rdi), %ymm2
; AVX512VL-FALLBACK-NEXT:    vmovdqa 32(%rdi), %ymm3
; AVX512VL-FALLBACK-NEXT:    vpcmpgtw %ymm1, %ymm3, %ymm4
; AVX512VL-FALLBACK-NEXT:    vpcmpgtw %ymm0, %ymm2, %ymm5
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm4, %zmm5, %zmm4
; AVX512VL-FALLBACK-NEXT:    vpminsw %ymm1, %ymm3, %ymm5
; AVX512VL-FALLBACK-NEXT:    vpmaxsw %ymm1, %ymm3, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpsubw %ymm5, %ymm1, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpminsw %ymm0, %ymm2, %ymm5
; AVX512VL-FALLBACK-NEXT:    vpmaxsw %ymm0, %ymm2, %ymm0
; AVX512VL-FALLBACK-NEXT:    vpsubw %ymm5, %ymm0, %ymm0
; AVX512VL-FALLBACK-NEXT:    vpsrlw $1, %ymm0, %ymm0
; AVX512VL-FALLBACK-NEXT:    vpsrlw $1, %ymm1, %ymm1
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm5
; AVX512VL-FALLBACK-NEXT:    vpxor %xmm6, %xmm6, %xmm6
; AVX512VL-FALLBACK-NEXT:    vpsubw %ymm1, %ymm6, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpsubw %ymm0, %ymm6, %ymm0
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512VL-FALLBACK-NEXT:    vpternlogq $226, %zmm5, %zmm4, %zmm0
; AVX512VL-FALLBACK-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpaddw %ymm3, %ymm1, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpaddw %ymm2, %ymm0, %ymm0
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512VL-FALLBACK-NEXT:    retq
;
; AVX512BW-LABEL: vec512_i16_signed_mem_mem:
; AVX512BW:       # %bb.0:
; AVX512BW-NEXT:    vmovdqa64 (%rdi), %zmm0
; AVX512BW-NEXT:    vmovdqa64 (%rsi), %zmm1
; AVX512BW-NEXT:    vpcmpgtw %zmm1, %zmm0, %k1
; AVX512BW-NEXT:    vpminsw %zmm1, %zmm0, %zmm2
; AVX512BW-NEXT:    vpmaxsw %zmm1, %zmm0, %zmm1
; AVX512BW-NEXT:    vpsubw %zmm2, %zmm1, %zmm1
; AVX512BW-NEXT:    vpsrlw $1, %zmm1, %zmm1
; AVX512BW-NEXT:    vpxor %xmm2, %xmm2, %xmm2
; AVX512BW-NEXT:    vpsubw %zmm1, %zmm2, %zmm1 {%k1}
; AVX512BW-NEXT:    vpaddw %zmm0, %zmm1, %zmm0
; AVX512BW-NEXT:    retq
  %a1 = load <32 x i16>, ptr %a1_addr
  %a2 = load <32 x i16>, ptr %a2_addr
  %t3 = icmp sgt <32 x i16> %a1, %a2 ; signed
  %t4 = select <32 x i1> %t3, <32 x i16> <i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1>, <32 x i16> <i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1>
  %t5 = select <32 x i1> %t3, <32 x i16> %a2, <32 x i16> %a1
  %t6 = select <32 x i1> %t3, <32 x i16> %a1, <32 x i16> %a2
  %t7 = sub <32 x i16> %t6, %t5
  %t16 = lshr <32 x i16> %t7, <i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1>
  %t9 = mul nsw <32 x i16> %t16, %t4 ; signed
  %a10 = add nsw <32 x i16> %t9, %a1 ; signed
  ret <32 x i16> %a10
}

; ---------------------------------------------------------------------------- ;
; 8-bit width. 512 / 8 = 64 elts.
