[llvm] 4beb4d5 - [CostModel][X86] Add test coverage for icmp vs zero

Wed Feb 21 08:20:01 PST 2024

Author: Simon Pilgrim
Date: 2024-02-21T16:19:39Z
New Revision: 4beb4d5c72880ec69ef36bdebaed06c90cc5309b

URL: https://github.com/llvm/llvm-project/commit/4beb4d5c72880ec69ef36bdebaed06c90cc5309b
DIFF: https://github.com/llvm/llvm-project/commit/4beb4d5c72880ec69ef36bdebaed06c90cc5309b.diff

LOG: [CostModel][X86] Add test coverage for icmp vs zero

This is really to test for icmp vs constant - some icmp unsigned could fold to simpler comparisons, but costmodel analysis won't do this

Added: 
    llvm/test/Analysis/CostModel/X86/icmp0.ll

Modified: 
    

Removed: 
    


################################################################################
diff  --git a/llvm/test/Analysis/CostModel/X86/icmp0.ll b/llvm/test/Analysis/CostModel/X86/icmp0.ll
new file mode 100644
index 00000000000000..0d8a25207e781e

--- /dev/null
+++ b/llvm/test/Analysis/CostModel/X86/icmp0.ll
@@ -0,0 +1,3057 @@
+; NOTE: Assertions have been autogenerated by utils/update_analyze_test_checks.py
+; RUN: opt < %s -mtriple=x86_64-- -passes="print<cost-model>" 2>&1 -disable-output -mattr=+sse2 | FileCheck %s --check-prefixes=SSE2
+; RUN: opt < %s -mtriple=x86_64-- -passes="print<cost-model>" 2>&1 -disable-output -mattr=+sse3 | FileCheck %s --check-prefixes=SSE3
+; RUN: opt < %s -mtriple=x86_64-- -passes="print<cost-model>" 2>&1 -disable-output -mattr=+ssse3 | FileCheck %s --check-prefixes=SSSE3
+; RUN: opt < %s -mtriple=x86_64-- -passes="print<cost-model>" 2>&1 -disable-output -mattr=+sse4.1 | FileCheck %s --check-prefixes=SSE41
+; RUN: opt < %s -mtriple=x86_64-- -passes="print<cost-model>" 2>&1 -disable-output -mattr=+sse4.2 | FileCheck %s --check-prefixes=SSE42
+; RUN: opt < %s -mtriple=x86_64-- -passes="print<cost-model>" 2>&1 -disable-output -mattr=+avx | FileCheck %s --check-prefixes=AVX1
+; RUN: opt < %s -mtriple=x86_64-- -passes="print<cost-model>" 2>&1 -disable-output -mattr=+avx2 | FileCheck %s --check-prefixes=AVX2
+; RUN: opt < %s -mtriple=x86_64-- -passes="print<cost-model>" 2>&1 -disable-output -mattr=+avx512f | FileCheck %s --check-prefixes=AVX512F
+; RUN: opt < %s -mtriple=x86_64-- -passes="print<cost-model>" 2>&1 -disable-output -mattr=+avx512f,+avx512bw | FileCheck %s --check-prefixes=AVX512BW
+; RUN: opt < %s -mtriple=x86_64-- -passes="print<cost-model>" 2>&1 -disable-output -mattr=+xop,+avx | FileCheck %s -check-prefixes=XOPAVX1
+; RUN: opt < %s -mtriple=x86_64-- -passes="print<cost-model>" 2>&1 -disable-output -mattr=+xop,+avx2 | FileCheck %s -check-prefixes=XOPAVX2
+;
+; RUN: opt < %s -mtriple=x86_64-- -passes="print<cost-model>" 2>&1 -disable-output -mcpu=slm | FileCheck %s --check-prefixes=SLM
+; RUN: opt < %s -mtriple=x86_64-- -passes="print<cost-model>" 2>&1 -disable-output -mcpu=goldmont | FileCheck %s --check-prefixes=SSE42
+; RUN: opt < %s -mtriple=x86_64-- -passes="print<cost-model>" 2>&1 -disable-output -mcpu=btver2 | FileCheck %s --check-prefixes=AVX1
+
+define i32 @cmp_int_eq(i8 %arg8, <16 x i8> %argv16i8, <32 x i8> %argv32i8, <64 x i8> %argv64i8, <128 x i8> %argv128i8, i16 %arg16, <8 x i16> %argv8i16, <16 x i16> %argv16i16, <32 x i16> %argv32i16, <64 x i16> %argv64i16, i32 %arg32, <4 x i32> %argv4i32, <8 x i32> %argv8i32, <16 x i32> %argv16i32, <32 x i32> %argv32i32, i64 %arg64, <2 x i64> %argv2i64, <4 x i64> %argv4i64, <8 x i64> %argv8i64, <16 x i64> %argv16i64) {
+; SSE2-LABEL: 'cmp_int_eq'
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp eq i8 %arg8, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp eq <16 x i8> %argv16i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I8 = icmp eq <32 x i8> %argv32i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I8 = icmp eq <64 x i8> %argv64i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V128I8 = icmp eq <128 x i8> %argv128i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp eq i16 %arg16, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp eq <8 x i16> %argv8i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I16 = icmp eq <16 x i16> %argv16i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I16 = icmp eq <32 x i16> %argv32i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I16 = icmp eq <64 x i16> %argv64i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp eq i32 %arg32, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp eq <4 x i32> %argv4i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I32 = icmp eq <8 x i32> %argv8i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I32 = icmp eq <16 x i32> %argv16i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I32 = icmp eq <32 x i32> %argv32i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp eq i64 %arg64, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V2I64 = icmp eq <2 x i64> %argv2i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V4I64 = icmp eq <4 x i64> %argv4i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 20 for instruction: %V8I64 = icmp eq <8 x i64> %argv8i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 40 for instruction: %V16I64 = icmp eq <16 x i64> %argv16i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSE3-LABEL: 'cmp_int_eq'
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp eq i8 %arg8, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp eq <16 x i8> %argv16i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I8 = icmp eq <32 x i8> %argv32i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I8 = icmp eq <64 x i8> %argv64i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V128I8 = icmp eq <128 x i8> %argv128i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp eq i16 %arg16, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp eq <8 x i16> %argv8i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I16 = icmp eq <16 x i16> %argv16i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I16 = icmp eq <32 x i16> %argv32i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I16 = icmp eq <64 x i16> %argv64i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp eq i32 %arg32, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp eq <4 x i32> %argv4i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I32 = icmp eq <8 x i32> %argv8i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I32 = icmp eq <16 x i32> %argv16i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I32 = icmp eq <32 x i32> %argv32i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp eq i64 %arg64, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V2I64 = icmp eq <2 x i64> %argv2i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V4I64 = icmp eq <4 x i64> %argv4i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 20 for instruction: %V8I64 = icmp eq <8 x i64> %argv8i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 40 for instruction: %V16I64 = icmp eq <16 x i64> %argv16i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSSE3-LABEL: 'cmp_int_eq'
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp eq i8 %arg8, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp eq <16 x i8> %argv16i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I8 = icmp eq <32 x i8> %argv32i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I8 = icmp eq <64 x i8> %argv64i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V128I8 = icmp eq <128 x i8> %argv128i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp eq i16 %arg16, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp eq <8 x i16> %argv8i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I16 = icmp eq <16 x i16> %argv16i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I16 = icmp eq <32 x i16> %argv32i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I16 = icmp eq <64 x i16> %argv64i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp eq i32 %arg32, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp eq <4 x i32> %argv4i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I32 = icmp eq <8 x i32> %argv8i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I32 = icmp eq <16 x i32> %argv16i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I32 = icmp eq <32 x i32> %argv32i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp eq i64 %arg64, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V2I64 = icmp eq <2 x i64> %argv2i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V4I64 = icmp eq <4 x i64> %argv4i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 20 for instruction: %V8I64 = icmp eq <8 x i64> %argv8i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 40 for instruction: %V16I64 = icmp eq <16 x i64> %argv16i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSE41-LABEL: 'cmp_int_eq'
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp eq i8 %arg8, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp eq <16 x i8> %argv16i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I8 = icmp eq <32 x i8> %argv32i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I8 = icmp eq <64 x i8> %argv64i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V128I8 = icmp eq <128 x i8> %argv128i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp eq i16 %arg16, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp eq <8 x i16> %argv8i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I16 = icmp eq <16 x i16> %argv16i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I16 = icmp eq <32 x i16> %argv32i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I16 = icmp eq <64 x i16> %argv64i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp eq i32 %arg32, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp eq <4 x i32> %argv4i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I32 = icmp eq <8 x i32> %argv8i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I32 = icmp eq <16 x i32> %argv16i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I32 = icmp eq <32 x i32> %argv32i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp eq i64 %arg64, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V2I64 = icmp eq <2 x i64> %argv2i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V4I64 = icmp eq <4 x i64> %argv4i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 20 for instruction: %V8I64 = icmp eq <8 x i64> %argv8i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 40 for instruction: %V16I64 = icmp eq <16 x i64> %argv16i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSE42-LABEL: 'cmp_int_eq'
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp eq i8 %arg8, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp eq <16 x i8> %argv16i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I8 = icmp eq <32 x i8> %argv32i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I8 = icmp eq <64 x i8> %argv64i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V128I8 = icmp eq <128 x i8> %argv128i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp eq i16 %arg16, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp eq <8 x i16> %argv8i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I16 = icmp eq <16 x i16> %argv16i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I16 = icmp eq <32 x i16> %argv32i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I16 = icmp eq <64 x i16> %argv64i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp eq i32 %arg32, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp eq <4 x i32> %argv4i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I32 = icmp eq <8 x i32> %argv8i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I32 = icmp eq <16 x i32> %argv16i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I32 = icmp eq <32 x i32> %argv32i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp eq i64 %arg64, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp eq <2 x i64> %argv2i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I64 = icmp eq <4 x i64> %argv4i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I64 = icmp eq <8 x i64> %argv8i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I64 = icmp eq <16 x i64> %argv16i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX1-LABEL: 'cmp_int_eq'
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp eq i8 %arg8, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp eq <16 x i8> %argv16i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp eq <32 x i8> %argv32i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp eq <64 x i8> %argv64i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp eq <128 x i8> %argv128i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp eq i16 %arg16, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp eq <8 x i16> %argv8i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp eq <16 x i16> %argv16i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp eq <32 x i16> %argv32i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp eq <64 x i16> %argv64i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp eq i32 %arg32, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp eq <4 x i32> %argv4i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I32 = icmp eq <8 x i32> %argv8i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I32 = icmp eq <16 x i32> %argv16i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V32I32 = icmp eq <32 x i32> %argv32i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp eq i64 %arg64, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp eq <2 x i64> %argv2i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V4I64 = icmp eq <4 x i64> %argv4i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V8I64 = icmp eq <8 x i64> %argv8i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V16I64 = icmp eq <16 x i64> %argv16i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX2-LABEL: 'cmp_int_eq'
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp eq i8 %arg8, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp eq <16 x i8> %argv16i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V32I8 = icmp eq <32 x i8> %argv32i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V64I8 = icmp eq <64 x i8> %argv64i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V128I8 = icmp eq <128 x i8> %argv128i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp eq i16 %arg16, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp eq <8 x i16> %argv8i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I16 = icmp eq <16 x i16> %argv16i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I16 = icmp eq <32 x i16> %argv32i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I16 = icmp eq <64 x i16> %argv64i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp eq i32 %arg32, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp eq <4 x i32> %argv4i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I32 = icmp eq <8 x i32> %argv8i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I32 = icmp eq <16 x i32> %argv16i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I32 = icmp eq <32 x i32> %argv32i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp eq i64 %arg64, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp eq <2 x i64> %argv2i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I64 = icmp eq <4 x i64> %argv4i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I64 = icmp eq <8 x i64> %argv8i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I64 = icmp eq <16 x i64> %argv16i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX512F-LABEL: 'cmp_int_eq'
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp eq i8 %arg8, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp eq <16 x i8> %argv16i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V32I8 = icmp eq <32 x i8> %argv32i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V64I8 = icmp eq <64 x i8> %argv64i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V128I8 = icmp eq <128 x i8> %argv128i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp eq i16 %arg16, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp eq <8 x i16> %argv8i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I16 = icmp eq <16 x i16> %argv16i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V32I16 = icmp eq <32 x i16> %argv32i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V64I16 = icmp eq <64 x i16> %argv64i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp eq i32 %arg32, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp eq <4 x i32> %argv4i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I32 = icmp eq <8 x i32> %argv8i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I32 = icmp eq <16 x i32> %argv16i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I32 = icmp eq <32 x i32> %argv32i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp eq i64 %arg64, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp eq <2 x i64> %argv2i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I64 = icmp eq <4 x i64> %argv4i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I64 = icmp eq <8 x i64> %argv8i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I64 = icmp eq <16 x i64> %argv16i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX512BW-LABEL: 'cmp_int_eq'
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp eq i8 %arg8, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp eq <16 x i8> %argv16i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V32I8 = icmp eq <32 x i8> %argv32i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V64I8 = icmp eq <64 x i8> %argv64i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V128I8 = icmp eq <128 x i8> %argv128i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp eq i16 %arg16, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp eq <8 x i16> %argv8i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I16 = icmp eq <16 x i16> %argv16i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V32I16 = icmp eq <32 x i16> %argv32i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V64I16 = icmp eq <64 x i16> %argv64i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp eq i32 %arg32, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp eq <4 x i32> %argv4i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I32 = icmp eq <8 x i32> %argv8i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I32 = icmp eq <16 x i32> %argv16i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I32 = icmp eq <32 x i32> %argv32i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp eq i64 %arg64, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp eq <2 x i64> %argv2i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I64 = icmp eq <4 x i64> %argv4i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I64 = icmp eq <8 x i64> %argv8i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I64 = icmp eq <16 x i64> %argv16i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; XOPAVX1-LABEL: 'cmp_int_eq'
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp eq i8 %arg8, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp eq <16 x i8> %argv16i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp eq <32 x i8> %argv32i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp eq <64 x i8> %argv64i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp eq <128 x i8> %argv128i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp eq i16 %arg16, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp eq <8 x i16> %argv8i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp eq <16 x i16> %argv16i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp eq <32 x i16> %argv32i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp eq <64 x i16> %argv64i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp eq i32 %arg32, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp eq <4 x i32> %argv4i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I32 = icmp eq <8 x i32> %argv8i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I32 = icmp eq <16 x i32> %argv16i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V32I32 = icmp eq <32 x i32> %argv32i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp eq i64 %arg64, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp eq <2 x i64> %argv2i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V4I64 = icmp eq <4 x i64> %argv4i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V8I64 = icmp eq <8 x i64> %argv8i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V16I64 = icmp eq <16 x i64> %argv16i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; XOPAVX2-LABEL: 'cmp_int_eq'
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp eq i8 %arg8, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp eq <16 x i8> %argv16i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V32I8 = icmp eq <32 x i8> %argv32i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V64I8 = icmp eq <64 x i8> %argv64i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V128I8 = icmp eq <128 x i8> %argv128i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp eq i16 %arg16, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp eq <8 x i16> %argv8i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I16 = icmp eq <16 x i16> %argv16i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I16 = icmp eq <32 x i16> %argv32i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I16 = icmp eq <64 x i16> %argv64i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp eq i32 %arg32, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp eq <4 x i32> %argv4i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I32 = icmp eq <8 x i32> %argv8i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I32 = icmp eq <16 x i32> %argv16i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I32 = icmp eq <32 x i32> %argv32i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp eq i64 %arg64, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp eq <2 x i64> %argv2i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I64 = icmp eq <4 x i64> %argv4i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I64 = icmp eq <8 x i64> %argv8i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I64 = icmp eq <16 x i64> %argv16i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SLM-LABEL: 'cmp_int_eq'
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp eq i8 %arg8, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp eq <16 x i8> %argv16i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I8 = icmp eq <32 x i8> %argv32i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I8 = icmp eq <64 x i8> %argv64i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V128I8 = icmp eq <128 x i8> %argv128i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp eq i16 %arg16, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp eq <8 x i16> %argv8i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I16 = icmp eq <16 x i16> %argv16i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I16 = icmp eq <32 x i16> %argv32i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I16 = icmp eq <64 x i16> %argv64i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp eq i32 %arg32, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp eq <4 x i32> %argv4i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I32 = icmp eq <8 x i32> %argv8i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I32 = icmp eq <16 x i32> %argv16i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I32 = icmp eq <32 x i32> %argv32i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp eq i64 %arg64, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V2I64 = icmp eq <2 x i64> %argv2i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V4I64 = icmp eq <4 x i64> %argv4i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V8I64 = icmp eq <8 x i64> %argv8i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V16I64 = icmp eq <16 x i64> %argv16i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+  %I8 = icmp eq i8 %arg8, zeroinitializer
+  %V16I8 = icmp eq <16 x i8> %argv16i8, zeroinitializer
+  %V32I8 = icmp eq <32 x i8> %argv32i8, zeroinitializer
+  %V64I8 = icmp eq <64 x i8> %argv64i8, zeroinitializer
+  %V128I8 = icmp eq <128 x i8> %argv128i8, zeroinitializer
+
+  %I16 = icmp eq i16 %arg16, zeroinitializer
+  %V8I16  = icmp eq <8 x i16> %argv8i16, zeroinitializer
+  %V16I16 = icmp eq <16 x i16> %argv16i16, zeroinitializer
+  %V32I16 = icmp eq <32 x i16> %argv32i16, zeroinitializer
+  %V64I16 = icmp eq <64 x i16> %argv64i16, zeroinitializer
+
+  %I32 = icmp eq i32 %arg32, zeroinitializer
+  %V4I32 = icmp eq <4 x i32> %argv4i32, zeroinitializer
+  %V8I32 = icmp eq <8 x i32> %argv8i32, zeroinitializer
+  %V16I32 = icmp eq <16 x i32> %argv16i32, zeroinitializer
+  %V32I32 = icmp eq <32 x i32> %argv32i32, zeroinitializer
+
+  %I64 = icmp eq i64 %arg64, zeroinitializer
+  %V2I64 = icmp eq <2 x i64> %argv2i64, zeroinitializer
+  %V4I64 = icmp eq <4 x i64> %argv4i64, zeroinitializer
+  %V8I64 = icmp eq <8 x i64> %argv8i64, zeroinitializer
+  %V16I64 = icmp eq <16 x i64> %argv16i64, zeroinitializer
+
+  ret i32 undef
+}
+
+define i32 @cmp_int_ne(i8 %arg8, <16 x i8> %argv16i8, <32 x i8> %argv32i8, <64 x i8> %argv64i8, <128 x i8> %argv128i8, i16 %arg16, <8 x i16> %argv8i16, <16 x i16> %argv16i16, <32 x i16> %argv32i16, <64 x i16> %argv64i16, i32 %arg32, <4 x i32> %argv4i32, <8 x i32> %argv8i32, <16 x i32> %argv16i32, <32 x i32> %argv32i32, i64 %arg64, <2 x i64> %argv2i64, <4 x i64> %argv4i64, <8 x i64> %argv8i64, <16 x i64> %argv16i64) {
+; SSE2-LABEL: 'cmp_int_ne'
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ne i8 %arg8, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp ne <16 x i8> %argv16i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp ne <32 x i8> %argv32i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp ne <64 x i8> %argv64i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp ne <128 x i8> %argv128i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ne i16 %arg16, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp ne <8 x i16> %argv8i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp ne <16 x i16> %argv16i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp ne <32 x i16> %argv32i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp ne <64 x i16> %argv64i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ne i32 %arg32, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I32 = icmp ne <4 x i32> %argv4i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I32 = icmp ne <8 x i32> %argv8i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I32 = icmp ne <16 x i32> %argv16i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V32I32 = icmp ne <32 x i32> %argv32i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ne i64 %arg64, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V2I64 = icmp ne <2 x i64> %argv2i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V4I64 = icmp ne <4 x i64> %argv4i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V8I64 = icmp ne <8 x i64> %argv8i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 48 for instruction: %V16I64 = icmp ne <16 x i64> %argv16i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSE3-LABEL: 'cmp_int_ne'
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ne i8 %arg8, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp ne <16 x i8> %argv16i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp ne <32 x i8> %argv32i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp ne <64 x i8> %argv64i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp ne <128 x i8> %argv128i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ne i16 %arg16, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp ne <8 x i16> %argv8i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp ne <16 x i16> %argv16i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp ne <32 x i16> %argv32i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp ne <64 x i16> %argv64i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ne i32 %arg32, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I32 = icmp ne <4 x i32> %argv4i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I32 = icmp ne <8 x i32> %argv8i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I32 = icmp ne <16 x i32> %argv16i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V32I32 = icmp ne <32 x i32> %argv32i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ne i64 %arg64, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V2I64 = icmp ne <2 x i64> %argv2i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V4I64 = icmp ne <4 x i64> %argv4i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V8I64 = icmp ne <8 x i64> %argv8i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 48 for instruction: %V16I64 = icmp ne <16 x i64> %argv16i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSSE3-LABEL: 'cmp_int_ne'
