[X86][Skylake][KNL] Updating code gen regression test to use the KNL and SKYLAKE prefixes. NFC,

[gadi.haber]

Summary

NFC.
Updating the code gen regression test bmi2-schedule.ll to use the KNL and SKYLAKE prefixes for the run commands that use the knl and Skylake mcpu options.
The fix is in preparation for a large patch of adding all SKL scheduling information.

[zvi]

Can you please move the NFC to the Title:
[X86][Skylake][KNL] Updating code gen regression test to use the KNL and SKYLAKE prefixes. NFC.

test/CodeGen/X86/bmi2-schedule.ll

2–3

Can you please add a common --check-prefix to allow common CHECK:'s?

[RKSimon]

LGTM

test/CodeGen/X86/bmi2-schedule.ll

2–3

Given that these are per-target-cpu scheduler tests these are highly unlikely to be used, but won't do any harm. You could add a common prefix for HW/SKX/KNL as well if its important.

[gadi.haber]

Updated diff to include the --check-prefix=COMMON flag in the run commands.

[RKSimon]

This was committed at rL313138

test/CodeGen/X86/bmi2-schedule.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=x86-64 -mattr=+bmi2 | FileCheck %s --check-prefix=CHECK --check-prefix=GENERIC
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=haswell | FileCheck %s --check-prefix=CHECK --check-prefix=HASWELL
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=skylake | FileCheck %s --check-prefix=CHECK --check-prefix=HASWELL
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=knl     | FileCheck %s --check-prefix=CHECK --check-prefix=HASWELL
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=znver1  | FileCheck %s --check-prefix=CHECK --check-prefix=ZNVER1
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=haswell | FileCheck %s --check-prefix=CHECK --check-prefix=COMMON --check-prefix=HASWELL
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=skylake | FileCheck %s --check-prefix=CHECK --check-prefix=COMMON --check-prefix=SKYLAKE
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=knl     | FileCheck %s --check-prefix=CHECK --check-prefix=COMMON --check-prefix=KNL
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=znver1  | FileCheck %s --check-prefix=CHECK --check-prefix=COMMON --check-prefix=ZNVER1
 
 define i32 @test_bzhi_i32(i32 %a0, i32 %a1, i32 *%a2) {
 ; GENERIC-LABEL: test_bzhi_i32:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    bzhil %edi, (%rdx), %ecx
 ; GENERIC-NEXT:    bzhil %edi, %esi, %eax
 ; GENERIC-NEXT:    addl %ecx, %eax # sched: [1:0.33]
 ; GENERIC-NEXT:    retq # sched: [1:1.00]
 ;
 ; HASWELL-LABEL: test_bzhi_i32:
 ; HASWELL:       # BB#0:
 ; HASWELL-NEXT:    bzhil %edi, (%rdx), %ecx # sched: [1:0.50]
 ; HASWELL-NEXT:    bzhil %edi, %esi, %eax # sched: [1:0.50]
 ; HASWELL-NEXT:    addl %ecx, %eax # sched: [1:0.25]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_bzhi_i32:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    bzhil %edi, (%rdx), %ecx # sched: [1:0.50]
+; SKYLAKE-NEXT:    bzhil %edi, %esi, %eax # sched: [1:0.50]
+; SKYLAKE-NEXT:    addl %ecx, %eax # sched: [1:0.25]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
+; KNL-LABEL: test_bzhi_i32:
+; KNL:       # BB#0:
+; KNL-NEXT:    bzhil %edi, (%rdx), %ecx # sched: [1:0.50]
+; KNL-NEXT:    bzhil %edi, %esi, %eax # sched: [1:0.50]
+; KNL-NEXT:    addl %ecx, %eax # sched: [1:0.25]
+; KNL-NEXT:    retq # sched: [2:1.00]
+;
 ; ZNVER1-LABEL: test_bzhi_i32:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    bzhil %edi, (%rdx), %ecx # sched: [5:0.50]
 ; ZNVER1-NEXT:    bzhil %edi, %esi, %eax # sched: [1:0.25]
 ; ZNVER1-NEXT:    addl %ecx, %eax # sched: [1:0.25]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = load i32, i32 *%a2
   %2 = tail call i32 @llvm.x86.bmi.bzhi.32(i32 %1, i32 %a0)