; ---------------------------------------------------------------------------- ;

; Values come from regs

define <64 x i8> @vec512_i8_signed_reg_reg(<64 x i8> %a1, <64 x i8> %a2) nounwind {
; AVX512F-LABEL: vec512_i8_signed_reg_reg:
; AVX512F:       # %bb.0:
; AVX512F-NEXT:    vextracti64x4 $1, %zmm1, %ymm2
; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm3
; AVX512F-NEXT:    vpcmpgtb %ymm2, %ymm3, %ymm4
; AVX512F-NEXT:    vpcmpgtb %ymm1, %ymm0, %ymm5
; AVX512F-NEXT:    vinserti64x4 $1, %ymm4, %zmm5, %zmm4
; AVX512F-NEXT:    vpminsb %ymm2, %ymm3, %ymm5
; AVX512F-NEXT:    vpmaxsb %ymm2, %ymm3, %ymm2
; AVX512F-NEXT:    vpsubb %ymm5, %ymm2, %ymm2
; AVX512F-NEXT:    vpminsb %ymm1, %ymm0, %ymm5
; AVX512F-NEXT:    vpmaxsb %ymm1, %ymm0, %ymm1
; AVX512F-NEXT:    vpsubb %ymm5, %ymm1, %ymm1
; AVX512F-NEXT:    vpsrlw $1, %ymm1, %ymm1
; AVX512F-NEXT:    vpsrlw $1, %ymm2, %ymm2
; AVX512F-NEXT:    vinserti64x4 $1, %ymm2, %zmm1, %zmm1
; AVX512F-NEXT:    vpandd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to16}, %zmm1, %zmm1
; AVX512F-NEXT:    vextracti64x4 $1, %zmm1, %ymm2
; AVX512F-NEXT:    vpxor %xmm5, %xmm5, %xmm5
; AVX512F-NEXT:    vpsubb %ymm2, %ymm5, %ymm2
; AVX512F-NEXT:    vpsubb %ymm1, %ymm5, %ymm5
; AVX512F-NEXT:    vinserti64x4 $1, %ymm2, %zmm5, %zmm2
; AVX512F-NEXT:    vpternlogq $226, %zmm1, %zmm4, %zmm2
; AVX512F-NEXT:    vextracti64x4 $1, %zmm2, %ymm1
; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
; AVX512F-NEXT:    vpaddb %ymm0, %ymm2, %ymm0
; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512F-NEXT:    retq
;
; AVX512VL-FALLBACK-LABEL: vec512_i8_signed_reg_reg:
; AVX512VL-FALLBACK:       # %bb.0:
; AVX512VL-FALLBACK-NEXT:    vextracti64x4 $1, %zmm1, %ymm2
; AVX512VL-FALLBACK-NEXT:    vextracti64x4 $1, %zmm0, %ymm3
; AVX512VL-FALLBACK-NEXT:    vpcmpgtb %ymm2, %ymm3, %ymm4
; AVX512VL-FALLBACK-NEXT:    vpcmpgtb %ymm1, %ymm0, %ymm5
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm4, %zmm5, %zmm4
; AVX512VL-FALLBACK-NEXT:    vpminsb %ymm2, %ymm3, %ymm5
; AVX512VL-FALLBACK-NEXT:    vpmaxsb %ymm2, %ymm3, %ymm2
; AVX512VL-FALLBACK-NEXT:    vpsubb %ymm5, %ymm2, %ymm2
; AVX512VL-FALLBACK-NEXT:    vpminsb %ymm1, %ymm0, %ymm5
; AVX512VL-FALLBACK-NEXT:    vpmaxsb %ymm1, %ymm0, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpsubb %ymm5, %ymm1, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpsrlw $1, %ymm1, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpsrlw $1, %ymm2, %ymm2
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm2, %zmm1, %zmm1
; AVX512VL-FALLBACK-NEXT:    vpandd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to16}, %zmm1, %zmm1
; AVX512VL-FALLBACK-NEXT:    vextracti64x4 $1, %zmm1, %ymm2
; AVX512VL-FALLBACK-NEXT:    vpxor %xmm5, %xmm5, %xmm5
; AVX512VL-FALLBACK-NEXT:    vpsubb %ymm2, %ymm5, %ymm2
; AVX512VL-FALLBACK-NEXT:    vpsubb %ymm1, %ymm5, %ymm5
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm2, %zmm5, %zmm2
; AVX512VL-FALLBACK-NEXT:    vpternlogq $226, %zmm1, %zmm4, %zmm2
; AVX512VL-FALLBACK-NEXT:    vextracti64x4 $1, %zmm2, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpaddb %ymm0, %ymm2, %ymm0