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ne i8 %arg8, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp ne <16 x i8> %argv16i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp ne <32 x i8> %argv32i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp ne <64 x i8> %argv64i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp ne <128 x i8> %argv128i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ne i16 %arg16, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp ne <8 x i16> %argv8i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp ne <16 x i16> %argv16i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp ne <32 x i16> %argv32i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp ne <64 x i16> %argv64i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ne i32 %arg32, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I32 = icmp ne <4 x i32> %argv4i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I32 = icmp ne <8 x i32> %argv8i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I32 = icmp ne <16 x i32> %argv16i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V32I32 = icmp ne <32 x i32> %argv32i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ne i64 %arg64, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V2I64 = icmp ne <2 x i64> %argv2i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V4I64 = icmp ne <4 x i64> %argv4i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V8I64 = icmp ne <8 x i64> %argv8i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 48 for instruction: %V16I64 = icmp ne <16 x i64> %argv16i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSE41-LABEL: 'cmp_int_ne'
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ne i8 %arg8, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp ne <16 x i8> %argv16i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp ne <32 x i8> %argv32i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp ne <64 x i8> %argv64i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp ne <128 x i8> %argv128i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ne i16 %arg16, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp ne <8 x i16> %argv8i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp ne <16 x i16> %argv16i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp ne <32 x i16> %argv32i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp ne <64 x i16> %argv64i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ne i32 %arg32, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I32 = icmp ne <4 x i32> %argv4i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I32 = icmp ne <8 x i32> %argv8i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I32 = icmp ne <16 x i32> %argv16i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V32I32 = icmp ne <32 x i32> %argv32i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ne i64 %arg64, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V2I64 = icmp ne <2 x i64> %argv2i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V4I64 = icmp ne <4 x i64> %argv4i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V8I64 = icmp ne <8 x i64> %argv8i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 48 for instruction: %V16I64 = icmp ne <16 x i64> %argv16i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSE42-LABEL: 'cmp_int_ne'
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ne i8 %arg8, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp ne <16 x i8> %argv16i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp ne <32 x i8> %argv32i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp ne <64 x i8> %argv64i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp ne <128 x i8> %argv128i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ne i16 %arg16, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp ne <8 x i16> %argv8i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp ne <16 x i16> %argv16i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp ne <32 x i16> %argv32i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp ne <64 x i16> %argv64i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ne i32 %arg32, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I32 = icmp ne <4 x i32> %argv4i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I32 = icmp ne <8 x i32> %argv8i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I32 = icmp ne <16 x i32> %argv16i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V32I32 = icmp ne <32 x i32> %argv32i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ne i64 %arg64, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V2I64 = icmp ne <2 x i64> %argv2i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V4I64 = icmp ne <4 x i64> %argv4i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V8I64 = icmp ne <8 x i64> %argv8i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V16I64 = icmp ne <16 x i64> %argv16i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX1-LABEL: 'cmp_int_ne'
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ne i8 %arg8, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp ne <16 x i8> %argv16i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V32I8 = icmp ne <32 x i8> %argv32i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V64I8 = icmp ne <64 x i8> %argv64i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 20 for instruction: %V128I8 = icmp ne <128 x i8> %argv128i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ne i16 %arg16, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp ne <8 x i16> %argv8i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V16I16 = icmp ne <16 x i16> %argv16i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V32I16 = icmp ne <32 x i16> %argv32i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 20 for instruction: %V64I16 = icmp ne <64 x i16> %argv64i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ne i32 %arg32, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I32 = icmp ne <4 x i32> %argv4i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V8I32 = icmp ne <8 x i32> %argv8i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V16I32 = icmp ne <16 x i32> %argv16i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 20 for instruction: %V32I32 = icmp ne <32 x i32> %argv32i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ne i64 %arg64, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V2I64 = icmp ne <2 x i64> %argv2i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V4I64 = icmp ne <4 x i64> %argv4i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V8I64 = icmp ne <8 x i64> %argv8i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 20 for instruction: %V16I64 = icmp ne <16 x i64> %argv16i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX2-LABEL: 'cmp_int_ne'
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ne i8 %arg8, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp ne <16 x i8> %argv16i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I8 = icmp ne <32 x i8> %argv32i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I8 = icmp ne <64 x i8> %argv64i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V128I8 = icmp ne <128 x i8> %argv128i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ne i16 %arg16, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp ne <8 x i16> %argv8i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I16 = icmp ne <16 x i16> %argv16i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I16 = icmp ne <32 x i16> %argv32i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I16 = icmp ne <64 x i16> %argv64i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ne i32 %arg32, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I32 = icmp ne <4 x i32> %argv4i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I32 = icmp ne <8 x i32> %argv8i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I32 = icmp ne <16 x i32> %argv16i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I32 = icmp ne <32 x i32> %argv32i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ne i64 %arg64, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V2I64 = icmp ne <2 x i64> %argv2i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I64 = icmp ne <4 x i64> %argv4i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I64 = icmp ne <8 x i64> %argv8i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I64 = icmp ne <16 x i64> %argv16i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX512F-LABEL: 'cmp_int_ne'
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ne i8 %arg8, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp ne <16 x i8> %argv16i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I8 = icmp ne <32 x i8> %argv32i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I8 = icmp ne <64 x i8> %argv64i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V128I8 = icmp ne <128 x i8> %argv128i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ne i16 %arg16, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp ne <8 x i16> %argv8i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I16 = icmp ne <16 x i16> %argv16i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I16 = icmp ne <32 x i16> %argv32i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I16 = icmp ne <64 x i16> %argv64i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ne i32 %arg32, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp ne <4 x i32> %argv4i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I32 = icmp ne <8 x i32> %argv8i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I32 = icmp ne <16 x i32> %argv16i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I32 = icmp ne <32 x i32> %argv32i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ne i64 %arg64, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp ne <2 x i64> %argv2i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I64 = icmp ne <4 x i64> %argv4i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I64 = icmp ne <8 x i64> %argv8i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I64 = icmp ne <16 x i64> %argv16i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX512BW-LABEL: 'cmp_int_ne'
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ne i8 %arg8, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp ne <16 x i8> %argv16i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V32I8 = icmp ne <32 x i8> %argv32i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V64I8 = icmp ne <64 x i8> %argv64i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V128I8 = icmp ne <128 x i8> %argv128i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ne i16 %arg16, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp ne <8 x i16> %argv8i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I16 = icmp ne <16 x i16> %argv16i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V32I16 = icmp ne <32 x i16> %argv32i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V64I16 = icmp ne <64 x i16> %argv64i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ne i32 %arg32, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp ne <4 x i32> %argv4i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I32 = icmp ne <8 x i32> %argv8i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I32 = icmp ne <16 x i32> %argv16i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I32 = icmp ne <32 x i32> %argv32i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ne i64 %arg64, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp ne <2 x i64> %argv2i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I64 = icmp ne <4 x i64> %argv4i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I64 = icmp ne <8 x i64> %argv8i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I64 = icmp ne <16 x i64> %argv16i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; XOPAVX1-LABEL: 'cmp_int_ne'
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ne i8 %arg8, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp ne <16 x i8> %argv16i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp ne <32 x i8> %argv32i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp ne <64 x i8> %argv64i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp ne <128 x i8> %argv128i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ne i16 %arg16, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp ne <8 x i16> %argv8i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp ne <16 x i16> %argv16i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp ne <32 x i16> %argv32i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp ne <64 x i16> %argv64i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ne i32 %arg32, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp ne <4 x i32> %argv4i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I32 = icmp ne <8 x i32> %argv8i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I32 = icmp ne <16 x i32> %argv16i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V32I32 = icmp ne <32 x i32> %argv32i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ne i64 %arg64, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp ne <2 x i64> %argv2i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V4I64 = icmp ne <4 x i64> %argv4i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V8I64 = icmp ne <8 x i64> %argv8i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V16I64 = icmp ne <16 x i64> %argv16i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; XOPAVX2-LABEL: 'cmp_int_ne'
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ne i8 %arg8, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp ne <16 x i8> %argv16i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I8 = icmp ne <32 x i8> %argv32i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I8 = icmp ne <64 x i8> %argv64i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V128I8 = icmp ne <128 x i8> %argv128i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ne i16 %arg16, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp ne <8 x i16> %argv8i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I16 = icmp ne <16 x i16> %argv16i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I16 = icmp ne <32 x i16> %argv32i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I16 = icmp ne <64 x i16> %argv64i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ne i32 %arg32, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp ne <4 x i32> %argv4i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I32 = icmp ne <8 x i32> %argv8i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I32 = icmp ne <16 x i32> %argv16i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I32 = icmp ne <32 x i32> %argv32i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ne i64 %arg64, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp ne <2 x i64> %argv2i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I64 = icmp ne <4 x i64> %argv4i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I64 = icmp ne <8 x i64> %argv8i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I64 = icmp ne <16 x i64> %argv16i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SLM-LABEL: 'cmp_int_ne'
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ne i8 %arg8, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp ne <16 x i8> %argv16i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp ne <32 x i8> %argv32i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp ne <64 x i8> %argv64i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp ne <128 x i8> %argv128i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ne i16 %arg16, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp ne <8 x i16> %argv8i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp ne <16 x i16> %argv16i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp ne <32 x i16> %argv32i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp ne <64 x i16> %argv64i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ne i32 %arg32, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I32 = icmp ne <4 x i32> %argv4i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I32 = icmp ne <8 x i32> %argv8i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I32 = icmp ne <16 x i32> %argv16i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V32I32 = icmp ne <32 x i32> %argv32i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ne i64 %arg64, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V2I64 = icmp ne <2 x i64> %argv2i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V4I64 = icmp ne <4 x i64> %argv4i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V8I64 = icmp ne <8 x i64> %argv8i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V16I64 = icmp ne <16 x i64> %argv16i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+  %I8 = icmp ne i8 %arg8, zeroinitializer
+  %V16I8 = icmp ne <16 x i8> %argv16i8, zeroinitializer
+  %V32I8 = icmp ne <32 x i8> %argv32i8, zeroinitializer
+  %V64I8 = icmp ne <64 x i8> %argv64i8, zeroinitializer
+  %V128I8 = icmp ne <128 x i8> %argv128i8, zeroinitializer
+
+  %I16 = icmp ne i16 %arg16, zeroinitializer
+  %V8I16  = icmp ne <8 x i16> %argv8i16, zeroinitializer
+  %V16I16 = icmp ne <16 x i16> %argv16i16, zeroinitializer
+  %V32I16 = icmp ne <32 x i16> %argv32i16, zeroinitializer
+  %V64I16 = icmp ne <64 x i16> %argv64i16, zeroinitializer
+
+  %I32 = icmp ne i32 %arg32, zeroinitializer
+  %V4I32 = icmp ne <4 x i32> %argv4i32, zeroinitializer
+  %V8I32 = icmp ne <8 x i32> %argv8i32, zeroinitializer
+  %V16I32 = icmp ne <16 x i32> %argv16i32, zeroinitializer
+  %V32I32 = icmp ne <32 x i32> %argv32i32, zeroinitializer
+
+  %I64 = icmp ne i64 %arg64, zeroinitializer
+  %V2I64 = icmp ne <2 x i64> %argv2i64, zeroinitializer
+  %V4I64 = icmp ne <4 x i64> %argv4i64, zeroinitializer
+  %V8I64 = icmp ne <8 x i64> %argv8i64, zeroinitializer
+  %V16I64 = icmp ne <16 x i64> %argv16i64, zeroinitializer
+
+  ret i32 undef
+}
+
+define i32 @cmp_int_sge(i8 %arg8, <16 x i8> %argv16i8, <32 x i8> %argv32i8, <64 x i8> %argv64i8, <128 x i8> %argv128i8, i16 %arg16, <8 x i16> %argv8i16, <16 x i16> %argv16i16, <32 x i16> %argv32i16, <64 x i16> %argv64i16, i32 %arg32, <4 x i32> %argv4i32, <8 x i32> %argv8i32, <16 x i32> %argv16i32, <32 x i32> %argv32i32, i64 %arg64, <2 x i64> %argv2i64, <4 x i64> %argv4i64, <8 x i64> %argv8i64, <16 x i64> %argv16i64) {
+; SSE2-LABEL: 'cmp_int_sge'
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp sge i8 %arg8, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp sge <16 x i8> %argv16i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp sge <32 x i8> %argv32i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp sge <64 x i8> %argv64i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp sge <128 x i8> %argv128i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp sge i16 %arg16, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp sge <8 x i16> %argv8i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp sge <16 x i16> %argv16i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp sge <32 x i16> %argv32i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp sge <64 x i16> %argv64i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp sge i32 %arg32, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I32 = icmp sge <4 x i32> %argv4i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I32 = icmp sge <8 x i32> %argv8i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I32 = icmp sge <16 x i32> %argv16i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V32I32 = icmp sge <32 x i32> %argv32i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp sge i64 %arg64, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V2I64 = icmp sge <2 x i64> %argv2i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V4I64 = icmp sge <4 x i64> %argv4i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V8I64 = icmp sge <8 x i64> %argv8i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 48 for instruction: %V16I64 = icmp sge <16 x i64> %argv16i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSE3-LABEL: 'cmp_int_sge'
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp sge i8 %arg8, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp sge <16 x i8> %argv16i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp sge <32 x i8> %argv32i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp sge <64 x i8> %argv64i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp sge <128 x i8> %argv128i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp sge i16 %arg16, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp sge <8 x i16> %argv8i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp sge <16 x i16> %argv16i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp sge <32 x i16> %argv32i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp sge <64 x i16> %argv64i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp sge i32 %arg32, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I32 = icmp sge <4 x i32> %argv4i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I32 = icmp sge <8 x i32> %argv8i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I32 = icmp sge <16 x i32> %argv16i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V32I32 = icmp sge <32 x i32> %argv32i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp sge i64 %arg64, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V2I64 = icmp sge <2 x i64> %argv2i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V4I64 = icmp sge <4 x i64> %argv4i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V8I64 = icmp sge <8 x i64> %argv8i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 48 for instruction: %V16I64 = icmp sge <16 x i64> %argv16i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSSE3-LABEL: 'cmp_int_sge'
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp sge i8 %arg8, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp sge <16 x i8> %argv16i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp sge <32 x i8> %argv32i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp sge <64 x i8> %argv64i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp sge <128 x i8> %argv128i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp sge i16 %arg16, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp sge <8 x i16> %argv8i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp sge <16 x i16> %argv16i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp sge <32 x i16> %argv32i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp sge <64 x i16> %argv64i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp sge i32 %arg32, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I32 = icmp sge <4 x i32> %argv4i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I32 = icmp sge <8 x i32> %argv8i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I32 = icmp sge <16 x i32> %argv16i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V32I32 = icmp sge <32 x i32> %argv32i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp sge i64 %arg64, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V2I64 = icmp sge <2 x i64> %argv2i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V4I64 = icmp sge <4 x i64> %argv4i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V8I64 = icmp sge <8 x i64> %argv8i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 48 for instruction: %V16I64 = icmp sge <16 x i64> %argv16i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSE41-LABEL: 'cmp_int_sge'
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp sge i8 %arg8, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp sge <16 x i8> %argv16i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp sge <32 x i8> %argv32i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp sge <64 x i8> %argv64i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp sge <128 x i8> %argv128i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp sge i16 %arg16, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp sge <8 x i16> %argv8i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp sge <16 x i16> %argv16i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp sge <32 x i16> %argv32i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp sge <64 x i16> %argv64i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp sge i32 %arg32, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I32 = icmp sge <4 x i32> %argv4i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I32 = icmp sge <8 x i32> %argv8i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I32 = icmp sge <16 x i32> %argv16i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V32I32 = icmp sge <32 x i32> %argv32i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp sge i64 %arg64, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V2I64 = icmp sge <2 x i64> %argv2i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V4I64 = icmp sge <4 x i64> %argv4i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V8I64 = icmp sge <8 x i64> %argv8i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 48 for instruction: %V16I64 = icmp sge <16 x i64> %argv16i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSE42-LABEL: 'cmp_int_sge'
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp sge i8 %arg8, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp sge <16 x i8> %argv16i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp sge <32 x i8> %argv32i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp sge <64 x i8> %argv64i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp sge <128 x i8> %argv128i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp sge i16 %arg16, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp sge <8 x i16> %argv8i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp sge <16 x i16> %argv16i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp sge <32 x i16> %argv32i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp sge <64 x i16> %argv64i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp sge i32 %arg32, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I32 = icmp sge <4 x i32> %argv4i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I32 = icmp sge <8 x i32> %argv8i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I32 = icmp sge <16 x i32> %argv16i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V32I32 = icmp sge <32 x i32> %argv32i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp sge i64 %arg64, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V2I64 = icmp sge <2 x i64> %argv2i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V4I64 = icmp sge <4 x i64> %argv4i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V8I64 = icmp sge <8 x i64> %argv8i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V16I64 = icmp sge <16 x i64> %argv16i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX1-LABEL: 'cmp_int_sge'
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp sge i8 %arg8, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp sge <16 x i8> %argv16i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V32I8 = icmp sge <32 x i8> %argv32i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V64I8 = icmp sge <64 x i8> %argv64i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 20 for instruction: %V128I8 = icmp sge <128 x i8> %argv128i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp sge i16 %arg16, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp sge <8 x i16> %argv8i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V16I16 = icmp sge <16 x i16> %argv16i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V32I16 = icmp sge <32 x i16> %argv32i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 20 for instruction: %V64I16 = icmp sge <64 x i16> %argv64i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp sge i32 %arg32, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I32 = icmp sge <4 x i32> %argv4i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V8I32 = icmp sge <8 x i32> %argv8i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V16I32 = icmp sge <16 x i32> %argv16i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 20 for instruction: %V32I32 = icmp sge <32 x i32> %argv32i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp sge i64 %arg64, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V2I64 = icmp sge <2 x i64> %argv2i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V4I64 = icmp sge <4 x i64> %argv4i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V8I64 = icmp sge <8 x i64> %argv8i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 20 for instruction: %V16I64 = icmp sge <16 x i64> %argv16i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX2-LABEL: 'cmp_int_sge'