 ;
 ; HASWELL-LABEL: test_bzhi_i64:
 ; HASWELL:       # BB#0:
 ; HASWELL-NEXT:    bzhiq %rdi, (%rdx), %rcx # sched: [1:0.50]
 ; HASWELL-NEXT:    bzhiq %rdi, %rsi, %rax # sched: [1:0.50]
 ; HASWELL-NEXT:    addq %rcx, %rax # sched: [1:0.25]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_bzhi_i64:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    bzhiq %rdi, (%rdx), %rcx # sched: [1:0.50]
+; SKYLAKE-NEXT:    bzhiq %rdi, %rsi, %rax # sched: [1:0.50]
+; SKYLAKE-NEXT:    addq %rcx, %rax # sched: [1:0.25]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
+; KNL-LABEL: test_bzhi_i64:
+; KNL:       # BB#0:
+; KNL-NEXT:    bzhiq %rdi, (%rdx), %rcx # sched: [1:0.50]
+; KNL-NEXT:    bzhiq %rdi, %rsi, %rax # sched: [1:0.50]
+; KNL-NEXT:    addq %rcx, %rax # sched: [1:0.25]
+; KNL-NEXT:    retq # sched: [2:1.00]
+;
 ; ZNVER1-LABEL: test_bzhi_i64:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    bzhiq %rdi, (%rdx), %rcx # sched: [5:0.50]
 ; ZNVER1-NEXT:    bzhiq %rdi, %rsi, %rax # sched: [1:0.25]
 ; ZNVER1-NEXT:    addq %rcx, %rax # sched: [1:0.25]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = load i64, i64 *%a2
   %2 = tail call i64 @llvm.x86.bmi.bzhi.64(i64 %1, i64 %a0)

 ; HASWELL:       # BB#0:
 ; HASWELL-NEXT:    movq %rdx, %rax # sched: [1:0.25]
 ; HASWELL-NEXT:    movq %rdi, %rdx # sched: [1:0.25]
 ; HASWELL-NEXT:    mulxq %rsi, %rsi, %rcx # sched: [4:1.00]
 ; HASWELL-NEXT:    mulxq (%rax), %rdx, %rax # sched: [4:1.00]
 ; HASWELL-NEXT:    orq %rcx, %rax # sched: [1:0.25]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_mulx_i64:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    movq %rdx, %rax # sched: [1:0.25]
+; SKYLAKE-NEXT:    movq %rdi, %rdx # sched: [1:0.25]
+; SKYLAKE-NEXT:    mulxq %rsi, %rsi, %rcx # sched: [4:1.00]
+; SKYLAKE-NEXT:    mulxq (%rax), %rdx, %rax # sched: [4:1.00]
+; SKYLAKE-NEXT:    orq %rcx, %rax # sched: [1:0.25]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
+; KNL-LABEL: test_mulx_i64:
+; KNL:       # BB#0:
+; KNL-NEXT:    movq %rdx, %rax # sched: [1:0.25]
+; KNL-NEXT:    movq %rdi, %rdx # sched: [1:0.25]
+; KNL-NEXT:    mulxq %rsi, %rsi, %rcx # sched: [4:1.00]
+; KNL-NEXT:    mulxq (%rax), %rdx, %rax # sched: [4:1.00]
+; KNL-NEXT:    orq %rcx, %rax # sched: [1:0.25]
+; KNL-NEXT:    retq # sched: [2:1.00]
+;
 ; ZNVER1-LABEL: test_mulx_i64:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    movq %rdx, %rax # sched: [1:0.25]
 ; ZNVER1-NEXT:    movq %rdi, %rdx # sched: [1:0.25]
 ; ZNVER1-NEXT:    mulxq %rsi, %rsi, %rcx # sched: [3:1.00]
 ; ZNVER1-NEXT:    mulxq (%rax), %rdx, %rax # sched: [8:1.00]
 ; ZNVER1-NEXT:    orq %rcx, %rax # sched: [1:0.25]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]

 ;
 ; HASWELL-LABEL: test_pdep_i32:
 ; HASWELL:       # BB#0:
 ; HASWELL-NEXT:    pdepl (%rdx), %edi, %ecx # sched: [3:1.00]
 ; HASWELL-NEXT:    pdepl %esi, %edi, %eax # sched: [3:1.00]
 ; HASWELL-NEXT:    addl %ecx, %eax # sched: [1:0.25]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pdep_i32:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    pdepl (%rdx), %edi, %ecx # sched: [3:1.00]
+; SKYLAKE-NEXT:    pdepl %esi, %edi, %eax # sched: [3:1.00]
+; SKYLAKE-NEXT:    addl %ecx, %eax # sched: [1:0.25]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
+; KNL-LABEL: test_pdep_i32:
+; KNL:       # BB#0:
+; KNL-NEXT:    pdepl (%rdx), %edi, %ecx # sched: [3:1.00]
+; KNL-NEXT:    pdepl %esi, %edi, %eax # sched: [3:1.00]
+; KNL-NEXT:    addl %ecx, %eax # sched: [1:0.25]
+; KNL-NEXT:    retq # sched: [2:1.00]
+;
 ; ZNVER1-LABEL: test_pdep_i32:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    pdepl (%rdx), %edi, %ecx # sched: [100:?]
 ; ZNVER1-NEXT:    pdepl %esi, %edi, %eax # sched: [100:?]
 ; ZNVER1-NEXT:    addl %ecx, %eax # sched: [1:0.25]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = load i32, i32 *%a2
   %2 = tail call i32 @llvm.x86.bmi.pdep.32(i32 %a0, i32 %1)