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512VL-FALLBACK-NEXT:    retq
;
; AVX512BW-LABEL: vec512_i8_signed_reg_reg:
; AVX512BW:       # %bb.0:
; AVX512BW-NEXT:    vpcmpgtb %zmm1, %zmm0, %k1
; AVX512BW-NEXT:    vpminsb %zmm1, %zmm0, %zmm2
; AVX512BW-NEXT:    vpmaxsb %zmm1, %zmm0, %zmm1
; AVX512BW-NEXT:    vpsubb %zmm2, %zmm1, %zmm1
; AVX512BW-NEXT:    vpsrlw $1, %zmm1, %zmm1
; AVX512BW-NEXT:    vpandd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to16}, %zmm1, %zmm1
; AVX512BW-NEXT:    vpxor %xmm2, %xmm2, %xmm2
; AVX512BW-NEXT:    vpsubb %zmm1, %zmm2, %zmm1 {%k1}
; AVX512BW-NEXT:    vpaddb %zmm0, %zmm1, %zmm0
; AVX512BW-NEXT:    retq
  %t3 = icmp sgt <64 x i8> %a1, %a2 ; signed
  %t4 = select <64 x i1> %t3, <64 x i8> <i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1>, <64 x i8> <i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1>
  %t5 = select <64 x i1> %t3, <64 x i8> %a2, <64 x i8> %a1
  %t6 = select <64 x i1> %t3, <64 x i8> %a1, <64 x i8> %a2
  %t7 = sub <64 x i8> %t6, %t5
  %t8 = lshr <64 x i8> %t7, <i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1>
  %t9 = mul nsw <64 x i8> %t8, %t4 ; signed
  %a10 = add nsw <64 x i8> %t9, %a1 ; signed
  ret <64 x i8> %a10
}

define <64 x i8> @vec512_i8_unsigned_reg_reg(<64 x i8> %a1, <64 x i8> %a2) nounwind {
; AVX512F-LABEL: vec512_i8_unsigned_reg_reg:
; AVX512F:       # %bb.0:
; AVX512F-NEXT:    vextracti64x4 $1, %zmm1, %ymm2
; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm3
; AVX512F-NEXT:    vpminub %ymm2, %ymm3, %ymm4
; AVX512F-NEXT:    vpcmpeqb %ymm4, %ymm3, %ymm5
; AVX512F-NEXT:    vpminub %ymm1, %ymm0, %ymm6
; AVX512F-NEXT:    vpcmpeqb %ymm6, %ymm0, %ymm7
; AVX512F-NEXT:    vinserti64x4 $1, %ymm5, %zmm7, %zmm5
; AVX512F-NEXT:    vpmaxub %ymm2, %ymm3, %ymm2
; AVX512F-NEXT:    vpsubb %ymm4, %ymm2, %ymm2
; AVX512F-NEXT:    vpmaxub %ymm1, %ymm0, %ymm1
; AVX512F-NEXT:    vpsubb %ymm6, %ymm1, %ymm1
; AVX512F-NEXT:    vpsrlw $1, %ymm1, %ymm1
; AVX512F-NEXT:    vpsrlw $1, %ymm2, %ymm2
; AVX512F-NEXT:    vinserti64x4 $1, %ymm2, %zmm1, %zmm1
; AVX512F-NEXT:    vpandd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to16}, %zmm1, %zmm1
; AVX512F-NEXT:    vextracti64x4 $1, %zmm1, %ymm2
; AVX512F-NEXT:    vpxor %xmm4, %xmm4, %xmm4
; AVX512F-NEXT:    vpsubb %ymm2, %ymm4, %ymm2
; AVX512F-NEXT:    vpsubb %ymm1, %ymm4, %ymm4
; AVX512F-NEXT:    vinserti64x4 $1, %ymm2, %zmm4, %zmm2
; AVX512F-NEXT:    vpternlogq $216, %zmm5, %zmm1, %zmm2
; AVX512F-NEXT:    vextracti64x4 $1, %zmm2, %ymm1
; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
; AVX512F-NEXT:    vpaddb %ymm0, %ymm2, %ymm0
; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512F-NEXT:    retq
;
; AVX512VL-FALLBACK-LABEL: vec512_i8_unsigned_reg_reg:
; AVX512VL-FALLBACK:       # %bb.0:
; AVX512VL-FALLBACK-NEXT:    vextracti64x4 $1, %zmm1, %ymm2
; AVX512VL-FALLBACK-NEXT:    vextracti64x4 $1, %zmm0, %ymm3
; AVX512VL-FALLBACK-NEXT:    vpminub %ymm2, %ymm3, %ymm4
; AVX512VL-FALLBACK-NEXT:    vpcmpeqb %ymm4, %ymm3, %ymm5
; AVX512VL-FALLBACK-NEXT:    vpminub %ymm1, %ymm0, %ymm6
; AVX512VL-FALLBACK-NEXT:    vpcmpeqb %ymm6, %ymm0, %ymm7
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm5, %zmm7, %zmm5