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp sge i8 %arg8, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp sge <16 x i8> %argv16i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I8 = icmp sge <32 x i8> %argv32i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I8 = icmp sge <64 x i8> %argv64i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V128I8 = icmp sge <128 x i8> %argv128i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp sge i16 %arg16, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp sge <8 x i16> %argv8i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I16 = icmp sge <16 x i16> %argv16i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I16 = icmp sge <32 x i16> %argv32i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I16 = icmp sge <64 x i16> %argv64i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp sge i32 %arg32, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I32 = icmp sge <4 x i32> %argv4i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I32 = icmp sge <8 x i32> %argv8i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I32 = icmp sge <16 x i32> %argv16i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I32 = icmp sge <32 x i32> %argv32i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp sge i64 %arg64, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V2I64 = icmp sge <2 x i64> %argv2i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I64 = icmp sge <4 x i64> %argv4i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I64 = icmp sge <8 x i64> %argv8i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I64 = icmp sge <16 x i64> %argv16i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX512F-LABEL: 'cmp_int_sge'
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp sge i8 %arg8, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp sge <16 x i8> %argv16i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I8 = icmp sge <32 x i8> %argv32i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I8 = icmp sge <64 x i8> %argv64i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V128I8 = icmp sge <128 x i8> %argv128i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp sge i16 %arg16, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp sge <8 x i16> %argv8i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I16 = icmp sge <16 x i16> %argv16i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I16 = icmp sge <32 x i16> %argv32i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I16 = icmp sge <64 x i16> %argv64i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp sge i32 %arg32, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp sge <4 x i32> %argv4i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I32 = icmp sge <8 x i32> %argv8i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I32 = icmp sge <16 x i32> %argv16i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I32 = icmp sge <32 x i32> %argv32i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp sge i64 %arg64, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp sge <2 x i64> %argv2i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I64 = icmp sge <4 x i64> %argv4i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I64 = icmp sge <8 x i64> %argv8i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I64 = icmp sge <16 x i64> %argv16i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX512BW-LABEL: 'cmp_int_sge'
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp sge i8 %arg8, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp sge <16 x i8> %argv16i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V32I8 = icmp sge <32 x i8> %argv32i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V64I8 = icmp sge <64 x i8> %argv64i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V128I8 = icmp sge <128 x i8> %argv128i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp sge i16 %arg16, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp sge <8 x i16> %argv8i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I16 = icmp sge <16 x i16> %argv16i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V32I16 = icmp sge <32 x i16> %argv32i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V64I16 = icmp sge <64 x i16> %argv64i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp sge i32 %arg32, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp sge <4 x i32> %argv4i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I32 = icmp sge <8 x i32> %argv8i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I32 = icmp sge <16 x i32> %argv16i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I32 = icmp sge <32 x i32> %argv32i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp sge i64 %arg64, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp sge <2 x i64> %argv2i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I64 = icmp sge <4 x i64> %argv4i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I64 = icmp sge <8 x i64> %argv8i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I64 = icmp sge <16 x i64> %argv16i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; XOPAVX1-LABEL: 'cmp_int_sge'
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp sge i8 %arg8, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp sge <16 x i8> %argv16i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp sge <32 x i8> %argv32i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp sge <64 x i8> %argv64i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp sge <128 x i8> %argv128i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp sge i16 %arg16, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp sge <8 x i16> %argv8i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp sge <16 x i16> %argv16i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp sge <32 x i16> %argv32i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp sge <64 x i16> %argv64i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp sge i32 %arg32, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp sge <4 x i32> %argv4i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I32 = icmp sge <8 x i32> %argv8i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I32 = icmp sge <16 x i32> %argv16i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V32I32 = icmp sge <32 x i32> %argv32i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp sge i64 %arg64, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp sge <2 x i64> %argv2i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V4I64 = icmp sge <4 x i64> %argv4i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V8I64 = icmp sge <8 x i64> %argv8i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V16I64 = icmp sge <16 x i64> %argv16i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; XOPAVX2-LABEL: 'cmp_int_sge'
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp sge i8 %arg8, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp sge <16 x i8> %argv16i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I8 = icmp sge <32 x i8> %argv32i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I8 = icmp sge <64 x i8> %argv64i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V128I8 = icmp sge <128 x i8> %argv128i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp sge i16 %arg16, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp sge <8 x i16> %argv8i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I16 = icmp sge <16 x i16> %argv16i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I16 = icmp sge <32 x i16> %argv32i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I16 = icmp sge <64 x i16> %argv64i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp sge i32 %arg32, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp sge <4 x i32> %argv4i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I32 = icmp sge <8 x i32> %argv8i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I32 = icmp sge <16 x i32> %argv16i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I32 = icmp sge <32 x i32> %argv32i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp sge i64 %arg64, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp sge <2 x i64> %argv2i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I64 = icmp sge <4 x i64> %argv4i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I64 = icmp sge <8 x i64> %argv8i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I64 = icmp sge <16 x i64> %argv16i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SLM-LABEL: 'cmp_int_sge'
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp sge i8 %arg8, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp sge <16 x i8> %argv16i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp sge <32 x i8> %argv32i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp sge <64 x i8> %argv64i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp sge <128 x i8> %argv128i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp sge i16 %arg16, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp sge <8 x i16> %argv8i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp sge <16 x i16> %argv16i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp sge <32 x i16> %argv32i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp sge <64 x i16> %argv64i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp sge i32 %arg32, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I32 = icmp sge <4 x i32> %argv4i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I32 = icmp sge <8 x i32> %argv8i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I32 = icmp sge <16 x i32> %argv16i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V32I32 = icmp sge <32 x i32> %argv32i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp sge i64 %arg64, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V2I64 = icmp sge <2 x i64> %argv2i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V4I64 = icmp sge <4 x i64> %argv4i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V8I64 = icmp sge <8 x i64> %argv8i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V16I64 = icmp sge <16 x i64> %argv16i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+  %I8 = icmp sge i8 %arg8, zeroinitializer
+  %V16I8 = icmp sge <16 x i8> %argv16i8, zeroinitializer
+  %V32I8 = icmp sge <32 x i8> %argv32i8, zeroinitializer
+  %V64I8 = icmp sge <64 x i8> %argv64i8, zeroinitializer
+  %V128I8 = icmp sge <128 x i8> %argv128i8, zeroinitializer
+
+  %I16 = icmp sge i16 %arg16, zeroinitializer
+  %V8I16  = icmp sge <8 x i16> %argv8i16, zeroinitializer
+  %V16I16 = icmp sge <16 x i16> %argv16i16, zeroinitializer
+  %V32I16 = icmp sge <32 x i16> %argv32i16, zeroinitializer
+  %V64I16 = icmp sge <64 x i16> %argv64i16, zeroinitializer
+
+  %I32 = icmp sge i32 %arg32, zeroinitializer
+  %V4I32 = icmp sge <4 x i32> %argv4i32, zeroinitializer
+  %V8I32 = icmp sge <8 x i32> %argv8i32, zeroinitializer
+  %V16I32 = icmp sge <16 x i32> %argv16i32, zeroinitializer
+  %V32I32 = icmp sge <32 x i32> %argv32i32, zeroinitializer
+
+  %I64 = icmp sge i64 %arg64, zeroinitializer
+  %V2I64 = icmp sge <2 x i64> %argv2i64, zeroinitializer
+  %V4I64 = icmp sge <4 x i64> %argv4i64, zeroinitializer
+  %V8I64 = icmp sge <8 x i64> %argv8i64, zeroinitializer
+  %V16I64 = icmp sge <16 x i64> %argv16i64, zeroinitializer
+
+  ret i32 undef
+}
+
+define i32 @cmp_int_uge(i8 %arg8, <16 x i8> %argv16i8, <32 x i8> %argv32i8, <64 x i8> %argv64i8, <128 x i8> %argv128i8, i16 %arg16, <8 x i16> %argv8i16, <16 x i16> %argv16i16, <32 x i16> %argv32i16, <64 x i16> %argv64i16, i32 %arg32, <4 x i32> %argv4i32, <8 x i32> %argv8i32, <16 x i32> %argv16i32, <32 x i32> %argv32i32, i64 %arg64, <2 x i64> %argv2i64, <4 x i64> %argv4i64, <8 x i64> %argv8i64, <16 x i64> %argv16i64) {
+; SSE2-LABEL: 'cmp_int_uge'
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp uge i8 %arg8, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp uge <16 x i8> %argv16i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp uge <32 x i8> %argv32i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp uge <64 x i8> %argv64i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp uge <128 x i8> %argv128i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp uge i16 %arg16, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp uge <8 x i16> %argv8i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp uge <16 x i16> %argv16i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp uge <32 x i16> %argv32i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp uge <64 x i16> %argv64i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp uge i32 %arg32, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V4I32 = icmp uge <4 x i32> %argv4i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V8I32 = icmp uge <8 x i32> %argv8i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V16I32 = icmp uge <16 x i32> %argv16i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 32 for instruction: %V32I32 = icmp uge <32 x i32> %argv32i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp uge i64 %arg64, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V2I64 = icmp uge <2 x i64> %argv2i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V4I64 = icmp uge <4 x i64> %argv4i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 32 for instruction: %V8I64 = icmp uge <8 x i64> %argv8i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 64 for instruction: %V16I64 = icmp uge <16 x i64> %argv16i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSE3-LABEL: 'cmp_int_uge'
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp uge i8 %arg8, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp uge <16 x i8> %argv16i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp uge <32 x i8> %argv32i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp uge <64 x i8> %argv64i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp uge <128 x i8> %argv128i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp uge i16 %arg16, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp uge <8 x i16> %argv8i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp uge <16 x i16> %argv16i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp uge <32 x i16> %argv32i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp uge <64 x i16> %argv64i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp uge i32 %arg32, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V4I32 = icmp uge <4 x i32> %argv4i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V8I32 = icmp uge <8 x i32> %argv8i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V16I32 = icmp uge <16 x i32> %argv16i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 32 for instruction: %V32I32 = icmp uge <32 x i32> %argv32i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp uge i64 %arg64, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V2I64 = icmp uge <2 x i64> %argv2i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V4I64 = icmp uge <4 x i64> %argv4i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 32 for instruction: %V8I64 = icmp uge <8 x i64> %argv8i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 64 for instruction: %V16I64 = icmp uge <16 x i64> %argv16i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSSE3-LABEL: 'cmp_int_uge'
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp uge i8 %arg8, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp uge <16 x i8> %argv16i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp uge <32 x i8> %argv32i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp uge <64 x i8> %argv64i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp uge <128 x i8> %argv128i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp uge i16 %arg16, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp uge <8 x i16> %argv8i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp uge <16 x i16> %argv16i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp uge <32 x i16> %argv32i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp uge <64 x i16> %argv64i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp uge i32 %arg32, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V4I32 = icmp uge <4 x i32> %argv4i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V8I32 = icmp uge <8 x i32> %argv8i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V16I32 = icmp uge <16 x i32> %argv16i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 32 for instruction: %V32I32 = icmp uge <32 x i32> %argv32i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp uge i64 %arg64, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V2I64 = icmp uge <2 x i64> %argv2i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V4I64 = icmp uge <4 x i64> %argv4i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 32 for instruction: %V8I64 = icmp uge <8 x i64> %argv8i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 64 for instruction: %V16I64 = icmp uge <16 x i64> %argv16i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSE41-LABEL: 'cmp_int_uge'
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp uge i8 %arg8, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp uge <16 x i8> %argv16i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp uge <32 x i8> %argv32i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp uge <64 x i8> %argv64i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp uge <128 x i8> %argv128i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp uge i16 %arg16, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp uge <8 x i16> %argv8i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp uge <16 x i16> %argv16i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp uge <32 x i16> %argv32i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp uge <64 x i16> %argv64i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp uge i32 %arg32, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I32 = icmp uge <4 x i32> %argv4i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I32 = icmp uge <8 x i32> %argv8i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I32 = icmp uge <16 x i32> %argv16i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V32I32 = icmp uge <32 x i32> %argv32i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp uge i64 %arg64, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V2I64 = icmp uge <2 x i64> %argv2i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V4I64 = icmp uge <4 x i64> %argv4i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 32 for instruction: %V8I64 = icmp uge <8 x i64> %argv8i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 64 for instruction: %V16I64 = icmp uge <16 x i64> %argv16i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSE42-LABEL: 'cmp_int_uge'
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp uge i8 %arg8, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp uge <16 x i8> %argv16i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp uge <32 x i8> %argv32i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp uge <64 x i8> %argv64i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp uge <128 x i8> %argv128i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp uge i16 %arg16, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp uge <8 x i16> %argv8i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp uge <16 x i16> %argv16i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp uge <32 x i16> %argv32i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp uge <64 x i16> %argv64i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp uge i32 %arg32, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I32 = icmp uge <4 x i32> %argv4i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I32 = icmp uge <8 x i32> %argv8i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I32 = icmp uge <16 x i32> %argv16i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V32I32 = icmp uge <32 x i32> %argv32i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp uge i64 %arg64, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V2I64 = icmp uge <2 x i64> %argv2i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V4I64 = icmp uge <4 x i64> %argv4i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V8I64 = icmp uge <8 x i64> %argv8i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 32 for instruction: %V16I64 = icmp uge <16 x i64> %argv16i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX1-LABEL: 'cmp_int_uge'
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp uge i8 %arg8, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp uge <16 x i8> %argv16i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V32I8 = icmp uge <32 x i8> %argv32i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V64I8 = icmp uge <64 x i8> %argv64i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 20 for instruction: %V128I8 = icmp uge <128 x i8> %argv128i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp uge i16 %arg16, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp uge <8 x i16> %argv8i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V16I16 = icmp uge <16 x i16> %argv16i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V32I16 = icmp uge <32 x i16> %argv32i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 20 for instruction: %V64I16 = icmp uge <64 x i16> %argv64i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp uge i32 %arg32, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I32 = icmp uge <4 x i32> %argv4i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V8I32 = icmp uge <8 x i32> %argv8i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V16I32 = icmp uge <16 x i32> %argv16i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 20 for instruction: %V32I32 = icmp uge <32 x i32> %argv32i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp uge i64 %arg64, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V2I64 = icmp uge <2 x i64> %argv2i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 7 for instruction: %V4I64 = icmp uge <4 x i64> %argv4i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 14 for instruction: %V8I64 = icmp uge <8 x i64> %argv8i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 28 for instruction: %V16I64 = icmp uge <16 x i64> %argv16i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX2-LABEL: 'cmp_int_uge'
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp uge i8 %arg8, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp uge <16 x i8> %argv16i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I8 = icmp uge <32 x i8> %argv32i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I8 = icmp uge <64 x i8> %argv64i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V128I8 = icmp uge <128 x i8> %argv128i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp uge i16 %arg16, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp uge <8 x i16> %argv8i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I16 = icmp uge <16 x i16> %argv16i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I16 = icmp uge <32 x i16> %argv32i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I16 = icmp uge <64 x i16> %argv64i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp uge i32 %arg32, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I32 = icmp uge <4 x i32> %argv4i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I32 = icmp uge <8 x i32> %argv8i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I32 = icmp uge <16 x i32> %argv16i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I32 = icmp uge <32 x i32> %argv32i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp uge i64 %arg64, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V2I64 = icmp uge <2 x i64> %argv2i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V4I64 = icmp uge <4 x i64> %argv4i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V8I64 = icmp uge <8 x i64> %argv8i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V16I64 = icmp uge <16 x i64> %argv16i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX512F-LABEL: 'cmp_int_uge'
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp uge i8 %arg8, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp uge <16 x i8> %argv16i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I8 = icmp uge <32 x i8> %argv32i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I8 = icmp uge <64 x i8> %argv64i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V128I8 = icmp uge <128 x i8> %argv128i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp uge i16 %arg16, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp uge <8 x i16> %argv8i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I16 = icmp uge <16 x i16> %argv16i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I16 = icmp uge <32 x i16> %argv32i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I16 = icmp uge <64 x i16> %argv64i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp uge i32 %arg32, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp uge <4 x i32> %argv4i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I32 = icmp uge <8 x i32> %argv8i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I32 = icmp uge <16 x i32> %argv16i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I32 = icmp uge <32 x i32> %argv32i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp uge i64 %arg64, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp uge <2 x i64> %argv2i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I64 = icmp uge <4 x i64> %argv4i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I64 = icmp uge <8 x i64> %argv8i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I64 = icmp uge <16 x i64> %argv16i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX512BW-LABEL: 'cmp_int_uge'
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp uge i8 %arg8, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp uge <16 x i8> %argv16i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V32I8 = icmp uge <32 x i8> %argv32i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V64I8 = icmp uge <64 x i8> %argv64i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V128I8 = icmp uge <128 x i8> %argv128i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp uge i16 %arg16, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp uge <8 x i16> %argv8i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I16 = icmp uge <16 x i16> %argv16i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V32I16 = icmp uge <32 x i16> %argv32i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V64I16 = icmp uge <64 x i16> %argv64i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp uge i32 %arg32, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp uge <4 x i32> %argv4i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I32 = icmp uge <8 x i32> %argv8i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I32 = icmp uge <16 x i32> %argv16i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I32 = icmp uge <32 x i32> %argv32i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp uge i64 %arg64, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp uge <2 x i64> %argv2i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I64 = icmp uge <4 x i64> %argv4i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I64 = icmp uge <8 x i64> %argv8i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I64 = icmp uge <16 x i64> %argv16i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; XOPAVX1-LABEL: 'cmp_int_uge'