; AVX512VL-FALLBACK-NEXT:    vpmaxub %ymm2, %ymm3, %ymm2
; AVX512VL-FALLBACK-NEXT:    vpsubb %ymm4, %ymm2, %ymm2
; AVX512VL-FALLBACK-NEXT:    vpmaxub %ymm1, %ymm0, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpsubb %ymm6, %ymm1, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpsrlw $1, %ymm1, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpsrlw $1, %ymm2, %ymm2
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm2, %zmm1, %zmm1
; AVX512VL-FALLBACK-NEXT:    vpandd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to16}, %zmm1, %zmm1
; AVX512VL-FALLBACK-NEXT:    vextracti64x4 $1, %zmm1, %ymm2
; AVX512VL-FALLBACK-NEXT:    vpxor %xmm4, %xmm4, %xmm4
; AVX512VL-FALLBACK-NEXT:    vpsubb %ymm2, %ymm4, %ymm2
; AVX512VL-FALLBACK-NEXT:    vpsubb %ymm1, %ymm4, %ymm4
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm2, %zmm4, %zmm2
; AVX512VL-FALLBACK-NEXT:    vpternlogq $216, %zmm5, %zmm1, %zmm2
; AVX512VL-FALLBACK-NEXT:    vextracti64x4 $1, %zmm2, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpaddb %ymm0, %ymm2, %ymm0
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512VL-FALLBACK-NEXT:    retq
;
; AVX512BW-LABEL: vec512_i8_unsigned_reg_reg:
; AVX512BW:       # %bb.0:
; AVX512BW-NEXT:    vpcmpnleub %zmm1, %zmm0, %k1
; AVX512BW-NEXT:    vpminub %zmm1, %zmm0, %zmm2
; AVX512BW-NEXT:    vpmaxub %zmm1, %zmm0, %zmm1
; AVX512BW-NEXT:    vpsubb %zmm2, %zmm1, %zmm1
; AVX512BW-NEXT:    vpsrlw $1, %zmm1, %zmm1
; AVX512BW-NEXT:    vpandd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to16}, %zmm1, %zmm1
; AVX512BW-NEXT:    vpxor %xmm2, %xmm2, %xmm2
; AVX512BW-NEXT:    vpsubb %zmm1, %zmm2, %zmm1 {%k1}
; AVX512BW-NEXT:    vpaddb %zmm0, %zmm1, %zmm0
; AVX512BW-NEXT:    retq
  %t3 = icmp ugt <64 x i8> %a1, %a2
  %t4 = select <64 x i1> %t3, <64 x i8> <i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1>, <64 x i8> <i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1>
  %t5 = select <64 x i1> %t3, <64 x i8> %a2, <64 x i8> %a1
  %t6 = select <64 x i1> %t3, <64 x i8> %a1, <64 x i8> %a2
  %t7 = sub <64 x i8> %t6, %t5
  %t8 = lshr <64 x i8> %t7, <i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1>
  %t9 = mul <64 x i8> %t8, %t4
  %a10 = add <64 x i8> %t9, %a1
  ret <64 x i8> %a10
}

; Values are loaded. Only check signed case.

define <64 x i8> @vec512_i8_signed_mem_reg(ptr %a1_addr, <64 x i8> %a2) nounwind {
; AVX512F-LABEL: vec512_i8_signed_mem_reg:
; AVX512F:       # %bb.0:
; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
; AVX512F-NEXT:    vmovdqa (%rdi), %ymm2
; AVX512F-NEXT:    vmovdqa 32(%rdi), %ymm3
; AVX512F-NEXT:    vpcmpgtb %ymm1, %ymm3, %ymm4
; AVX512F-NEXT:    vpcmpgtb %ymm0, %ymm2, %ymm5
; AVX512F-NEXT:    vinserti64x4 $1, %ymm4, %zmm5, %zmm4
; AVX512F-NEXT:    vpminsb %ymm1, %ymm3, %ymm5
; AVX512F-NEXT:    vpmaxsb %ymm1, %ymm3, %ymm1
; AVX512F-NEXT:    vpsubb %ymm5, %ymm1, %ymm1
; AVX512F-NEXT:    vpminsb %ymm0, %ymm2, %ymm5
; AVX512F-NEXT:    vpmaxsb %ymm0, %ymm2, %ymm0
; AVX512F-NEXT:    vpsubb %ymm5, %ymm0, %ymm0
; AVX512F-NEXT:    vpsrlw $1, %ymm0, %ymm0
; AVX512F-NEXT:    vpsrlw $1, %ymm1, %ymm1
; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512F-NEXT:    vpandd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to16}, %zmm0, %zmm0
; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
; AVX512F-NEXT:    vpxor %xmm5, %xmm5, %xmm5
; AVX512F-NEXT:    vpsubb %ymm1, %ymm5, %ymm1
; AVX512F-NEXT:    vpsubb %ymm0, %ymm5, %ymm5
; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm5, %zmm1
; AVX512F-NEXT:    vpternlogq $226, %zmm0, %zmm4, %zmm1
; AVX512F-NEXT:    vextracti64x4 $1, %zmm1, %ymm0
; AVX512F-NEXT:    vpaddb %ymm3, %ymm0, %ymm0
; AVX512F-NEXT:    vpaddb %ymm2, %ymm1, %ymm1
; AVX512F-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512F-NEXT:    retq
;
; AVX512VL-FALLBACK-LABEL: vec512_i8_signed_mem_reg:
; AVX512VL-FALLBACK:       # %bb.0:
; AVX512VL-FALLBACK-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
; AVX512VL-FALLBACK-NEXT:    vmovdqa (%rdi), %ymm2
; AVX512VL-FALLBACK-NEXT:    vmovdqa 32(%rdi), %ymm3
; AVX512VL-FALLBACK-NEXT:    vpcmpgtb %ymm1, %ymm3, %ymm4
; AVX512VL-FALLBACK-NEXT:    vpcmpgtb %ymm0, %ymm2, %ymm5
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm4, %zmm5, %zmm4
; AVX512VL-FALLBACK-NEXT:    vpminsb %ymm1, %ymm3, %ymm5
; AVX512VL-FALLBACK-NEXT:    vpmaxsb %ymm1, %ymm3, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpsubb %ymm5, %ymm1, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpminsb %ymm0, %ymm2, %ymm5
; AVX512VL-FALLBACK-NEXT:    vpmaxsb %ymm0, %ymm2, %ymm0
; AVX512VL-FALLBACK-NEXT:    vpsubb %ymm5, %ymm0, %ymm0
; AVX512VL-FALLBACK-NEXT:    vpsrlw $1, %ymm0, %ymm0
; AVX512VL-FALLBACK-NEXT:    vpsrlw $1, %ymm1, %ymm1
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512VL-FALLBACK-NEXT:    vpandd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to16}, %zmm0, %zmm0
; AVX512VL-FALLBACK-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpxor %xmm5, %xmm5, %xmm5
; AVX512VL-FALLBACK-NEXT:    vpsubb %ymm1, %ymm5, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpsubb %ymm0, %ymm5, %ymm5
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm1, %zmm5, %zmm1
; AVX512VL-FALLBACK-NEXT:    vpternlogq $226, %zmm0, %zmm4, %zmm1
; AVX512VL-FALLBACK-NEXT:    vextracti64x4 $1, %zmm1, %ymm0
; AVX512VL-FALLBACK-NEXT:    vpaddb %ymm3, %ymm0, %ymm0
; AVX512VL-FALLBACK-NEXT:    vpaddb %ymm2, %ymm1, %ymm1
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512VL-FALLBACK-NEXT:    retq
;
; AVX512BW-LABEL: vec512_i8_signed_mem_reg:
; AVX512BW:       # %bb.0:
; AVX512BW-NEXT:    vmovdqa64 (%rdi), %zmm1
; AVX512BW-NEXT:    vpcmpgtb %zmm0, %zmm1, %k1
; AVX512BW-NEXT:    vpminsb %zmm0, %zmm1, %zmm2
; AVX512BW-NEXT:    vpmaxsb %zmm0, %zmm1, %zmm0
; AVX512BW-NEXT:    vpsubb %zmm2, %zmm0, %zmm0
; AVX512BW-NEXT:    vpsrlw $1, %zmm0, %zmm0
; AVX512BW-NEXT:    vpandd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to16}, %zmm0, %zmm0
; AVX512BW-NEXT:    vpxor %xmm2, %xmm2, %xmm2
; AVX512BW-NEXT:    vpsubb %zmm0, %zmm2, %zmm0 {%k1}
; AVX512BW-NEXT:    vpaddb %zmm1, %zmm0, %zmm0
; AVX512BW-NEXT:    retq
  %a1 = load <64 x i8>, ptr %a1_addr
  %t3 = icmp sgt <64 x i8> %a1, %a2 ; signed
  %t4 = select <64 x i1> %t3, <64 x i8> <i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1>, <64 x i8> <i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1>