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp uge i8 %arg8, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp uge <16 x i8> %argv16i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp uge <32 x i8> %argv32i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp uge <64 x i8> %argv64i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp uge <128 x i8> %argv128i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp uge i16 %arg16, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp uge <8 x i16> %argv8i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp uge <16 x i16> %argv16i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp uge <32 x i16> %argv32i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp uge <64 x i16> %argv64i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp uge i32 %arg32, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp uge <4 x i32> %argv4i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I32 = icmp uge <8 x i32> %argv8i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I32 = icmp uge <16 x i32> %argv16i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V32I32 = icmp uge <32 x i32> %argv32i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp uge i64 %arg64, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp uge <2 x i64> %argv2i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V4I64 = icmp uge <4 x i64> %argv4i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V8I64 = icmp uge <8 x i64> %argv8i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V16I64 = icmp uge <16 x i64> %argv16i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; XOPAVX2-LABEL: 'cmp_int_uge'
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp uge i8 %arg8, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp uge <16 x i8> %argv16i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I8 = icmp uge <32 x i8> %argv32i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I8 = icmp uge <64 x i8> %argv64i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V128I8 = icmp uge <128 x i8> %argv128i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp uge i16 %arg16, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp uge <8 x i16> %argv8i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I16 = icmp uge <16 x i16> %argv16i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I16 = icmp uge <32 x i16> %argv32i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I16 = icmp uge <64 x i16> %argv64i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp uge i32 %arg32, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp uge <4 x i32> %argv4i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I32 = icmp uge <8 x i32> %argv8i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I32 = icmp uge <16 x i32> %argv16i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I32 = icmp uge <32 x i32> %argv32i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp uge i64 %arg64, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp uge <2 x i64> %argv2i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V4I64 = icmp uge <4 x i64> %argv4i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V8I64 = icmp uge <8 x i64> %argv8i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V16I64 = icmp uge <16 x i64> %argv16i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SLM-LABEL: 'cmp_int_uge'
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp uge i8 %arg8, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp uge <16 x i8> %argv16i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp uge <32 x i8> %argv32i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp uge <64 x i8> %argv64i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp uge <128 x i8> %argv128i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp uge i16 %arg16, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp uge <8 x i16> %argv8i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp uge <16 x i16> %argv16i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp uge <32 x i16> %argv32i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp uge <64 x i16> %argv64i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp uge i32 %arg32, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I32 = icmp uge <4 x i32> %argv4i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I32 = icmp uge <8 x i32> %argv8i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I32 = icmp uge <16 x i32> %argv16i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V32I32 = icmp uge <32 x i32> %argv32i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp uge i64 %arg64, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V2I64 = icmp uge <2 x i64> %argv2i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V4I64 = icmp uge <4 x i64> %argv4i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 20 for instruction: %V8I64 = icmp uge <8 x i64> %argv8i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 40 for instruction: %V16I64 = icmp uge <16 x i64> %argv16i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+  %I8 = icmp uge i8 %arg8, zeroinitializer
+  %V16I8 = icmp uge <16 x i8> %argv16i8, zeroinitializer
+  %V32I8 = icmp uge <32 x i8> %argv32i8, zeroinitializer
+  %V64I8 = icmp uge <64 x i8> %argv64i8, zeroinitializer
+  %V128I8 = icmp uge <128 x i8> %argv128i8, zeroinitializer
+
+  %I16 = icmp uge i16 %arg16, zeroinitializer
+  %V8I16  = icmp uge <8 x i16> %argv8i16, zeroinitializer
+  %V16I16 = icmp uge <16 x i16> %argv16i16, zeroinitializer
+  %V32I16 = icmp uge <32 x i16> %argv32i16, zeroinitializer
+  %V64I16 = icmp uge <64 x i16> %argv64i16, zeroinitializer
+
+  %I32 = icmp uge i32 %arg32, zeroinitializer
+  %V4I32 = icmp uge <4 x i32> %argv4i32, zeroinitializer
+  %V8I32 = icmp uge <8 x i32> %argv8i32, zeroinitializer
+  %V16I32 = icmp uge <16 x i32> %argv16i32, zeroinitializer
+  %V32I32 = icmp uge <32 x i32> %argv32i32, zeroinitializer
+
+  %I64 = icmp uge i64 %arg64, zeroinitializer
+  %V2I64 = icmp uge <2 x i64> %argv2i64, zeroinitializer
+  %V4I64 = icmp uge <4 x i64> %argv4i64, zeroinitializer
+  %V8I64 = icmp uge <8 x i64> %argv8i64, zeroinitializer
+  %V16I64 = icmp uge <16 x i64> %argv16i64, zeroinitializer
+
+  ret i32 undef
+}
+
+define i32 @cmp_int_sgt(i8 %arg8, <16 x i8> %argv16i8, <32 x i8> %argv32i8, <64 x i8> %argv64i8, <128 x i8> %argv128i8, i16 %arg16, <8 x i16> %argv8i16, <16 x i16> %argv16i16, <32 x i16> %argv32i16, <64 x i16> %argv64i16, i32 %arg32, <4 x i32> %argv4i32, <8 x i32> %argv8i32, <16 x i32> %argv16i32, <32 x i32> %argv32i32, i64 %arg64, <2 x i64> %argv2i64, <4 x i64> %argv4i64, <8 x i64> %argv8i64, <16 x i64> %argv16i64) {
+; SSE2-LABEL: 'cmp_int_sgt'
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp sgt i8 %arg8, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp sgt <16 x i8> %argv16i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I8 = icmp sgt <32 x i8> %argv32i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I8 = icmp sgt <64 x i8> %argv64i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V128I8 = icmp sgt <128 x i8> %argv128i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp sgt i16 %arg16, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp sgt <8 x i16> %argv8i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I16 = icmp sgt <16 x i16> %argv16i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I16 = icmp sgt <32 x i16> %argv32i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I16 = icmp sgt <64 x i16> %argv64i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp sgt i32 %arg32, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp sgt <4 x i32> %argv4i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I32 = icmp sgt <8 x i32> %argv8i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I32 = icmp sgt <16 x i32> %argv16i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I32 = icmp sgt <32 x i32> %argv32i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp sgt i64 %arg64, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V2I64 = icmp sgt <2 x i64> %argv2i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V4I64 = icmp sgt <4 x i64> %argv4i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 20 for instruction: %V8I64 = icmp sgt <8 x i64> %argv8i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 40 for instruction: %V16I64 = icmp sgt <16 x i64> %argv16i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSE3-LABEL: 'cmp_int_sgt'
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp sgt i8 %arg8, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp sgt <16 x i8> %argv16i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I8 = icmp sgt <32 x i8> %argv32i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I8 = icmp sgt <64 x i8> %argv64i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V128I8 = icmp sgt <128 x i8> %argv128i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp sgt i16 %arg16, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp sgt <8 x i16> %argv8i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I16 = icmp sgt <16 x i16> %argv16i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I16 = icmp sgt <32 x i16> %argv32i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I16 = icmp sgt <64 x i16> %argv64i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp sgt i32 %arg32, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp sgt <4 x i32> %argv4i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I32 = icmp sgt <8 x i32> %argv8i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I32 = icmp sgt <16 x i32> %argv16i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I32 = icmp sgt <32 x i32> %argv32i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp sgt i64 %arg64, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V2I64 = icmp sgt <2 x i64> %argv2i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V4I64 = icmp sgt <4 x i64> %argv4i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 20 for instruction: %V8I64 = icmp sgt <8 x i64> %argv8i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 40 for instruction: %V16I64 = icmp sgt <16 x i64> %argv16i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSSE3-LABEL: 'cmp_int_sgt'
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp sgt i8 %arg8, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp sgt <16 x i8> %argv16i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I8 = icmp sgt <32 x i8> %argv32i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I8 = icmp sgt <64 x i8> %argv64i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V128I8 = icmp sgt <128 x i8> %argv128i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp sgt i16 %arg16, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp sgt <8 x i16> %argv8i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I16 = icmp sgt <16 x i16> %argv16i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I16 = icmp sgt <32 x i16> %argv32i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I16 = icmp sgt <64 x i16> %argv64i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp sgt i32 %arg32, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp sgt <4 x i32> %argv4i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I32 = icmp sgt <8 x i32> %argv8i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I32 = icmp sgt <16 x i32> %argv16i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I32 = icmp sgt <32 x i32> %argv32i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp sgt i64 %arg64, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V2I64 = icmp sgt <2 x i64> %argv2i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V4I64 = icmp sgt <4 x i64> %argv4i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 20 for instruction: %V8I64 = icmp sgt <8 x i64> %argv8i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 40 for instruction: %V16I64 = icmp sgt <16 x i64> %argv16i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSE41-LABEL: 'cmp_int_sgt'
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp sgt i8 %arg8, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp sgt <16 x i8> %argv16i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I8 = icmp sgt <32 x i8> %argv32i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I8 = icmp sgt <64 x i8> %argv64i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V128I8 = icmp sgt <128 x i8> %argv128i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp sgt i16 %arg16, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp sgt <8 x i16> %argv8i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I16 = icmp sgt <16 x i16> %argv16i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I16 = icmp sgt <32 x i16> %argv32i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I16 = icmp sgt <64 x i16> %argv64i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp sgt i32 %arg32, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp sgt <4 x i32> %argv4i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I32 = icmp sgt <8 x i32> %argv8i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I32 = icmp sgt <16 x i32> %argv16i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I32 = icmp sgt <32 x i32> %argv32i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp sgt i64 %arg64, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V2I64 = icmp sgt <2 x i64> %argv2i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V4I64 = icmp sgt <4 x i64> %argv4i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 20 for instruction: %V8I64 = icmp sgt <8 x i64> %argv8i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 40 for instruction: %V16I64 = icmp sgt <16 x i64> %argv16i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSE42-LABEL: 'cmp_int_sgt'
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp sgt i8 %arg8, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp sgt <16 x i8> %argv16i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I8 = icmp sgt <32 x i8> %argv32i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I8 = icmp sgt <64 x i8> %argv64i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V128I8 = icmp sgt <128 x i8> %argv128i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp sgt i16 %arg16, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp sgt <8 x i16> %argv8i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I16 = icmp sgt <16 x i16> %argv16i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I16 = icmp sgt <32 x i16> %argv32i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I16 = icmp sgt <64 x i16> %argv64i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp sgt i32 %arg32, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp sgt <4 x i32> %argv4i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I32 = icmp sgt <8 x i32> %argv8i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I32 = icmp sgt <16 x i32> %argv16i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I32 = icmp sgt <32 x i32> %argv32i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp sgt i64 %arg64, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp sgt <2 x i64> %argv2i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I64 = icmp sgt <4 x i64> %argv4i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I64 = icmp sgt <8 x i64> %argv8i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I64 = icmp sgt <16 x i64> %argv16i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX1-LABEL: 'cmp_int_sgt'
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp sgt i8 %arg8, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp sgt <16 x i8> %argv16i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp sgt <32 x i8> %argv32i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp sgt <64 x i8> %argv64i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp sgt <128 x i8> %argv128i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp sgt i16 %arg16, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp sgt <8 x i16> %argv8i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp sgt <16 x i16> %argv16i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp sgt <32 x i16> %argv32i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp sgt <64 x i16> %argv64i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp sgt i32 %arg32, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp sgt <4 x i32> %argv4i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I32 = icmp sgt <8 x i32> %argv8i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I32 = icmp sgt <16 x i32> %argv16i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V32I32 = icmp sgt <32 x i32> %argv32i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp sgt i64 %arg64, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp sgt <2 x i64> %argv2i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V4I64 = icmp sgt <4 x i64> %argv4i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V8I64 = icmp sgt <8 x i64> %argv8i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V16I64 = icmp sgt <16 x i64> %argv16i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX2-LABEL: 'cmp_int_sgt'
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp sgt i8 %arg8, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp sgt <16 x i8> %argv16i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V32I8 = icmp sgt <32 x i8> %argv32i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V64I8 = icmp sgt <64 x i8> %argv64i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V128I8 = icmp sgt <128 x i8> %argv128i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp sgt i16 %arg16, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp sgt <8 x i16> %argv8i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I16 = icmp sgt <16 x i16> %argv16i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I16 = icmp sgt <32 x i16> %argv32i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I16 = icmp sgt <64 x i16> %argv64i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp sgt i32 %arg32, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp sgt <4 x i32> %argv4i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I32 = icmp sgt <8 x i32> %argv8i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I32 = icmp sgt <16 x i32> %argv16i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I32 = icmp sgt <32 x i32> %argv32i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp sgt i64 %arg64, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp sgt <2 x i64> %argv2i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I64 = icmp sgt <4 x i64> %argv4i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I64 = icmp sgt <8 x i64> %argv8i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I64 = icmp sgt <16 x i64> %argv16i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX512F-LABEL: 'cmp_int_sgt'
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp sgt i8 %arg8, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp sgt <16 x i8> %argv16i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V32I8 = icmp sgt <32 x i8> %argv32i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V64I8 = icmp sgt <64 x i8> %argv64i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V128I8 = icmp sgt <128 x i8> %argv128i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp sgt i16 %arg16, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp sgt <8 x i16> %argv8i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I16 = icmp sgt <16 x i16> %argv16i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V32I16 = icmp sgt <32 x i16> %argv32i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V64I16 = icmp sgt <64 x i16> %argv64i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp sgt i32 %arg32, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp sgt <4 x i32> %argv4i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I32 = icmp sgt <8 x i32> %argv8i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I32 = icmp sgt <16 x i32> %argv16i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I32 = icmp sgt <32 x i32> %argv32i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp sgt i64 %arg64, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp sgt <2 x i64> %argv2i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I64 = icmp sgt <4 x i64> %argv4i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I64 = icmp sgt <8 x i64> %argv8i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I64 = icmp sgt <16 x i64> %argv16i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX512BW-LABEL: 'cmp_int_sgt'
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp sgt i8 %arg8, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp sgt <16 x i8> %argv16i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V32I8 = icmp sgt <32 x i8> %argv32i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V64I8 = icmp sgt <64 x i8> %argv64i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V128I8 = icmp sgt <128 x i8> %argv128i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp sgt i16 %arg16, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp sgt <8 x i16> %argv8i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I16 = icmp sgt <16 x i16> %argv16i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V32I16 = icmp sgt <32 x i16> %argv32i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V64I16 = icmp sgt <64 x i16> %argv64i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp sgt i32 %arg32, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp sgt <4 x i32> %argv4i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I32 = icmp sgt <8 x i32> %argv8i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I32 = icmp sgt <16 x i32> %argv16i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I32 = icmp sgt <32 x i32> %argv32i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp sgt i64 %arg64, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp sgt <2 x i64> %argv2i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I64 = icmp sgt <4 x i64> %argv4i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I64 = icmp sgt <8 x i64> %argv8i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I64 = icmp sgt <16 x i64> %argv16i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; XOPAVX1-LABEL: 'cmp_int_sgt'
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp sgt i8 %arg8, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp sgt <16 x i8> %argv16i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp sgt <32 x i8> %argv32i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp sgt <64 x i8> %argv64i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp sgt <128 x i8> %argv128i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp sgt i16 %arg16, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp sgt <8 x i16> %argv8i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp sgt <16 x i16> %argv16i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp sgt <32 x i16> %argv32i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp sgt <64 x i16> %argv64i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp sgt i32 %arg32, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp sgt <4 x i32> %argv4i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I32 = icmp sgt <8 x i32> %argv8i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I32 = icmp sgt <16 x i32> %argv16i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V32I32 = icmp sgt <32 x i32> %argv32i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp sgt i64 %arg64, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp sgt <2 x i64> %argv2i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V4I64 = icmp sgt <4 x i64> %argv4i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V8I64 = icmp sgt <8 x i64> %argv8i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V16I64 = icmp sgt <16 x i64> %argv16i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; XOPAVX2-LABEL: 'cmp_int_sgt'
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp sgt i8 %arg8, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp sgt <16 x i8> %argv16i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V32I8 = icmp sgt <32 x i8> %argv32i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V64I8 = icmp sgt <64 x i8> %argv64i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V128I8 = icmp sgt <128 x i8> %argv128i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp sgt i16 %arg16, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp sgt <8 x i16> %argv8i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I16 = icmp sgt <16 x i16> %argv16i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I16 = icmp sgt <32 x i16> %argv32i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I16 = icmp sgt <64 x i16> %argv64i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp sgt i32 %arg32, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp sgt <4 x i32> %argv4i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I32 = icmp sgt <8 x i32> %argv8i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I32 = icmp sgt <16 x i32> %argv16i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I32 = icmp sgt <32 x i32> %argv32i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp sgt i64 %arg64, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp sgt <2 x i64> %argv2i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I64 = icmp sgt <4 x i64> %argv4i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I64 = icmp sgt <8 x i64> %argv8i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I64 = icmp sgt <16 x i64> %argv16i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SLM-LABEL: 'cmp_int_sgt'
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp sgt i8 %arg8, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp sgt <16 x i8> %argv16i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I8 = icmp sgt <32 x i8> %argv32i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I8 = icmp sgt <64 x i8> %argv64i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V128I8 = icmp sgt <128 x i8> %argv128i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp sgt i16 %arg16, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp sgt <8 x i16> %argv8i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I16 = icmp sgt <16 x i16> %argv16i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I16 = icmp sgt <32 x i16> %argv32i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I16 = icmp sgt <64 x i16> %argv64i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp sgt i32 %arg32, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp sgt <4 x i32> %argv4i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I32 = icmp sgt <8 x i32> %argv8i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I32 = icmp sgt <16 x i32> %argv16i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I32 = icmp sgt <32 x i32> %argv32i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp sgt i64 %arg64, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V2I64 = icmp sgt <2 x i64> %argv2i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V4I64 = icmp sgt <4 x i64> %argv4i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V8I64 = icmp sgt <8 x i64> %argv8i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V16I64 = icmp sgt <16 x i64> %argv16i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+  %I8 = icmp sgt i8 %arg8, zeroinitializer
+  %V16I8 = icmp sgt <16 x i8> %argv16i8, zeroinitializer
+  %V32I8 = icmp sgt <32 x i8> %argv32i8, zeroinitializer
+  %V64I8 = icmp sgt <64 x i8> %argv64i8, zeroinitializer
+  %V128I8 = icmp sgt <128 x i8> %argv128i8, zeroinitializer
+
+  %I16 = icmp sgt i16 %arg16, zeroinitializer
+  %V8I16  = icmp sgt <8 x i16> %argv8i16, zeroinitializer
+  %V16I16 = icmp sgt <16 x i16> %argv16i16, zeroinitializer
+  %V32I16 = icmp sgt <32 x i16> %argv32i16, zeroinitializer
+  %V64I16 = icmp sgt <64 x i16> %argv64i16, zeroinitializer
+
+  %I32 = icmp sgt i32 %arg32, zeroinitializer
+  %V4I32 = icmp sgt <4 x i32> %argv4i32, zeroinitializer
+  %V8I32 = icmp sgt <8 x i32> %argv8i32, zeroinitializer
+  %V16I32 = icmp sgt <16 x i32> %argv16i32, zeroinitializer
+  %V32I32 = icmp sgt <32 x i32> %argv32i32, zeroinitializer
+
+  %I64 = icmp sgt i64 %arg64, zeroinitializer
+  %V2I64 = icmp sgt <2 x i64> %argv2i64, zeroinitializer
+  %V4I64 = icmp sgt <4 x i64> %argv4i64, zeroinitializer