  %t5 = select <64 x i1> %t3, <64 x i8> %a2, <64 x i8> %a1
  %t6 = select <64 x i1> %t3, <64 x i8> %a1, <64 x i8> %a2
  %t7 = sub <64 x i8> %t6, %t5
  %t8 = lshr <64 x i8> %t7, <i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1>
  %t9 = mul nsw <64 x i8> %t8, %t4 ; signed
  %a10 = add nsw <64 x i8> %t9, %a1 ; signed
  ret <64 x i8> %a10
}

define <64 x i8> @vec512_i8_signed_reg_mem(<64 x i8> %a1, ptr %a2_addr) nounwind {
; AVX512F-LABEL: vec512_i8_signed_reg_mem:
; AVX512F:       # %bb.0:
; AVX512F-NEXT:    vmovdqa (%rdi), %ymm1
; AVX512F-NEXT:    vmovdqa 32(%rdi), %ymm2
; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm3
; AVX512F-NEXT:    vpcmpgtb %ymm2, %ymm3, %ymm4
; AVX512F-NEXT:    vpcmpgtb %ymm1, %ymm0, %ymm5
; AVX512F-NEXT:    vinserti64x4 $1, %ymm4, %zmm5, %zmm4
; AVX512F-NEXT:    vpminsb %ymm2, %ymm3, %ymm5
; AVX512F-NEXT:    vpmaxsb %ymm2, %ymm3, %ymm2
; AVX512F-NEXT:    vpsubb %ymm5, %ymm2, %ymm2
; AVX512F-NEXT:    vpminsb %ymm1, %ymm0, %ymm5
; AVX512F-NEXT:    vpmaxsb %ymm1, %ymm0, %ymm1
; AVX512F-NEXT:    vpsubb %ymm5, %ymm1, %ymm1
; AVX512F-NEXT:    vpsrlw $1, %ymm1, %ymm1
; AVX512F-NEXT:    vpsrlw $1, %ymm2, %ymm2
; AVX512F-NEXT:    vinserti64x4 $1, %ymm2, %zmm1, %zmm1
; AVX512F-NEXT:    vpandd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to16}, %zmm1, %zmm1
; AVX512F-NEXT:    vextracti64x4 $1, %zmm1, %ymm2
; AVX512F-NEXT:    vpxor %xmm5, %xmm5, %xmm5
; AVX512F-NEXT:    vpsubb %ymm2, %ymm5, %ymm2
; AVX512F-NEXT:    vpsubb %ymm1, %ymm5, %ymm5
; AVX512F-NEXT:    vinserti64x4 $1, %ymm2, %zmm5, %zmm2
; AVX512F-NEXT:    vpternlogq $226, %zmm1, %zmm4, %zmm2
; AVX512F-NEXT:    vextracti64x4 $1, %zmm2, %ymm1
; AVX512F-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
; AVX512F-NEXT:    vpaddb %ymm0, %ymm2, %ymm0
; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512F-NEXT:    retq
;
; AVX512VL-FALLBACK-LABEL: vec512_i8_signed_reg_mem:
; AVX512VL-FALLBACK:       # %bb.0:
; AVX512VL-FALLBACK-NEXT:    vmovdqa (%rdi), %ymm1
; AVX512VL-FALLBACK-NEXT:    vmovdqa 32(%rdi), %ymm2
; AVX512VL-FALLBACK-NEXT:    vextracti64x4 $1, %zmm0, %ymm3
; AVX512VL-FALLBACK-NEXT:    vpcmpgtb %ymm2, %ymm3, %ymm4
; AVX512VL-FALLBACK-NEXT:    vpcmpgtb %ymm1, %ymm0, %ymm5
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm4, %zmm5, %zmm4
; AVX512VL-FALLBACK-NEXT:    vpminsb %ymm2, %ymm3, %ymm5
; AVX512VL-FALLBACK-NEXT:    vpmaxsb %ymm2, %ymm3, %ymm2
; AVX512VL-FALLBACK-NEXT:    vpsubb %ymm5, %ymm2, %ymm2
; AVX512VL-FALLBACK-NEXT:    vpminsb %ymm1, %ymm0, %ymm5
; AVX512VL-FALLBACK-NEXT:    vpmaxsb %ymm1, %ymm0, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpsubb %ymm5, %ymm1, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpsrlw $1, %ymm1, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpsrlw $1, %ymm2, %ymm2
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm2, %zmm1, %zmm1
; AVX512VL-FALLBACK-NEXT:    vpandd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to16}, %zmm1, %zmm1
; AVX512VL-FALLBACK-NEXT:    vextracti64x4 $1, %zmm1, %ymm2
; AVX512VL-FALLBACK-NEXT:    vpxor %xmm5, %xmm5, %xmm5