+  %V8I64 = icmp sgt <8 x i64> %argv8i64, zeroinitializer
+  %V16I64 = icmp sgt <16 x i64> %argv16i64, zeroinitializer
+
+  ret i32 undef
+}
+
+define i32 @cmp_int_ugt(i8 %arg8, <16 x i8> %argv16i8, <32 x i8> %argv32i8, <64 x i8> %argv64i8, <128 x i8> %argv128i8, i16 %arg16, <8 x i16> %argv8i16, <16 x i16> %argv16i16, <32 x i16> %argv32i16, <64 x i16> %argv64i16, i32 %arg32, <4 x i32> %argv4i32, <8 x i32> %argv8i32, <16 x i32> %argv16i32, <32 x i32> %argv32i32, i64 %arg64, <2 x i64> %argv2i64, <4 x i64> %argv4i64, <8 x i64> %argv8i64, <16 x i64> %argv16i64) {
+; SSE2-LABEL: 'cmp_int_ugt'
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ugt i8 %arg8, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V16I8 = icmp ugt <16 x i8> %argv16i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V32I8 = icmp ugt <32 x i8> %argv32i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V64I8 = icmp ugt <64 x i8> %argv64i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V128I8 = icmp ugt <128 x i8> %argv128i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ugt i16 %arg16, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V8I16 = icmp ugt <8 x i16> %argv8i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V16I16 = icmp ugt <16 x i16> %argv16i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V32I16 = icmp ugt <32 x i16> %argv32i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V64I16 = icmp ugt <64 x i16> %argv64i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ugt i32 %arg32, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V4I32 = icmp ugt <4 x i32> %argv4i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V8I32 = icmp ugt <8 x i32> %argv8i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V16I32 = icmp ugt <16 x i32> %argv16i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V32I32 = icmp ugt <32 x i32> %argv32i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ugt i64 %arg64, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 7 for instruction: %V2I64 = icmp ugt <2 x i64> %argv2i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 14 for instruction: %V4I64 = icmp ugt <4 x i64> %argv4i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 28 for instruction: %V8I64 = icmp ugt <8 x i64> %argv8i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 56 for instruction: %V16I64 = icmp ugt <16 x i64> %argv16i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSE3-LABEL: 'cmp_int_ugt'
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ugt i8 %arg8, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V16I8 = icmp ugt <16 x i8> %argv16i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V32I8 = icmp ugt <32 x i8> %argv32i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V64I8 = icmp ugt <64 x i8> %argv64i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V128I8 = icmp ugt <128 x i8> %argv128i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ugt i16 %arg16, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V8I16 = icmp ugt <8 x i16> %argv8i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V16I16 = icmp ugt <16 x i16> %argv16i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V32I16 = icmp ugt <32 x i16> %argv32i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V64I16 = icmp ugt <64 x i16> %argv64i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ugt i32 %arg32, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V4I32 = icmp ugt <4 x i32> %argv4i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V8I32 = icmp ugt <8 x i32> %argv8i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V16I32 = icmp ugt <16 x i32> %argv16i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V32I32 = icmp ugt <32 x i32> %argv32i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ugt i64 %arg64, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 7 for instruction: %V2I64 = icmp ugt <2 x i64> %argv2i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 14 for instruction: %V4I64 = icmp ugt <4 x i64> %argv4i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 28 for instruction: %V8I64 = icmp ugt <8 x i64> %argv8i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 56 for instruction: %V16I64 = icmp ugt <16 x i64> %argv16i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSSE3-LABEL: 'cmp_int_ugt'
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ugt i8 %arg8, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V16I8 = icmp ugt <16 x i8> %argv16i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V32I8 = icmp ugt <32 x i8> %argv32i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V64I8 = icmp ugt <64 x i8> %argv64i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V128I8 = icmp ugt <128 x i8> %argv128i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ugt i16 %arg16, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V8I16 = icmp ugt <8 x i16> %argv8i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V16I16 = icmp ugt <16 x i16> %argv16i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V32I16 = icmp ugt <32 x i16> %argv32i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V64I16 = icmp ugt <64 x i16> %argv64i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ugt i32 %arg32, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V4I32 = icmp ugt <4 x i32> %argv4i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V8I32 = icmp ugt <8 x i32> %argv8i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V16I32 = icmp ugt <16 x i32> %argv16i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V32I32 = icmp ugt <32 x i32> %argv32i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ugt i64 %arg64, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 7 for instruction: %V2I64 = icmp ugt <2 x i64> %argv2i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 14 for instruction: %V4I64 = icmp ugt <4 x i64> %argv4i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 28 for instruction: %V8I64 = icmp ugt <8 x i64> %argv8i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 56 for instruction: %V16I64 = icmp ugt <16 x i64> %argv16i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSE41-LABEL: 'cmp_int_ugt'
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ugt i8 %arg8, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V16I8 = icmp ugt <16 x i8> %argv16i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V32I8 = icmp ugt <32 x i8> %argv32i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V64I8 = icmp ugt <64 x i8> %argv64i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V128I8 = icmp ugt <128 x i8> %argv128i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ugt i16 %arg16, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V8I16 = icmp ugt <8 x i16> %argv8i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V16I16 = icmp ugt <16 x i16> %argv16i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V32I16 = icmp ugt <32 x i16> %argv32i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V64I16 = icmp ugt <64 x i16> %argv64i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ugt i32 %arg32, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V4I32 = icmp ugt <4 x i32> %argv4i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V8I32 = icmp ugt <8 x i32> %argv8i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V16I32 = icmp ugt <16 x i32> %argv16i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V32I32 = icmp ugt <32 x i32> %argv32i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ugt i64 %arg64, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 7 for instruction: %V2I64 = icmp ugt <2 x i64> %argv2i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 14 for instruction: %V4I64 = icmp ugt <4 x i64> %argv4i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 28 for instruction: %V8I64 = icmp ugt <8 x i64> %argv8i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 56 for instruction: %V16I64 = icmp ugt <16 x i64> %argv16i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSE42-LABEL: 'cmp_int_ugt'
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ugt i8 %arg8, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V16I8 = icmp ugt <16 x i8> %argv16i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V32I8 = icmp ugt <32 x i8> %argv32i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V64I8 = icmp ugt <64 x i8> %argv64i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V128I8 = icmp ugt <128 x i8> %argv128i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ugt i16 %arg16, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V8I16 = icmp ugt <8 x i16> %argv8i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V16I16 = icmp ugt <16 x i16> %argv16i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V32I16 = icmp ugt <32 x i16> %argv32i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V64I16 = icmp ugt <64 x i16> %argv64i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ugt i32 %arg32, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V4I32 = icmp ugt <4 x i32> %argv4i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V8I32 = icmp ugt <8 x i32> %argv8i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V16I32 = icmp ugt <16 x i32> %argv16i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V32I32 = icmp ugt <32 x i32> %argv32i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ugt i64 %arg64, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V2I64 = icmp ugt <2 x i64> %argv2i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V4I64 = icmp ugt <4 x i64> %argv4i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V8I64 = icmp ugt <8 x i64> %argv8i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V16I64 = icmp ugt <16 x i64> %argv16i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX1-LABEL: 'cmp_int_ugt'
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ugt i8 %arg8, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V16I8 = icmp ugt <16 x i8> %argv16i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V32I8 = icmp ugt <32 x i8> %argv32i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V64I8 = icmp ugt <64 x i8> %argv64i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V128I8 = icmp ugt <128 x i8> %argv128i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ugt i16 %arg16, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V8I16 = icmp ugt <8 x i16> %argv8i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V16I16 = icmp ugt <16 x i16> %argv16i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V32I16 = icmp ugt <32 x i16> %argv32i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V64I16 = icmp ugt <64 x i16> %argv64i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ugt i32 %arg32, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V4I32 = icmp ugt <4 x i32> %argv4i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V8I32 = icmp ugt <8 x i32> %argv8i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V16I32 = icmp ugt <16 x i32> %argv16i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V32I32 = icmp ugt <32 x i32> %argv32i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ugt i64 %arg64, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V2I64 = icmp ugt <2 x i64> %argv2i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V4I64 = icmp ugt <4 x i64> %argv4i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V8I64 = icmp ugt <8 x i64> %argv8i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V16I64 = icmp ugt <16 x i64> %argv16i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX2-LABEL: 'cmp_int_ugt'
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ugt i8 %arg8, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V16I8 = icmp ugt <16 x i8> %argv16i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V32I8 = icmp ugt <32 x i8> %argv32i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V64I8 = icmp ugt <64 x i8> %argv64i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V128I8 = icmp ugt <128 x i8> %argv128i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ugt i16 %arg16, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V8I16 = icmp ugt <8 x i16> %argv8i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V16I16 = icmp ugt <16 x i16> %argv16i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V32I16 = icmp ugt <32 x i16> %argv32i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V64I16 = icmp ugt <64 x i16> %argv64i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ugt i32 %arg32, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V4I32 = icmp ugt <4 x i32> %argv4i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V8I32 = icmp ugt <8 x i32> %argv8i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V16I32 = icmp ugt <16 x i32> %argv16i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V32I32 = icmp ugt <32 x i32> %argv32i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ugt i64 %arg64, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V2I64 = icmp ugt <2 x i64> %argv2i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V4I64 = icmp ugt <4 x i64> %argv4i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V8I64 = icmp ugt <8 x i64> %argv8i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V16I64 = icmp ugt <16 x i64> %argv16i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX512F-LABEL: 'cmp_int_ugt'
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ugt i8 %arg8, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V16I8 = icmp ugt <16 x i8> %argv16i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V32I8 = icmp ugt <32 x i8> %argv32i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V64I8 = icmp ugt <64 x i8> %argv64i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V128I8 = icmp ugt <128 x i8> %argv128i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ugt i16 %arg16, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V8I16 = icmp ugt <8 x i16> %argv8i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V16I16 = icmp ugt <16 x i16> %argv16i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V32I16 = icmp ugt <32 x i16> %argv32i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V64I16 = icmp ugt <64 x i16> %argv64i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ugt i32 %arg32, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp ugt <4 x i32> %argv4i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I32 = icmp ugt <8 x i32> %argv8i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I32 = icmp ugt <16 x i32> %argv16i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I32 = icmp ugt <32 x i32> %argv32i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ugt i64 %arg64, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp ugt <2 x i64> %argv2i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I64 = icmp ugt <4 x i64> %argv4i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I64 = icmp ugt <8 x i64> %argv8i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I64 = icmp ugt <16 x i64> %argv16i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX512BW-LABEL: 'cmp_int_ugt'
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ugt i8 %arg8, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp ugt <16 x i8> %argv16i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V32I8 = icmp ugt <32 x i8> %argv32i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V64I8 = icmp ugt <64 x i8> %argv64i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V128I8 = icmp ugt <128 x i8> %argv128i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ugt i16 %arg16, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp ugt <8 x i16> %argv8i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I16 = icmp ugt <16 x i16> %argv16i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V32I16 = icmp ugt <32 x i16> %argv32i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V64I16 = icmp ugt <64 x i16> %argv64i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ugt i32 %arg32, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp ugt <4 x i32> %argv4i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I32 = icmp ugt <8 x i32> %argv8i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I32 = icmp ugt <16 x i32> %argv16i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I32 = icmp ugt <32 x i32> %argv32i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ugt i64 %arg64, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp ugt <2 x i64> %argv2i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I64 = icmp ugt <4 x i64> %argv4i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I64 = icmp ugt <8 x i64> %argv8i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I64 = icmp ugt <16 x i64> %argv16i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; XOPAVX1-LABEL: 'cmp_int_ugt'
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ugt i8 %arg8, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp ugt <16 x i8> %argv16i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp ugt <32 x i8> %argv32i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp ugt <64 x i8> %argv64i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp ugt <128 x i8> %argv128i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ugt i16 %arg16, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp ugt <8 x i16> %argv8i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp ugt <16 x i16> %argv16i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp ugt <32 x i16> %argv32i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp ugt <64 x i16> %argv64i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ugt i32 %arg32, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp ugt <4 x i32> %argv4i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I32 = icmp ugt <8 x i32> %argv8i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I32 = icmp ugt <16 x i32> %argv16i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V32I32 = icmp ugt <32 x i32> %argv32i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ugt i64 %arg64, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp ugt <2 x i64> %argv2i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V4I64 = icmp ugt <4 x i64> %argv4i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V8I64 = icmp ugt <8 x i64> %argv8i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V16I64 = icmp ugt <16 x i64> %argv16i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; XOPAVX2-LABEL: 'cmp_int_ugt'
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ugt i8 %arg8, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp ugt <16 x i8> %argv16i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V32I8 = icmp ugt <32 x i8> %argv32i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V64I8 = icmp ugt <64 x i8> %argv64i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V128I8 = icmp ugt <128 x i8> %argv128i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ugt i16 %arg16, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp ugt <8 x i16> %argv8i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V16I16 = icmp ugt <16 x i16> %argv16i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V32I16 = icmp ugt <32 x i16> %argv32i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V64I16 = icmp ugt <64 x i16> %argv64i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ugt i32 %arg32, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp ugt <4 x i32> %argv4i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V8I32 = icmp ugt <8 x i32> %argv8i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V16I32 = icmp ugt <16 x i32> %argv16i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V32I32 = icmp ugt <32 x i32> %argv32i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ugt i64 %arg64, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp ugt <2 x i64> %argv2i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V4I64 = icmp ugt <4 x i64> %argv4i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V8I64 = icmp ugt <8 x i64> %argv8i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V16I64 = icmp ugt <16 x i64> %argv16i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SLM-LABEL: 'cmp_int_ugt'
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ugt i8 %arg8, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V16I8 = icmp ugt <16 x i8> %argv16i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V32I8 = icmp ugt <32 x i8> %argv32i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V64I8 = icmp ugt <64 x i8> %argv64i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V128I8 = icmp ugt <128 x i8> %argv128i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ugt i16 %arg16, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V8I16 = icmp ugt <8 x i16> %argv8i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V16I16 = icmp ugt <16 x i16> %argv16i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V32I16 = icmp ugt <32 x i16> %argv32i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V64I16 = icmp ugt <64 x i16> %argv64i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ugt i32 %arg32, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V4I32 = icmp ugt <4 x i32> %argv4i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V8I32 = icmp ugt <8 x i32> %argv8i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V16I32 = icmp ugt <16 x i32> %argv16i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V32I32 = icmp ugt <32 x i32> %argv32i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ugt i64 %arg64, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V2I64 = icmp ugt <2 x i64> %argv2i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V4I64 = icmp ugt <4 x i64> %argv4i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V8I64 = icmp ugt <8 x i64> %argv8i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 32 for instruction: %V16I64 = icmp ugt <16 x i64> %argv16i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+  %I8 = icmp ugt i8 %arg8, zeroinitializer
+  %V16I8 = icmp ugt <16 x i8> %argv16i8, zeroinitializer
+  %V32I8 = icmp ugt <32 x i8> %argv32i8, zeroinitializer
+  %V64I8 = icmp ugt <64 x i8> %argv64i8, zeroinitializer
+  %V128I8 = icmp ugt <128 x i8> %argv128i8, zeroinitializer
+
+  %I16 = icmp ugt i16 %arg16, zeroinitializer
+  %V8I16  = icmp ugt <8 x i16> %argv8i16, zeroinitializer
+  %V16I16 = icmp ugt <16 x i16> %argv16i16, zeroinitializer
+  %V32I16 = icmp ugt <32 x i16> %argv32i16, zeroinitializer
+  %V64I16 = icmp ugt <64 x i16> %argv64i16, zeroinitializer
+
+  %I32 = icmp ugt i32 %arg32, zeroinitializer
+  %V4I32 = icmp ugt <4 x i32> %argv4i32, zeroinitializer
+  %V8I32 = icmp ugt <8 x i32> %argv8i32, zeroinitializer
+  %V16I32 = icmp ugt <16 x i32> %argv16i32, zeroinitializer
+  %V32I32 = icmp ugt <32 x i32> %argv32i32, zeroinitializer
+
+  %I64 = icmp ugt i64 %arg64, zeroinitializer
+  %V2I64 = icmp ugt <2 x i64> %argv2i64, zeroinitializer
+  %V4I64 = icmp ugt <4 x i64> %argv4i64, zeroinitializer
+  %V8I64 = icmp ugt <8 x i64> %argv8i64, zeroinitializer
+  %V16I64 = icmp ugt <16 x i64> %argv16i64, zeroinitializer
+
+  ret i32 undef
+}
+
+define i32 @cmp_int_sle(i8 %arg8, <16 x i8> %argv16i8, <32 x i8> %argv32i8, <64 x i8> %argv64i8, <128 x i8> %argv128i8, i16 %arg16, <8 x i16> %argv8i16, <16 x i16> %argv16i16, <32 x i16> %argv32i16, <64 x i16> %argv64i16, i32 %arg32, <4 x i32> %argv4i32, <8 x i32> %argv8i32, <16 x i32> %argv16i32, <32 x i32> %argv32i32, i64 %arg64, <2 x i64> %argv2i64, <4 x i64> %argv4i64, <8 x i64> %argv8i64, <16 x i64> %argv16i64) {
+; SSE2-LABEL: 'cmp_int_sle'
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp sle i8 %arg8, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp sle <16 x i8> %argv16i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp sle <32 x i8> %argv32i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp sle <64 x i8> %argv64i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp sle <128 x i8> %argv128i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp sle i16 %arg16, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp sle <8 x i16> %argv8i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp sle <16 x i16> %argv16i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp sle <32 x i16> %argv32i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp sle <64 x i16> %argv64i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp sle i32 %arg32, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I32 = icmp sle <4 x i32> %argv4i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I32 = icmp sle <8 x i32> %argv8i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I32 = icmp sle <16 x i32> %argv16i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V32I32 = icmp sle <32 x i32> %argv32i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp sle i64 %arg64, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V2I64 = icmp sle <2 x i64> %argv2i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V4I64 = icmp sle <4 x i64> %argv4i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V8I64 = icmp sle <8 x i64> %argv8i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 48 for instruction: %V16I64 = icmp sle <16 x i64> %argv16i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSE3-LABEL: 'cmp_int_sle'
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp sle i8 %arg8, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp sle <16 x i8> %argv16i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp sle <32 x i8> %argv32i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp sle <64 x i8> %argv64i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp sle <128 x i8> %argv128i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp sle i16 %arg16, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp sle <8 x i16> %argv8i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp sle <16 x i16> %argv16i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp sle <32 x i16> %argv32i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp sle <64 x i16> %argv64i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp sle i32 %arg32, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I32 = icmp sle <4 x i32> %argv4i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I32 = icmp sle <8 x i32> %argv8i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I32 = icmp sle <16 x i32> %argv16i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V32I32 = icmp sle <32 x i32> %argv32i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp sle i64 %arg64, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V2I64 = icmp sle <2 x i64> %argv2i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V4I64 = icmp sle <4 x i64> %argv4i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V8I64 = icmp sle <8 x i64> %argv8i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 48 for instruction: %V16I64 = icmp sle <16 x i64> %argv16i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSSE3-LABEL: 'cmp_int_sle'
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp sle i8 %arg8, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp sle <16 x i8> %argv16i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp sle <32 x i8> %argv32i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp sle <64 x i8> %argv64i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp sle <128 x i8> %argv128i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp sle i16 %arg16, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp sle <8 x i16> %argv8i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp sle <16 x i16> %argv16i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp sle <32 x i16> %argv32i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp sle <64 x i16> %argv64i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp sle i32 %arg32, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I32 = icmp sle <4 x i32> %argv4i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I32 = icmp sle <8 x i32> %argv8i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I32 = icmp sle <16 x i32> %argv16i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V32I32 = icmp sle <32 x i32> %argv32i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp sle i64 %arg64, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V2I64 = icmp sle <2 x i64> %argv2i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V4I64 = icmp sle <4 x i64> %argv4i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V8I64 = icmp sle <8 x i64> %argv8i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 48 for instruction: %V16I64 = icmp sle <16 x i64> %argv16i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSE41-LABEL: 'cmp_int_sle'
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp sle i8 %arg8, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp sle <16 x i8> %argv16i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp sle <32 x i8> %argv32i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp sle <64 x i8> %argv64i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp sle <128 x i8> %argv128i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp sle i16 %arg16, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp sle <8 x i16> %argv8i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp sle <16 x i16> %argv16i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp sle <32 x i16> %argv32i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp sle <64 x i16> %argv64i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp sle i32 %arg32, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I32 = icmp sle <4 x i32> %argv4i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I32 = icmp sle <8 x i32> %argv8i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I32 = icmp sle <16 x i32> %argv16i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V32I32 = icmp sle <32 x i32> %argv32i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp sle i64 %arg64, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V2I64 = icmp sle <2 x i64> %argv2i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V4I64 = icmp sle <4 x i64> %argv4i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V8I64 = icmp sle <8 x i64> %argv8i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 48 for instruction: %V16I64 = icmp sle <16 x i64> %argv16i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSE42-LABEL: 'cmp_int_sle'
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp sle i8 %arg8, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp sle <16 x i8> %argv16i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp sle <32 x i8> %argv32i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp sle <64 x i8> %argv64i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp sle <128 x i8> %argv128i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp sle i16 %arg16, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp sle <8 x i16> %argv8i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp sle <16 x i16> %argv16i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp sle <32 x i16> %argv32i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp sle <64 x i16> %argv64i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp sle i32 %arg32, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I32 = icmp sle <4 x i32> %argv4i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I32 = icmp sle <8 x i32> %argv8i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I32 = icmp sle <16 x i32> %argv16i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V32I32 = icmp sle <32 x i32> %argv32i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp sle i64 %arg64, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V2I64 = icmp sle <2 x i64> %argv2i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V4I64 = icmp sle <4 x i64> %argv4i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V8I64 = icmp sle <8 x i64> %argv8i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V16I64 = icmp sle <16 x i64> %argv16i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX1-LABEL: 'cmp_int_sle'
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp sle i8 %arg8, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp sle <16 x i8> %argv16i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V32I8 = icmp sle <32 x i8> %argv32i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V64I8 = icmp sle <64 x i8> %argv64i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 20 for instruction: %V128I8 = icmp sle <128 x i8> %argv128i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp sle i16 %arg16, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp sle <8 x i16> %argv8i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V16I16 = icmp sle <16 x i16> %argv16i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V32I16 = icmp sle <32 x i16> %argv32i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 20 for instruction: %V64I16 = icmp sle <64 x i16> %argv64i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp sle i32 %arg32, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I32 = icmp sle <4 x i32> %argv4i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V8I32 = icmp sle <8 x i32> %argv8i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V16I32 = icmp sle <16 x i32> %argv16i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 20 for instruction: %V32I32 = icmp sle <32 x i32> %argv32i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp sle i64 %arg64, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V2I64 = icmp sle <2 x i64> %argv2i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V4I64 = icmp sle <4 x i64> %argv4i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V8I64 = icmp sle <8 x i64> %argv8i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 20 for instruction: %V16I64 = icmp sle <16 x i64> %argv16i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX2-LABEL: 'cmp_int_sle'
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp sle i8 %arg8, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp sle <16 x i8> %argv16i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I8 = icmp sle <32 x i8> %argv32i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I8 = icmp sle <64 x i8> %argv64i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V128I8 = icmp sle <128 x i8> %argv128i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp sle i16 %arg16, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp sle <8 x i16> %argv8i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I16 = icmp sle <16 x i16> %argv16i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I16 = icmp sle <32 x i16> %argv32i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I16 = icmp sle <64 x i16> %argv64i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp sle i32 %arg32, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I32 = icmp sle <4 x i32> %argv4i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I32 = icmp sle <8 x i32> %argv8i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I32 = icmp sle <16 x i32> %argv16i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I32 = icmp sle <32 x i32> %argv32i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp sle i64 %arg64, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V2I64 = icmp sle <2 x i64> %argv2i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I64 = icmp sle <4 x i64> %argv4i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I64 = icmp sle <8 x i64> %argv8i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I64 = icmp sle <16 x i64> %argv16i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX512F-LABEL: 'cmp_int_sle'
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp sle i8 %arg8, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp sle <16 x i8> %argv16i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I8 = icmp sle <32 x i8> %argv32i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I8 = icmp sle <64 x i8> %argv64i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V128I8 = icmp sle <128 x i8> %argv128i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp sle i16 %arg16, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp sle <8 x i16> %argv8i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I16 = icmp sle <16 x i16> %argv16i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I16 = icmp sle <32 x i16> %argv32i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I16 = icmp sle <64 x i16> %argv64i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp sle i32 %arg32, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp sle <4 x i32> %argv4i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I32 = icmp sle <8 x i32> %argv8i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I32 = icmp sle <16 x i32> %argv16i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I32 = icmp sle <32 x i32> %argv32i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp sle i64 %arg64, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp sle <2 x i64> %argv2i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I64 = icmp sle <4 x i64> %argv4i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I64 = icmp sle <8 x i64> %argv8i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I64 = icmp sle <16 x i64> %argv16i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX512BW-LABEL: 'cmp_int_sle'
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp sle i8 %arg8, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp sle <16 x i8> %argv16i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V32I8 = icmp sle <32 x i8> %argv32i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V64I8 = icmp sle <64 x i8> %argv64i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V128I8 = icmp sle <128 x i8> %argv128i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp sle i16 %arg16, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp sle <8 x i16> %argv8i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I16 = icmp sle <16 x i16> %argv16i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V32I16 = icmp sle <32 x i16> %argv32i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V64I16 = icmp sle <64 x i16> %argv64i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp sle i32 %arg32, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp sle <4 x i32> %argv4i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I32 = icmp sle <8 x i32> %argv8i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I32 = icmp sle <16 x i32> %argv16i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I32 = icmp sle <32 x i32> %argv32i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp sle i64 %arg64, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp sle <2 x i64> %argv2i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I64 = icmp sle <4 x i64> %argv4i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I64 = icmp sle <8 x i64> %argv8i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I64 = icmp sle <16 x i64> %argv16i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; XOPAVX1-LABEL: 'cmp_int_sle'
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp sle i8 %arg8, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp sle <16 x i8> %argv16i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp sle <32 x i8> %argv32i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp sle <64 x i8> %argv64i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp sle <128 x i8> %argv128i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp sle i16 %arg16, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp sle <8 x i16> %argv8i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp sle <16 x i16> %argv16i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp sle <32 x i16> %argv32i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp sle <64 x i16> %argv64i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp sle i32 %arg32, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp sle <4 x i32> %argv4i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I32 = icmp sle <8 x i32> %argv8i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I32 = icmp sle <16 x i32> %argv16i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V32I32 = icmp sle <32 x i32> %argv32i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp sle i64 %arg64, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp sle <2 x i64> %argv2i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V4I64 = icmp sle <4 x i64> %argv4i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V8I64 = icmp sle <8 x i64> %argv8i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V16I64 = icmp sle <16 x i64> %argv16i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; XOPAVX2-LABEL: 'cmp_int_sle'
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp sle i8 %arg8, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp sle <16 x i8> %argv16i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I8 = icmp sle <32 x i8> %argv32i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I8 = icmp sle <64 x i8> %argv64i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V128I8 = icmp sle <128 x i8> %argv128i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp sle i16 %arg16, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp sle <8 x i16> %argv8i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I16 = icmp sle <16 x i16> %argv16i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I16 = icmp sle <32 x i16> %argv32i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I16 = icmp sle <64 x i16> %argv64i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp sle i32 %arg32, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp sle <4 x i32> %argv4i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I32 = icmp sle <8 x i32> %argv8i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I32 = icmp sle <16 x i32> %argv16i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I32 = icmp sle <32 x i32> %argv32i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp sle i64 %arg64, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp sle <2 x i64> %argv2i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I64 = icmp sle <4 x i64> %argv4i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I64 = icmp sle <8 x i64> %argv8i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I64 = icmp sle <16 x i64> %argv16i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SLM-LABEL: 'cmp_int_sle'
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp sle i8 %arg8, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp sle <16 x i8> %argv16i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp sle <32 x i8> %argv32i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp sle <64 x i8> %argv64i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp sle <128 x i8> %argv128i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp sle i16 %arg16, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp sle <8 x i16> %argv8i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp sle <16 x i16> %argv16i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp sle <32 x i16> %argv32i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp sle <64 x i16> %argv64i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp sle i32 %arg32, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I32 = icmp sle <4 x i32> %argv4i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I32 = icmp sle <8 x i32> %argv8i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I32 = icmp sle <16 x i32> %argv16i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V32I32 = icmp sle <32 x i32> %argv32i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp sle i64 %arg64, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V2I64 = icmp sle <2 x i64> %argv2i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V4I64 = icmp sle <4 x i64> %argv4i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V8I64 = icmp sle <8 x i64> %argv8i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V16I64 = icmp sle <16 x i64> %argv16i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+  %I8 = icmp sle i8 %arg8, zeroinitializer
+  %V16I8 = icmp sle <16 x i8> %argv16i8, zeroinitializer
+  %V32I8 = icmp sle <32 x i8> %argv32i8, zeroinitializer
+  %V64I8 = icmp sle <64 x i8> %argv64i8, zeroinitializer
+  %V128I8 = icmp sle <128 x i8> %argv128i8, zeroinitializer
+
+  %I16 = icmp sle i16 %arg16, zeroinitializer
+  %V8I16  = icmp sle <8 x i16> %argv8i16, zeroinitializer
+  %V16I16 = icmp sle <16 x i16> %argv16i16, zeroinitializer
+  %V32I16 = icmp sle <32 x i16> %argv32i16, zeroinitializer
+  %V64I16 = icmp sle <64 x i16> %argv64i16, zeroinitializer
+
+  %I32 = icmp sle i32 %arg32, zeroinitializer
+  %V4I32 = icmp sle <4 x i32> %argv4i32, zeroinitializer
+  %V8I32 = icmp sle <8 x i32> %argv8i32, zeroinitializer
+  %V16I32 = icmp sle <16 x i32> %argv16i32, zeroinitializer
+  %V32I32 = icmp sle <32 x i32> %argv32i32, zeroinitializer
+
+  %I64 = icmp sle i64 %arg64, zeroinitializer
+  %V2I64 = icmp sle <2 x i64> %argv2i64, zeroinitializer
+  %V4I64 = icmp sle <4 x i64> %argv4i64, zeroinitializer
+  %V8I64 = icmp sle <8 x i64> %argv8i64, zeroinitializer
+  %V16I64 = icmp sle <16 x i64> %argv16i64, zeroinitializer
+
+  ret i32 undef
+}
+
+define i32 @cmp_int_ule(i8 %arg8, <16 x i8> %argv16i8, <32 x i8> %argv32i8, <64 x i8> %argv64i8, <128 x i8> %argv128i8, i16 %arg16, <8 x i16> %argv8i16, <16 x i16> %argv16i16, <32 x i16> %argv32i16, <64 x i16> %argv64i16, i32 %arg32, <4 x i32> %argv4i32, <8 x i32> %argv8i32, <16 x i32> %argv16i32, <32 x i32> %argv32i32, i64 %arg64, <2 x i64> %argv2i64, <4 x i64> %argv4i64, <8 x i64> %argv8i64, <16 x i64> %argv16i64) {
+; SSE2-LABEL: 'cmp_int_ule'
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ule i8 %arg8, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp ule <16 x i8> %argv16i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp ule <32 x i8> %argv32i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp ule <64 x i8> %argv64i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp ule <128 x i8> %argv128i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ule i16 %arg16, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp ule <8 x i16> %argv8i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp ule <16 x i16> %argv16i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp ule <32 x i16> %argv32i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp ule <64 x i16> %argv64i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ule i32 %arg32, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V4I32 = icmp ule <4 x i32> %argv4i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V8I32 = icmp ule <8 x i32> %argv8i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V16I32 = icmp ule <16 x i32> %argv16i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 32 for instruction: %V32I32 = icmp ule <32 x i32> %argv32i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ule i64 %arg64, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V2I64 = icmp ule <2 x i64> %argv2i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V4I64 = icmp ule <4 x i64> %argv4i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 32 for instruction: %V8I64 = icmp ule <8 x i64> %argv8i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 64 for instruction: %V16I64 = icmp ule <16 x i64> %argv16i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSE3-LABEL: 'cmp_int_ule'
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ule i8 %arg8, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp ule <16 x i8> %argv16i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp ule <32 x i8> %argv32i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp ule <64 x i8> %argv64i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp ule <128 x i8> %argv128i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ule i16 %arg16, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp ule <8 x i16> %argv8i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp ule <16 x i16> %argv16i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp ule <32 x i16> %argv32i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp ule <64 x i16> %argv64i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ule i32 %arg32, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V4I32 = icmp ule <4 x i32> %argv4i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V8I32 = icmp ule <8 x i32> %argv8i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V16I32 = icmp ule <16 x i32> %argv16i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 32 for instruction: %V32I32 = icmp ule <32 x i32> %argv32i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ule i64 %arg64, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V2I64 = icmp ule <2 x i64> %argv2i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V4I64 = icmp ule <4 x i64> %argv4i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 32 for instruction: %V8I64 = icmp ule <8 x i64> %argv8i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 64 for instruction: %V16I64 = icmp ule <16 x i64> %argv16i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSSE3-LABEL: 'cmp_int_ule'
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ule i8 %arg8, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp ule <16 x i8> %argv16i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp ule <32 x i8> %argv32i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp ule <64 x i8> %argv64i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp ule <128 x i8> %argv128i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ule i16 %arg16, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp ule <8 x i16> %argv8i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp ule <16 x i16> %argv16i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp ule <32 x i16> %argv32i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp ule <64 x i16> %argv64i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ule i32 %arg32, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V4I32 = icmp ule <4 x i32> %argv4i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V8I32 = icmp ule <8 x i32> %argv8i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V16I32 = icmp ule <16 x i32> %argv16i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 32 for instruction: %V32I32 = icmp ule <32 x i32> %argv32i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ule i64 %arg64, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V2I64 = icmp ule <2 x i64> %argv2i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V4I64 = icmp ule <4 x i64> %argv4i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 32 for instruction: %V8I64 = icmp ule <8 x i64> %argv8i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 64 for instruction: %V16I64 = icmp ule <16 x i64> %argv16i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSE41-LABEL: 'cmp_int_ule'
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ule i8 %arg8, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp ule <16 x i8> %argv16i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp ule <32 x i8> %argv32i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp ule <64 x i8> %argv64i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp ule <128 x i8> %argv128i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ule i16 %arg16, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp ule <8 x i16> %argv8i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp ule <16 x i16> %argv16i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp ule <32 x i16> %argv32i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp ule <64 x i16> %argv64i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ule i32 %arg32, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I32 = icmp ule <4 x i32> %argv4i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I32 = icmp ule <8 x i32> %argv8i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I32 = icmp ule <16 x i32> %argv16i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V32I32 = icmp ule <32 x i32> %argv32i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ule i64 %arg64, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V2I64 = icmp ule <2 x i64> %argv2i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V4I64 = icmp ule <4 x i64> %argv4i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 32 for instruction: %V8I64 = icmp ule <8 x i64> %argv8i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 64 for instruction: %V16I64 = icmp ule <16 x i64> %argv16i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSE42-LABEL: 'cmp_int_ule'
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ule i8 %arg8, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp ule <16 x i8> %argv16i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp ule <32 x i8> %argv32i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp ule <64 x i8> %argv64i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp ule <128 x i8> %argv128i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ule i16 %arg16, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp ule <8 x i16> %argv8i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp ule <16 x i16> %argv16i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp ule <32 x i16> %argv32i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp ule <64 x i16> %argv64i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ule i32 %arg32, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I32 = icmp ule <4 x i32> %argv4i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I32 = icmp ule <8 x i32> %argv8i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I32 = icmp ule <16 x i32> %argv16i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V32I32 = icmp ule <32 x i32> %argv32i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ule i64 %arg64, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V2I64 = icmp ule <2 x i64> %argv2i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V4I64 = icmp ule <4 x i64> %argv4i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V8I64 = icmp ule <8 x i64> %argv8i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 32 for instruction: %V16I64 = icmp ule <16 x i64> %argv16i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX1-LABEL: 'cmp_int_ule'
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ule i8 %arg8, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp ule <16 x i8> %argv16i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V32I8 = icmp ule <32 x i8> %argv32i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V64I8 = icmp ule <64 x i8> %argv64i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 20 for instruction: %V128I8 = icmp ule <128 x i8> %argv128i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ule i16 %arg16, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp ule <8 x i16> %argv8i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V16I16 = icmp ule <16 x i16> %argv16i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V32I16 = icmp ule <32 x i16> %argv32i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 20 for instruction: %V64I16 = icmp ule <64 x i16> %argv64i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ule i32 %arg32, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I32 = icmp ule <4 x i32> %argv4i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V8I32 = icmp ule <8 x i32> %argv8i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V16I32 = icmp ule <16 x i32> %argv16i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 20 for instruction: %V32I32 = icmp ule <32 x i32> %argv32i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ule i64 %arg64, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V2I64 = icmp ule <2 x i64> %argv2i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 7 for instruction: %V4I64 = icmp ule <4 x i64> %argv4i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 14 for instruction: %V8I64 = icmp ule <8 x i64> %argv8i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 28 for instruction: %V16I64 = icmp ule <16 x i64> %argv16i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX2-LABEL: 'cmp_int_ule'