; AVX512VL-FALLBACK-NEXT:    vpsubb %ymm2, %ymm5, %ymm2
; AVX512VL-FALLBACK-NEXT:    vpsubb %ymm1, %ymm5, %ymm5
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm2, %zmm5, %zmm2
; AVX512VL-FALLBACK-NEXT:    vpternlogq $226, %zmm1, %zmm4, %zmm2
; AVX512VL-FALLBACK-NEXT:    vextracti64x4 $1, %zmm2, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpaddb %ymm3, %ymm1, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpaddb %ymm0, %ymm2, %ymm0
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512VL-FALLBACK-NEXT:    retq
;
; AVX512BW-LABEL: vec512_i8_signed_reg_mem:
; AVX512BW:       # %bb.0:
; AVX512BW-NEXT:    vmovdqa64 (%rdi), %zmm1
; AVX512BW-NEXT:    vpcmpgtb %zmm1, %zmm0, %k1
; AVX512BW-NEXT:    vpminsb %zmm1, %zmm0, %zmm2
; AVX512BW-NEXT:    vpmaxsb %zmm1, %zmm0, %zmm1
; AVX512BW-NEXT:    vpsubb %zmm2, %zmm1, %zmm1
; AVX512BW-NEXT:    vpsrlw $1, %zmm1, %zmm1
; AVX512BW-NEXT:    vpandd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to16}, %zmm1, %zmm1
; AVX512BW-NEXT:    vpxor %xmm2, %xmm2, %xmm2
; AVX512BW-NEXT:    vpsubb %zmm1, %zmm2, %zmm1 {%k1}
; AVX512BW-NEXT:    vpaddb %zmm0, %zmm1, %zmm0
; AVX512BW-NEXT:    retq
  %a2 = load <64 x i8>, ptr %a2_addr
  %t3 = icmp sgt <64 x i8> %a1, %a2 ; signed
  %t4 = select <64 x i1> %t3, <64 x i8> <i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1>, <64 x i8> <i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1>
  %t5 = select <64 x i1> %t3, <64 x i8> %a2, <64 x i8> %a1
  %t6 = select <64 x i1> %t3, <64 x i8> %a1, <64 x i8> %a2
  %t7 = sub <64 x i8> %t6, %t5
  %t8 = lshr <64 x i8> %t7, <i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1>
  %t9 = mul nsw <64 x i8> %t8, %t4 ; signed
  %a10 = add nsw <64 x i8> %t9, %a1 ; signed
  ret <64 x i8> %a10
}

define <64 x i8> @vec512_i8_signed_mem_mem(ptr %a1_addr, ptr %a2_addr) nounwind {
; AVX512F-LABEL: vec512_i8_signed_mem_mem:
; AVX512F:       # %bb.0:
; AVX512F-NEXT:    vmovdqa (%rsi), %ymm0
; AVX512F-NEXT:    vmovdqa 32(%rsi), %ymm1
; AVX512F-NEXT:    vmovdqa (%rdi), %ymm2
; AVX512F-NEXT:    vmovdqa 32(%rdi), %ymm3
; AVX512F-NEXT:    vpcmpgtb %ymm1, %ymm3, %ymm4
; AVX512F-NEXT:    vpcmpgtb %ymm0, %ymm2, %ymm5
; AVX512F-NEXT:    vinserti64x4 $1, %ymm4, %zmm5, %zmm4
; AVX512F-NEXT:    vpminsb %ymm1, %ymm3, %ymm5
; AVX512F-NEXT:    vpmaxsb %ymm1, %ymm3, %ymm1
; AVX512F-NEXT:    vpsubb %ymm5, %ymm1, %ymm1
; AVX512F-NEXT:    vpminsb %ymm0, %ymm2, %ymm5
; AVX512F-NEXT:    vpmaxsb %ymm0, %ymm2, %ymm0
; AVX512F-NEXT:    vpsubb %ymm5, %ymm0, %ymm0
; AVX512F-NEXT:    vpsrlw $1, %ymm0, %ymm0
; AVX512F-NEXT:    vpsrlw $1, %ymm1, %ymm1
; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512F-NEXT:    vpandd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to16}, %zmm0, %zmm0
; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
; AVX512F-NEXT:    vpxor %xmm5, %xmm5, %xmm5
; AVX512F-NEXT:    vpsubb %ymm1, %ymm5, %ymm1
; AVX512F-NEXT:    vpsubb %ymm0, %ymm5, %ymm5
; AVX512F-NEXT:    vinserti64x4 $1, %ymm1, %zmm5, %zmm1
; AVX512F-NEXT:    vpternlogq $226, %zmm0, %zmm4, %zmm1
; AVX512F-NEXT:    vextracti64x4 $1, %zmm1, %ymm0
; AVX512F-NEXT:    vpaddb %ymm3, %ymm0, %ymm0