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ule i8 %arg8, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp ule <16 x i8> %argv16i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I8 = icmp ule <32 x i8> %argv32i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I8 = icmp ule <64 x i8> %argv64i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V128I8 = icmp ule <128 x i8> %argv128i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ule i16 %arg16, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp ule <8 x i16> %argv8i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I16 = icmp ule <16 x i16> %argv16i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I16 = icmp ule <32 x i16> %argv32i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I16 = icmp ule <64 x i16> %argv64i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ule i32 %arg32, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I32 = icmp ule <4 x i32> %argv4i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I32 = icmp ule <8 x i32> %argv8i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I32 = icmp ule <16 x i32> %argv16i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I32 = icmp ule <32 x i32> %argv32i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ule i64 %arg64, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V2I64 = icmp ule <2 x i64> %argv2i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V4I64 = icmp ule <4 x i64> %argv4i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V8I64 = icmp ule <8 x i64> %argv8i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V16I64 = icmp ule <16 x i64> %argv16i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX512F-LABEL: 'cmp_int_ule'
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ule i8 %arg8, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp ule <16 x i8> %argv16i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I8 = icmp ule <32 x i8> %argv32i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I8 = icmp ule <64 x i8> %argv64i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V128I8 = icmp ule <128 x i8> %argv128i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ule i16 %arg16, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp ule <8 x i16> %argv8i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I16 = icmp ule <16 x i16> %argv16i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I16 = icmp ule <32 x i16> %argv32i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I16 = icmp ule <64 x i16> %argv64i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ule i32 %arg32, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp ule <4 x i32> %argv4i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I32 = icmp ule <8 x i32> %argv8i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I32 = icmp ule <16 x i32> %argv16i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I32 = icmp ule <32 x i32> %argv32i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ule i64 %arg64, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp ule <2 x i64> %argv2i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I64 = icmp ule <4 x i64> %argv4i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I64 = icmp ule <8 x i64> %argv8i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I64 = icmp ule <16 x i64> %argv16i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX512BW-LABEL: 'cmp_int_ule'
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ule i8 %arg8, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp ule <16 x i8> %argv16i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V32I8 = icmp ule <32 x i8> %argv32i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V64I8 = icmp ule <64 x i8> %argv64i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V128I8 = icmp ule <128 x i8> %argv128i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ule i16 %arg16, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp ule <8 x i16> %argv8i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I16 = icmp ule <16 x i16> %argv16i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V32I16 = icmp ule <32 x i16> %argv32i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V64I16 = icmp ule <64 x i16> %argv64i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ule i32 %arg32, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp ule <4 x i32> %argv4i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I32 = icmp ule <8 x i32> %argv8i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I32 = icmp ule <16 x i32> %argv16i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I32 = icmp ule <32 x i32> %argv32i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ule i64 %arg64, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp ule <2 x i64> %argv2i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I64 = icmp ule <4 x i64> %argv4i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I64 = icmp ule <8 x i64> %argv8i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I64 = icmp ule <16 x i64> %argv16i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; XOPAVX1-LABEL: 'cmp_int_ule'
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ule i8 %arg8, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp ule <16 x i8> %argv16i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp ule <32 x i8> %argv32i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp ule <64 x i8> %argv64i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp ule <128 x i8> %argv128i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ule i16 %arg16, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp ule <8 x i16> %argv8i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp ule <16 x i16> %argv16i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp ule <32 x i16> %argv32i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp ule <64 x i16> %argv64i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ule i32 %arg32, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp ule <4 x i32> %argv4i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I32 = icmp ule <8 x i32> %argv8i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I32 = icmp ule <16 x i32> %argv16i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V32I32 = icmp ule <32 x i32> %argv32i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ule i64 %arg64, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp ule <2 x i64> %argv2i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V4I64 = icmp ule <4 x i64> %argv4i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V8I64 = icmp ule <8 x i64> %argv8i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V16I64 = icmp ule <16 x i64> %argv16i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; XOPAVX2-LABEL: 'cmp_int_ule'
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ule i8 %arg8, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp ule <16 x i8> %argv16i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I8 = icmp ule <32 x i8> %argv32i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I8 = icmp ule <64 x i8> %argv64i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V128I8 = icmp ule <128 x i8> %argv128i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ule i16 %arg16, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp ule <8 x i16> %argv8i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I16 = icmp ule <16 x i16> %argv16i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I16 = icmp ule <32 x i16> %argv32i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I16 = icmp ule <64 x i16> %argv64i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ule i32 %arg32, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp ule <4 x i32> %argv4i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I32 = icmp ule <8 x i32> %argv8i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I32 = icmp ule <16 x i32> %argv16i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I32 = icmp ule <32 x i32> %argv32i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ule i64 %arg64, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp ule <2 x i64> %argv2i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V4I64 = icmp ule <4 x i64> %argv4i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V8I64 = icmp ule <8 x i64> %argv8i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V16I64 = icmp ule <16 x i64> %argv16i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SLM-LABEL: 'cmp_int_ule'
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ule i8 %arg8, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I8 = icmp ule <16 x i8> %argv16i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp ule <32 x i8> %argv32i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp ule <64 x i8> %argv64i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp ule <128 x i8> %argv128i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ule i16 %arg16, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I16 = icmp ule <8 x i16> %argv8i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp ule <16 x i16> %argv16i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp ule <32 x i16> %argv32i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp ule <64 x i16> %argv64i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ule i32 %arg32, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I32 = icmp ule <4 x i32> %argv4i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I32 = icmp ule <8 x i32> %argv8i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I32 = icmp ule <16 x i32> %argv16i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V32I32 = icmp ule <32 x i32> %argv32i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ule i64 %arg64, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V2I64 = icmp ule <2 x i64> %argv2i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V4I64 = icmp ule <4 x i64> %argv4i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 20 for instruction: %V8I64 = icmp ule <8 x i64> %argv8i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 40 for instruction: %V16I64 = icmp ule <16 x i64> %argv16i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+  %I8 = icmp ule i8 %arg8, zeroinitializer
+  %V16I8 = icmp ule <16 x i8> %argv16i8, zeroinitializer
+  %V32I8 = icmp ule <32 x i8> %argv32i8, zeroinitializer
+  %V64I8 = icmp ule <64 x i8> %argv64i8, zeroinitializer
+  %V128I8 = icmp ule <128 x i8> %argv128i8, zeroinitializer
+
+  %I16 = icmp ule i16 %arg16, zeroinitializer
+  %V8I16  = icmp ule <8 x i16> %argv8i16, zeroinitializer
+  %V16I16 = icmp ule <16 x i16> %argv16i16, zeroinitializer
+  %V32I16 = icmp ule <32 x i16> %argv32i16, zeroinitializer
+  %V64I16 = icmp ule <64 x i16> %argv64i16, zeroinitializer
+
+  %I32 = icmp ule i32 %arg32, zeroinitializer
+  %V4I32 = icmp ule <4 x i32> %argv4i32, zeroinitializer
+  %V8I32 = icmp ule <8 x i32> %argv8i32, zeroinitializer
+  %V16I32 = icmp ule <16 x i32> %argv16i32, zeroinitializer
+  %V32I32 = icmp ule <32 x i32> %argv32i32, zeroinitializer
+
+  %I64 = icmp ule i64 %arg64, zeroinitializer
+  %V2I64 = icmp ule <2 x i64> %argv2i64, zeroinitializer
+  %V4I64 = icmp ule <4 x i64> %argv4i64, zeroinitializer
+  %V8I64 = icmp ule <8 x i64> %argv8i64, zeroinitializer
+  %V16I64 = icmp ule <16 x i64> %argv16i64, zeroinitializer
+
+  ret i32 undef
+}
+
+define i32 @cmp_int_slt(i8 %arg8, <16 x i8> %argv16i8, <32 x i8> %argv32i8, <64 x i8> %argv64i8, <128 x i8> %argv128i8, i16 %arg16, <8 x i16> %argv8i16, <16 x i16> %argv16i16, <32 x i16> %argv32i16, <64 x i16> %argv64i16, i32 %arg32, <4 x i32> %argv4i32, <8 x i32> %argv8i32, <16 x i32> %argv16i32, <32 x i32> %argv32i32, i64 %arg64, <2 x i64> %argv2i64, <4 x i64> %argv4i64, <8 x i64> %argv8i64, <16 x i64> %argv16i64) {
+; SSE2-LABEL: 'cmp_int_slt'
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp slt i8 %arg8, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp slt <16 x i8> %argv16i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I8 = icmp slt <32 x i8> %argv32i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I8 = icmp slt <64 x i8> %argv64i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V128I8 = icmp slt <128 x i8> %argv128i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp slt i16 %arg16, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp slt <8 x i16> %argv8i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I16 = icmp slt <16 x i16> %argv16i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I16 = icmp slt <32 x i16> %argv32i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I16 = icmp slt <64 x i16> %argv64i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp slt i32 %arg32, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp slt <4 x i32> %argv4i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I32 = icmp slt <8 x i32> %argv8i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I32 = icmp slt <16 x i32> %argv16i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I32 = icmp slt <32 x i32> %argv32i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp slt i64 %arg64, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V2I64 = icmp slt <2 x i64> %argv2i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V4I64 = icmp slt <4 x i64> %argv4i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 20 for instruction: %V8I64 = icmp slt <8 x i64> %argv8i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 40 for instruction: %V16I64 = icmp slt <16 x i64> %argv16i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSE3-LABEL: 'cmp_int_slt'
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp slt i8 %arg8, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp slt <16 x i8> %argv16i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I8 = icmp slt <32 x i8> %argv32i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I8 = icmp slt <64 x i8> %argv64i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V128I8 = icmp slt <128 x i8> %argv128i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp slt i16 %arg16, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp slt <8 x i16> %argv8i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I16 = icmp slt <16 x i16> %argv16i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I16 = icmp slt <32 x i16> %argv32i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I16 = icmp slt <64 x i16> %argv64i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp slt i32 %arg32, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp slt <4 x i32> %argv4i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I32 = icmp slt <8 x i32> %argv8i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I32 = icmp slt <16 x i32> %argv16i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I32 = icmp slt <32 x i32> %argv32i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp slt i64 %arg64, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V2I64 = icmp slt <2 x i64> %argv2i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V4I64 = icmp slt <4 x i64> %argv4i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 20 for instruction: %V8I64 = icmp slt <8 x i64> %argv8i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 40 for instruction: %V16I64 = icmp slt <16 x i64> %argv16i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSSE3-LABEL: 'cmp_int_slt'
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp slt i8 %arg8, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp slt <16 x i8> %argv16i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I8 = icmp slt <32 x i8> %argv32i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I8 = icmp slt <64 x i8> %argv64i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V128I8 = icmp slt <128 x i8> %argv128i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp slt i16 %arg16, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp slt <8 x i16> %argv8i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I16 = icmp slt <16 x i16> %argv16i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I16 = icmp slt <32 x i16> %argv32i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I16 = icmp slt <64 x i16> %argv64i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp slt i32 %arg32, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp slt <4 x i32> %argv4i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I32 = icmp slt <8 x i32> %argv8i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I32 = icmp slt <16 x i32> %argv16i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I32 = icmp slt <32 x i32> %argv32i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp slt i64 %arg64, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V2I64 = icmp slt <2 x i64> %argv2i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V4I64 = icmp slt <4 x i64> %argv4i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 20 for instruction: %V8I64 = icmp slt <8 x i64> %argv8i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 40 for instruction: %V16I64 = icmp slt <16 x i64> %argv16i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSE41-LABEL: 'cmp_int_slt'
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp slt i8 %arg8, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp slt <16 x i8> %argv16i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I8 = icmp slt <32 x i8> %argv32i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I8 = icmp slt <64 x i8> %argv64i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V128I8 = icmp slt <128 x i8> %argv128i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp slt i16 %arg16, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp slt <8 x i16> %argv8i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I16 = icmp slt <16 x i16> %argv16i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I16 = icmp slt <32 x i16> %argv32i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I16 = icmp slt <64 x i16> %argv64i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp slt i32 %arg32, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp slt <4 x i32> %argv4i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I32 = icmp slt <8 x i32> %argv8i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I32 = icmp slt <16 x i32> %argv16i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I32 = icmp slt <32 x i32> %argv32i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp slt i64 %arg64, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V2I64 = icmp slt <2 x i64> %argv2i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V4I64 = icmp slt <4 x i64> %argv4i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 20 for instruction: %V8I64 = icmp slt <8 x i64> %argv8i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 40 for instruction: %V16I64 = icmp slt <16 x i64> %argv16i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSE42-LABEL: 'cmp_int_slt'
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp slt i8 %arg8, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp slt <16 x i8> %argv16i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I8 = icmp slt <32 x i8> %argv32i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I8 = icmp slt <64 x i8> %argv64i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V128I8 = icmp slt <128 x i8> %argv128i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp slt i16 %arg16, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp slt <8 x i16> %argv8i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I16 = icmp slt <16 x i16> %argv16i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I16 = icmp slt <32 x i16> %argv32i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I16 = icmp slt <64 x i16> %argv64i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp slt i32 %arg32, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp slt <4 x i32> %argv4i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I32 = icmp slt <8 x i32> %argv8i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I32 = icmp slt <16 x i32> %argv16i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I32 = icmp slt <32 x i32> %argv32i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp slt i64 %arg64, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp slt <2 x i64> %argv2i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I64 = icmp slt <4 x i64> %argv4i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I64 = icmp slt <8 x i64> %argv8i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I64 = icmp slt <16 x i64> %argv16i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX1-LABEL: 'cmp_int_slt'
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp slt i8 %arg8, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp slt <16 x i8> %argv16i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp slt <32 x i8> %argv32i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp slt <64 x i8> %argv64i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp slt <128 x i8> %argv128i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp slt i16 %arg16, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp slt <8 x i16> %argv8i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp slt <16 x i16> %argv16i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp slt <32 x i16> %argv32i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp slt <64 x i16> %argv64i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp slt i32 %arg32, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp slt <4 x i32> %argv4i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I32 = icmp slt <8 x i32> %argv8i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I32 = icmp slt <16 x i32> %argv16i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V32I32 = icmp slt <32 x i32> %argv32i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp slt i64 %arg64, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp slt <2 x i64> %argv2i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V4I64 = icmp slt <4 x i64> %argv4i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V8I64 = icmp slt <8 x i64> %argv8i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V16I64 = icmp slt <16 x i64> %argv16i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX2-LABEL: 'cmp_int_slt'
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp slt i8 %arg8, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp slt <16 x i8> %argv16i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V32I8 = icmp slt <32 x i8> %argv32i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V64I8 = icmp slt <64 x i8> %argv64i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V128I8 = icmp slt <128 x i8> %argv128i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp slt i16 %arg16, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp slt <8 x i16> %argv8i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I16 = icmp slt <16 x i16> %argv16i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I16 = icmp slt <32 x i16> %argv32i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I16 = icmp slt <64 x i16> %argv64i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp slt i32 %arg32, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp slt <4 x i32> %argv4i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I32 = icmp slt <8 x i32> %argv8i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I32 = icmp slt <16 x i32> %argv16i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I32 = icmp slt <32 x i32> %argv32i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp slt i64 %arg64, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp slt <2 x i64> %argv2i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I64 = icmp slt <4 x i64> %argv4i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I64 = icmp slt <8 x i64> %argv8i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I64 = icmp slt <16 x i64> %argv16i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX512F-LABEL: 'cmp_int_slt'
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp slt i8 %arg8, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp slt <16 x i8> %argv16i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V32I8 = icmp slt <32 x i8> %argv32i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V64I8 = icmp slt <64 x i8> %argv64i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V128I8 = icmp slt <128 x i8> %argv128i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp slt i16 %arg16, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp slt <8 x i16> %argv8i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I16 = icmp slt <16 x i16> %argv16i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V32I16 = icmp slt <32 x i16> %argv32i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V64I16 = icmp slt <64 x i16> %argv64i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp slt i32 %arg32, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp slt <4 x i32> %argv4i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I32 = icmp slt <8 x i32> %argv8i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I32 = icmp slt <16 x i32> %argv16i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I32 = icmp slt <32 x i32> %argv32i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp slt i64 %arg64, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp slt <2 x i64> %argv2i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I64 = icmp slt <4 x i64> %argv4i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I64 = icmp slt <8 x i64> %argv8i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I64 = icmp slt <16 x i64> %argv16i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX512BW-LABEL: 'cmp_int_slt'
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp slt i8 %arg8, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp slt <16 x i8> %argv16i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V32I8 = icmp slt <32 x i8> %argv32i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V64I8 = icmp slt <64 x i8> %argv64i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V128I8 = icmp slt <128 x i8> %argv128i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp slt i16 %arg16, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp slt <8 x i16> %argv8i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I16 = icmp slt <16 x i16> %argv16i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V32I16 = icmp slt <32 x i16> %argv32i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V64I16 = icmp slt <64 x i16> %argv64i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp slt i32 %arg32, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp slt <4 x i32> %argv4i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I32 = icmp slt <8 x i32> %argv8i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I32 = icmp slt <16 x i32> %argv16i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I32 = icmp slt <32 x i32> %argv32i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp slt i64 %arg64, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp slt <2 x i64> %argv2i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I64 = icmp slt <4 x i64> %argv4i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I64 = icmp slt <8 x i64> %argv8i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I64 = icmp slt <16 x i64> %argv16i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; XOPAVX1-LABEL: 'cmp_int_slt'
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp slt i8 %arg8, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp slt <16 x i8> %argv16i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp slt <32 x i8> %argv32i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp slt <64 x i8> %argv64i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp slt <128 x i8> %argv128i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp slt i16 %arg16, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp slt <8 x i16> %argv8i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp slt <16 x i16> %argv16i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp slt <32 x i16> %argv32i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp slt <64 x i16> %argv64i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp slt i32 %arg32, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp slt <4 x i32> %argv4i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I32 = icmp slt <8 x i32> %argv8i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I32 = icmp slt <16 x i32> %argv16i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V32I32 = icmp slt <32 x i32> %argv32i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp slt i64 %arg64, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp slt <2 x i64> %argv2i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V4I64 = icmp slt <4 x i64> %argv4i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V8I64 = icmp slt <8 x i64> %argv8i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V16I64 = icmp slt <16 x i64> %argv16i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; XOPAVX2-LABEL: 'cmp_int_slt'
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp slt i8 %arg8, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp slt <16 x i8> %argv16i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V32I8 = icmp slt <32 x i8> %argv32i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V64I8 = icmp slt <64 x i8> %argv64i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V128I8 = icmp slt <128 x i8> %argv128i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp slt i16 %arg16, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp slt <8 x i16> %argv8i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I16 = icmp slt <16 x i16> %argv16i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I16 = icmp slt <32 x i16> %argv32i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I16 = icmp slt <64 x i16> %argv64i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp slt i32 %arg32, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp slt <4 x i32> %argv4i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I32 = icmp slt <8 x i32> %argv8i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I32 = icmp slt <16 x i32> %argv16i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I32 = icmp slt <32 x i32> %argv32i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp slt i64 %arg64, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp slt <2 x i64> %argv2i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I64 = icmp slt <4 x i64> %argv4i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I64 = icmp slt <8 x i64> %argv8i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I64 = icmp slt <16 x i64> %argv16i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SLM-LABEL: 'cmp_int_slt'
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp slt i8 %arg8, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp slt <16 x i8> %argv16i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I8 = icmp slt <32 x i8> %argv32i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V64I8 = icmp slt <64 x i8> %argv64i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V128I8 = icmp slt <128 x i8> %argv128i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp slt i16 %arg16, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp slt <8 x i16> %argv8i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I16 = icmp slt <16 x i16> %argv16i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I16 = icmp slt <32 x i16> %argv32i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I16 = icmp slt <64 x i16> %argv64i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp slt i32 %arg32, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp slt <4 x i32> %argv4i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V8I32 = icmp slt <8 x i32> %argv8i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I32 = icmp slt <16 x i32> %argv16i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I32 = icmp slt <32 x i32> %argv32i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp slt i64 %arg64, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V2I64 = icmp slt <2 x i64> %argv2i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V4I64 = icmp slt <4 x i64> %argv4i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V8I64 = icmp slt <8 x i64> %argv8i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V16I64 = icmp slt <16 x i64> %argv16i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+  %I8 = icmp slt i8 %arg8, zeroinitializer
+  %V16I8 = icmp slt <16 x i8> %argv16i8, zeroinitializer
+  %V32I8 = icmp slt <32 x i8> %argv32i8, zeroinitializer
+  %V64I8 = icmp slt <64 x i8> %argv64i8, zeroinitializer
+  %V128I8 = icmp slt <128 x i8> %argv128i8, zeroinitializer
+
+  %I16 = icmp slt i16 %arg16, zeroinitializer
+  %V8I16  = icmp slt <8 x i16> %argv8i16, zeroinitializer
+  %V16I16 = icmp slt <16 x i16> %argv16i16, zeroinitializer
+  %V32I16 = icmp slt <32 x i16> %argv32i16, zeroinitializer
+  %V64I16 = icmp slt <64 x i16> %argv64i16, zeroinitializer
+
+  %I32 = icmp slt i32 %arg32, zeroinitializer
+  %V4I32 = icmp slt <4 x i32> %argv4i32, zeroinitializer
+  %V8I32 = icmp slt <8 x i32> %argv8i32, zeroinitializer
+  %V16I32 = icmp slt <16 x i32> %argv16i32, zeroinitializer
+  %V32I32 = icmp slt <32 x i32> %argv32i32, zeroinitializer
+
+  %I64 = icmp slt i64 %arg64, zeroinitializer
+  %V2I64 = icmp slt <2 x i64> %argv2i64, zeroinitializer
+  %V4I64 = icmp slt <4 x i64> %argv4i64, zeroinitializer
+  %V8I64 = icmp slt <8 x i64> %argv8i64, zeroinitializer
+  %V16I64 = icmp slt <16 x i64> %argv16i64, zeroinitializer
+
+  ret i32 undef
+}
+
+define i32 @cmp_int_ult(i8 %arg8, <16 x i8> %argv16i8, <32 x i8> %argv32i8, <64 x i8> %argv64i8, <128 x i8> %argv128i8, i16 %arg16, <8 x i16> %argv8i16, <16 x i16> %argv16i16, <32 x i16> %argv32i16, <64 x i16> %argv64i16, i32 %arg32, <4 x i32> %argv4i32, <8 x i32> %argv8i32, <16 x i32> %argv16i32, <32 x i32> %argv32i32, i64 %arg64, <2 x i64> %argv2i64, <4 x i64> %argv4i64, <8 x i64> %argv8i64, <16 x i64> %argv16i64) {
+; SSE2-LABEL: 'cmp_int_ult'
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ult i8 %arg8, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V16I8 = icmp ult <16 x i8> %argv16i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V32I8 = icmp ult <32 x i8> %argv32i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V64I8 = icmp ult <64 x i8> %argv64i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V128I8 = icmp ult <128 x i8> %argv128i8, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ult i16 %arg16, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V8I16 = icmp ult <8 x i16> %argv8i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V16I16 = icmp ult <16 x i16> %argv16i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V32I16 = icmp ult <32 x i16> %argv32i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V64I16 = icmp ult <64 x i16> %argv64i16, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ult i32 %arg32, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V4I32 = icmp ult <4 x i32> %argv4i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V8I32 = icmp ult <8 x i32> %argv8i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V16I32 = icmp ult <16 x i32> %argv16i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V32I32 = icmp ult <32 x i32> %argv32i32, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ult i64 %arg64, 0
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 7 for instruction: %V2I64 = icmp ult <2 x i64> %argv2i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 14 for instruction: %V4I64 = icmp ult <4 x i64> %argv4i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 28 for instruction: %V8I64 = icmp ult <8 x i64> %argv8i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 56 for instruction: %V16I64 = icmp ult <16 x i64> %argv16i64, zeroinitializer
+; SSE2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSE3-LABEL: 'cmp_int_ult'
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ult i8 %arg8, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V16I8 = icmp ult <16 x i8> %argv16i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V32I8 = icmp ult <32 x i8> %argv32i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V64I8 = icmp ult <64 x i8> %argv64i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V128I8 = icmp ult <128 x i8> %argv128i8, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ult i16 %arg16, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V8I16 = icmp ult <8 x i16> %argv8i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V16I16 = icmp ult <16 x i16> %argv16i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V32I16 = icmp ult <32 x i16> %argv32i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V64I16 = icmp ult <64 x i16> %argv64i16, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ult i32 %arg32, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V4I32 = icmp ult <4 x i32> %argv4i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V8I32 = icmp ult <8 x i32> %argv8i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V16I32 = icmp ult <16 x i32> %argv16i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V32I32 = icmp ult <32 x i32> %argv32i32, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ult i64 %arg64, 0
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 7 for instruction: %V2I64 = icmp ult <2 x i64> %argv2i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 14 for instruction: %V4I64 = icmp ult <4 x i64> %argv4i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 28 for instruction: %V8I64 = icmp ult <8 x i64> %argv8i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 56 for instruction: %V16I64 = icmp ult <16 x i64> %argv16i64, zeroinitializer
+; SSE3-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSSE3-LABEL: 'cmp_int_ult'
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ult i8 %arg8, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V16I8 = icmp ult <16 x i8> %argv16i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V32I8 = icmp ult <32 x i8> %argv32i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V64I8 = icmp ult <64 x i8> %argv64i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V128I8 = icmp ult <128 x i8> %argv128i8, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ult i16 %arg16, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V8I16 = icmp ult <8 x i16> %argv8i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V16I16 = icmp ult <16 x i16> %argv16i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V32I16 = icmp ult <32 x i16> %argv32i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V64I16 = icmp ult <64 x i16> %argv64i16, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ult i32 %arg32, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V4I32 = icmp ult <4 x i32> %argv4i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V8I32 = icmp ult <8 x i32> %argv8i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V16I32 = icmp ult <16 x i32> %argv16i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V32I32 = icmp ult <32 x i32> %argv32i32, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ult i64 %arg64, 0
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 7 for instruction: %V2I64 = icmp ult <2 x i64> %argv2i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 14 for instruction: %V4I64 = icmp ult <4 x i64> %argv4i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 28 for instruction: %V8I64 = icmp ult <8 x i64> %argv8i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 56 for instruction: %V16I64 = icmp ult <16 x i64> %argv16i64, zeroinitializer
+; SSSE3-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSE41-LABEL: 'cmp_int_ult'
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ult i8 %arg8, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V16I8 = icmp ult <16 x i8> %argv16i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V32I8 = icmp ult <32 x i8> %argv32i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V64I8 = icmp ult <64 x i8> %argv64i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V128I8 = icmp ult <128 x i8> %argv128i8, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ult i16 %arg16, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V8I16 = icmp ult <8 x i16> %argv8i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V16I16 = icmp ult <16 x i16> %argv16i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V32I16 = icmp ult <32 x i16> %argv32i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V64I16 = icmp ult <64 x i16> %argv64i16, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ult i32 %arg32, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V4I32 = icmp ult <4 x i32> %argv4i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V8I32 = icmp ult <8 x i32> %argv8i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V16I32 = icmp ult <16 x i32> %argv16i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V32I32 = icmp ult <32 x i32> %argv32i32, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ult i64 %arg64, 0
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 7 for instruction: %V2I64 = icmp ult <2 x i64> %argv2i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 14 for instruction: %V4I64 = icmp ult <4 x i64> %argv4i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 28 for instruction: %V8I64 = icmp ult <8 x i64> %argv8i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 56 for instruction: %V16I64 = icmp ult <16 x i64> %argv16i64, zeroinitializer
+; SSE41-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SSE42-LABEL: 'cmp_int_ult'
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ult i8 %arg8, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V16I8 = icmp ult <16 x i8> %argv16i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V32I8 = icmp ult <32 x i8> %argv32i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V64I8 = icmp ult <64 x i8> %argv64i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V128I8 = icmp ult <128 x i8> %argv128i8, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ult i16 %arg16, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V8I16 = icmp ult <8 x i16> %argv8i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V16I16 = icmp ult <16 x i16> %argv16i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V32I16 = icmp ult <32 x i16> %argv32i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V64I16 = icmp ult <64 x i16> %argv64i16, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ult i32 %arg32, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V4I32 = icmp ult <4 x i32> %argv4i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V8I32 = icmp ult <8 x i32> %argv8i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V16I32 = icmp ult <16 x i32> %argv16i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V32I32 = icmp ult <32 x i32> %argv32i32, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ult i64 %arg64, 0
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V2I64 = icmp ult <2 x i64> %argv2i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V4I64 = icmp ult <4 x i64> %argv4i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V8I64 = icmp ult <8 x i64> %argv8i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V16I64 = icmp ult <16 x i64> %argv16i64, zeroinitializer
+; SSE42-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX1-LABEL: 'cmp_int_ult'
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ult i8 %arg8, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V16I8 = icmp ult <16 x i8> %argv16i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V32I8 = icmp ult <32 x i8> %argv32i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V64I8 = icmp ult <64 x i8> %argv64i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V128I8 = icmp ult <128 x i8> %argv128i8, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ult i16 %arg16, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V8I16 = icmp ult <8 x i16> %argv8i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V16I16 = icmp ult <16 x i16> %argv16i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V32I16 = icmp ult <32 x i16> %argv32i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V64I16 = icmp ult <64 x i16> %argv64i16, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ult i32 %arg32, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V4I32 = icmp ult <4 x i32> %argv4i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V8I32 = icmp ult <8 x i32> %argv8i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V16I32 = icmp ult <16 x i32> %argv16i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V32I32 = icmp ult <32 x i32> %argv32i32, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ult i64 %arg64, 0
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V2I64 = icmp ult <2 x i64> %argv2i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V4I64 = icmp ult <4 x i64> %argv4i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V8I64 = icmp ult <8 x i64> %argv8i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V16I64 = icmp ult <16 x i64> %argv16i64, zeroinitializer
+; AVX1-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX2-LABEL: 'cmp_int_ult'
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ult i8 %arg8, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V16I8 = icmp ult <16 x i8> %argv16i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V32I8 = icmp ult <32 x i8> %argv32i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V64I8 = icmp ult <64 x i8> %argv64i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V128I8 = icmp ult <128 x i8> %argv128i8, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ult i16 %arg16, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V8I16 = icmp ult <8 x i16> %argv8i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V16I16 = icmp ult <16 x i16> %argv16i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V32I16 = icmp ult <32 x i16> %argv32i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V64I16 = icmp ult <64 x i16> %argv64i16, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ult i32 %arg32, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V4I32 = icmp ult <4 x i32> %argv4i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V8I32 = icmp ult <8 x i32> %argv8i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V16I32 = icmp ult <16 x i32> %argv16i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V32I32 = icmp ult <32 x i32> %argv32i32, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ult i64 %arg64, 0
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V2I64 = icmp ult <2 x i64> %argv2i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V4I64 = icmp ult <4 x i64> %argv4i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V8I64 = icmp ult <8 x i64> %argv8i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V16I64 = icmp ult <16 x i64> %argv16i64, zeroinitializer
+; AVX2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX512F-LABEL: 'cmp_int_ult'
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ult i8 %arg8, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V16I8 = icmp ult <16 x i8> %argv16i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V32I8 = icmp ult <32 x i8> %argv32i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V64I8 = icmp ult <64 x i8> %argv64i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V128I8 = icmp ult <128 x i8> %argv128i8, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ult i16 %arg16, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V8I16 = icmp ult <8 x i16> %argv8i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V16I16 = icmp ult <16 x i16> %argv16i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V32I16 = icmp ult <32 x i16> %argv32i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %V64I16 = icmp ult <64 x i16> %argv64i16, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ult i32 %arg32, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp ult <4 x i32> %argv4i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I32 = icmp ult <8 x i32> %argv8i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I32 = icmp ult <16 x i32> %argv16i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I32 = icmp ult <32 x i32> %argv32i32, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ult i64 %arg64, 0
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp ult <2 x i64> %argv2i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I64 = icmp ult <4 x i64> %argv4i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I64 = icmp ult <8 x i64> %argv8i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I64 = icmp ult <16 x i64> %argv16i64, zeroinitializer
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; AVX512BW-LABEL: 'cmp_int_ult'
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ult i8 %arg8, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp ult <16 x i8> %argv16i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V32I8 = icmp ult <32 x i8> %argv32i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V64I8 = icmp ult <64 x i8> %argv64i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V128I8 = icmp ult <128 x i8> %argv128i8, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ult i16 %arg16, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp ult <8 x i16> %argv8i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I16 = icmp ult <16 x i16> %argv16i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V32I16 = icmp ult <32 x i16> %argv32i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V64I16 = icmp ult <64 x i16> %argv64i16, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ult i32 %arg32, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp ult <4 x i32> %argv4i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I32 = icmp ult <8 x i32> %argv8i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I32 = icmp ult <16 x i32> %argv16i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V32I32 = icmp ult <32 x i32> %argv32i32, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ult i64 %arg64, 0
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp ult <2 x i64> %argv2i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I64 = icmp ult <4 x i64> %argv4i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I64 = icmp ult <8 x i64> %argv8i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V16I64 = icmp ult <16 x i64> %argv16i64, zeroinitializer
+; AVX512BW-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; XOPAVX1-LABEL: 'cmp_int_ult'
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ult i8 %arg8, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp ult <16 x i8> %argv16i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V32I8 = icmp ult <32 x i8> %argv32i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V64I8 = icmp ult <64 x i8> %argv64i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V128I8 = icmp ult <128 x i8> %argv128i8, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ult i16 %arg16, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp ult <8 x i16> %argv8i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V16I16 = icmp ult <16 x i16> %argv16i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V32I16 = icmp ult <32 x i16> %argv32i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V64I16 = icmp ult <64 x i16> %argv64i16, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ult i32 %arg32, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp ult <4 x i32> %argv4i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V8I32 = icmp ult <8 x i32> %argv8i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V16I32 = icmp ult <16 x i32> %argv16i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V32I32 = icmp ult <32 x i32> %argv32i32, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ult i64 %arg64, 0
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp ult <2 x i64> %argv2i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V4I64 = icmp ult <4 x i64> %argv4i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V8I64 = icmp ult <8 x i64> %argv8i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V16I64 = icmp ult <16 x i64> %argv16i64, zeroinitializer
+; XOPAVX1-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; XOPAVX2-LABEL: 'cmp_int_ult'
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ult i8 %arg8, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = icmp ult <16 x i8> %argv16i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V32I8 = icmp ult <32 x i8> %argv32i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V64I8 = icmp ult <64 x i8> %argv64i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V128I8 = icmp ult <128 x i8> %argv128i8, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ult i16 %arg16, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = icmp ult <8 x i16> %argv8i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V16I16 = icmp ult <16 x i16> %argv16i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V32I16 = icmp ult <32 x i16> %argv32i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V64I16 = icmp ult <64 x i16> %argv64i16, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ult i32 %arg32, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = icmp ult <4 x i32> %argv4i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V8I32 = icmp ult <8 x i32> %argv8i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V16I32 = icmp ult <16 x i32> %argv16i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V32I32 = icmp ult <32 x i32> %argv32i32, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ult i64 %arg64, 0
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V2I64 = icmp ult <2 x i64> %argv2i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V4I64 = icmp ult <4 x i64> %argv4i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V8I64 = icmp ult <8 x i64> %argv8i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V16I64 = icmp ult <16 x i64> %argv16i64, zeroinitializer
+; XOPAVX2-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+; SLM-LABEL: 'cmp_int_ult'
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = icmp ult i8 %arg8, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V16I8 = icmp ult <16 x i8> %argv16i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V32I8 = icmp ult <32 x i8> %argv32i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V64I8 = icmp ult <64 x i8> %argv64i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V128I8 = icmp ult <128 x i8> %argv128i8, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = icmp ult i16 %arg16, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V8I16 = icmp ult <8 x i16> %argv8i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V16I16 = icmp ult <16 x i16> %argv16i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V32I16 = icmp ult <32 x i16> %argv32i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V64I16 = icmp ult <64 x i16> %argv64i16, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = icmp ult i32 %arg32, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V4I32 = icmp ult <4 x i32> %argv4i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V8I32 = icmp ult <8 x i32> %argv8i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V16I32 = icmp ult <16 x i32> %argv16i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V32I32 = icmp ult <32 x i32> %argv32i32, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = icmp ult i64 %arg64, 0
+; SLM-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %V2I64 = icmp ult <2 x i64> %argv2i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %V4I64 = icmp ult <4 x i64> %argv4i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %V8I64 = icmp ult <8 x i64> %argv8i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 32 for instruction: %V16I64 = icmp ult <16 x i64> %argv16i64, zeroinitializer
+; SLM-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
+  %I8 = icmp ult i8 %arg8, zeroinitializer
+  %V16I8 = icmp ult <16 x i8> %argv16i8, zeroinitializer
+  %V32I8 = icmp ult <32 x i8> %argv32i8, zeroinitializer
+  %V64I8 = icmp ult <64 x i8> %argv64i8, zeroinitializer
+  %V128I8 = icmp ult <128 x i8> %argv128i8, zeroinitializer
+
+  %I16 = icmp ult i16 %arg16, zeroinitializer
+  %V8I16  = icmp ult <8 x i16> %argv8i16, zeroinitializer
+  %V16I16 = icmp ult <16 x i16> %argv16i16, zeroinitializer
+  %V32I16 = icmp ult <32 x i16> %argv32i16, zeroinitializer
+  %V64I16 = icmp ult <64 x i16> %argv64i16, zeroinitializer
+
+  %I32 = icmp ult i32 %arg32, zeroinitializer
+  %V4I32 = icmp ult <4 x i32> %argv4i32, zeroinitializer
+  %V8I32 = icmp ult <8 x i32> %argv8i32, zeroinitializer
+  %V16I32 = icmp ult <16 x i32> %argv16i32, zeroinitializer
+  %V32I32 = icmp ult <32 x i32> %argv32i32, zeroinitializer
+
+  %I64 = icmp ult i64 %arg64, zeroinitializer
+  %V2I64 = icmp ult <2 x i64> %argv2i64, zeroinitializer
+  %V4I64 = icmp ult <4 x i64> %argv4i64, zeroinitializer
+  %V8I64 = icmp ult <8 x i64> %argv8i64, zeroinitializer
+  %V16I64 = icmp ult <16 x i64> %argv16i64, zeroinitializer
+
+  ret i32 undef
+}
+


        