; AVX512F-NEXT:    vpaddb %ymm2, %ymm1, %ymm1
; AVX512F-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512F-NEXT:    retq
;
; AVX512VL-FALLBACK-LABEL: vec512_i8_signed_mem_mem:
; AVX512VL-FALLBACK:       # %bb.0:
; AVX512VL-FALLBACK-NEXT:    vmovdqa (%rsi), %ymm0
; AVX512VL-FALLBACK-NEXT:    vmovdqa 32(%rsi), %ymm1
; AVX512VL-FALLBACK-NEXT:    vmovdqa (%rdi), %ymm2
; AVX512VL-FALLBACK-NEXT:    vmovdqa 32(%rdi), %ymm3
; AVX512VL-FALLBACK-NEXT:    vpcmpgtb %ymm1, %ymm3, %ymm4
; AVX512VL-FALLBACK-NEXT:    vpcmpgtb %ymm0, %ymm2, %ymm5
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm4, %zmm5, %zmm4
; AVX512VL-FALLBACK-NEXT:    vpminsb %ymm1, %ymm3, %ymm5
; AVX512VL-FALLBACK-NEXT:    vpmaxsb %ymm1, %ymm3, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpsubb %ymm5, %ymm1, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpminsb %ymm0, %ymm2, %ymm5
; AVX512VL-FALLBACK-NEXT:    vpmaxsb %ymm0, %ymm2, %ymm0
; AVX512VL-FALLBACK-NEXT:    vpsubb %ymm5, %ymm0, %ymm0
; AVX512VL-FALLBACK-NEXT:    vpsrlw $1, %ymm0, %ymm0
; AVX512VL-FALLBACK-NEXT:    vpsrlw $1, %ymm1, %ymm1
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512VL-FALLBACK-NEXT:    vpandd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to16}, %zmm0, %zmm0
; AVX512VL-FALLBACK-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpxor %xmm5, %xmm5, %xmm5
; AVX512VL-FALLBACK-NEXT:    vpsubb %ymm1, %ymm5, %ymm1
; AVX512VL-FALLBACK-NEXT:    vpsubb %ymm0, %ymm5, %ymm5
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm1, %zmm5, %zmm1
; AVX512VL-FALLBACK-NEXT:    vpternlogq $226, %zmm0, %zmm4, %zmm1
; AVX512VL-FALLBACK-NEXT:    vextracti64x4 $1, %zmm1, %ymm0
; AVX512VL-FALLBACK-NEXT:    vpaddb %ymm3, %ymm0, %ymm0
; AVX512VL-FALLBACK-NEXT:    vpaddb %ymm2, %ymm1, %ymm1
; AVX512VL-FALLBACK-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512VL-FALLBACK-NEXT:    retq
;
; AVX512BW-LABEL: vec512_i8_signed_mem_mem:
; AVX512BW:       # %bb.0:
; AVX512BW-NEXT:    vmovdqa64 (%rdi), %zmm0
; AVX512BW-NEXT:    vmovdqa64 (%rsi), %zmm1
; AVX512BW-NEXT:    vpcmpgtb %zmm1, %zmm0, %k1
; AVX512BW-NEXT:    vpminsb %zmm1, %zmm0, %zmm2
; AVX512BW-NEXT:    vpmaxsb %zmm1, %zmm0, %zmm1
; AVX512BW-NEXT:    vpsubb %zmm2, %zmm1, %zmm1
; AVX512BW-NEXT:    vpsrlw $1, %zmm1, %zmm1
; AVX512BW-NEXT:    vpandd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to16}, %zmm1, %zmm1
; AVX512BW-NEXT:    vpxor %xmm2, %xmm2, %xmm2
; AVX512BW-NEXT:    vpsubb %zmm1, %zmm2, %zmm1 {%k1}
; AVX512BW-NEXT:    vpaddb %zmm0, %zmm1, %zmm0
; AVX512BW-NEXT:    retq
  %a1 = load <64 x i8>, ptr %a1_addr
  %a2 = load <64 x i8>, ptr %a2_addr
  %t3 = icmp sgt <64 x i8> %a1, %a2 ; signed
  %t4 = select <64 x i1> %t3, <64 x i8> <i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1>, <64 x i8> <i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1>
  %t5 = select <64 x i1> %t3, <64 x i8> %a2, <64 x i8> %a1
  %t6 = select <64 x i1> %t3, <64 x i8> %a1, <64 x i8> %a2
  %t7 = sub <64 x i8> %t6, %t5
  %t8 = lshr <64 x i8> %t7, <i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1>
  %t9 = mul nsw <64 x i8> %t8, %t4 ; signed
  %a10 = add nsw <64 x i8> %t9, %a1 ; signed
  ret <64 x i8> %a10
}