 ;
 ; HASWELL-LABEL: test_pdep_i64:
 ; HASWELL:       # BB#0:
 ; HASWELL-NEXT:    pdepq (%rdx), %rdi, %rcx # sched: [3:1.00]
 ; HASWELL-NEXT:    pdepq %rsi, %rdi, %rax # sched: [3:1.00]
 ; HASWELL-NEXT:    addq %rcx, %rax # sched: [1:0.25]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pdep_i64:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    pdepq (%rdx), %rdi, %rcx # sched: [3:1.00]
+; SKYLAKE-NEXT:    pdepq %rsi, %rdi, %rax # sched: [3:1.00]
+; SKYLAKE-NEXT:    addq %rcx, %rax # sched: [1:0.25]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
+; KNL-LABEL: test_pdep_i64:
+; KNL:       # BB#0:
+; KNL-NEXT:    pdepq (%rdx), %rdi, %rcx # sched: [3:1.00]
+; KNL-NEXT:    pdepq %rsi, %rdi, %rax # sched: [3:1.00]
+; KNL-NEXT:    addq %rcx, %rax # sched: [1:0.25]
+; KNL-NEXT:    retq # sched: [2:1.00]
+;
 ; ZNVER1-LABEL: test_pdep_i64:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    pdepq (%rdx), %rdi, %rcx # sched: [100:?]
 ; ZNVER1-NEXT:    pdepq %rsi, %rdi, %rax # sched: [100:?]
 ; ZNVER1-NEXT:    addq %rcx, %rax # sched: [1:0.25]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = load i64, i64 *%a2
   %2 = tail call i64 @llvm.x86.bmi.pdep.64(i64 %a0, i64 %1)

 ;
 ; HASWELL-LABEL: test_pext_i32:
 ; HASWELL:       # BB#0:
 ; HASWELL-NEXT:    pextl (%rdx), %edi, %ecx # sched: [3:1.00]
 ; HASWELL-NEXT:    pextl %esi, %edi, %eax # sched: [3:1.00]
 ; HASWELL-NEXT:    addl %ecx, %eax # sched: [1:0.25]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pext_i32:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    pextl (%rdx), %edi, %ecx # sched: [3:1.00]
+; SKYLAKE-NEXT:    pextl %esi, %edi, %eax # sched: [3:1.00]
+; SKYLAKE-NEXT:    addl %ecx, %eax # sched: [1:0.25]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
+; KNL-LABEL: test_pext_i32:
+; KNL:       # BB#0:
+; KNL-NEXT:    pextl (%rdx), %edi, %ecx # sched: [3:1.00]
+; KNL-NEXT:    pextl %esi, %edi, %eax # sched: [3:1.00]
+; KNL-NEXT:    addl %ecx, %eax # sched: [1:0.25]
+; KNL-NEXT:    retq # sched: [2:1.00]
+;
 ; ZNVER1-LABEL: test_pext_i32:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    pextl (%rdx), %edi, %ecx # sched: [100:?]
 ; ZNVER1-NEXT:    pextl %esi, %edi, %eax # sched: [100:?]
 ; ZNVER1-NEXT:    addl %ecx, %eax # sched: [1:0.25]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = load i32, i32 *%a2
   %2 = tail call i32 @llvm.x86.bmi.pext.32(i32 %a0, i32 %1)

 ;
 ; HASWELL-LABEL: test_pext_i64:
 ; HASWELL:       # BB#0:
 ; HASWELL-NEXT:    pextq (%rdx), %rdi, %rcx # sched: [3:1.00]
 ; HASWELL-NEXT:    pextq %rsi, %rdi, %rax # sched: [3:1.00]
 ; HASWELL-NEXT:    addq %rcx, %rax # sched: [1:0.25]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_pext_i64:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    pextq (%rdx), %rdi, %rcx # sched: [3:1.00]
+; SKYLAKE-NEXT:    pextq %rsi, %rdi, %rax # sched: [3:1.00]
+; SKYLAKE-NEXT:    addq %rcx, %rax # sched: [1:0.25]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
+; KNL-LABEL: test_pext_i64:
+; KNL:       # BB#0:
+; KNL-NEXT:    pextq (%rdx), %rdi, %rcx # sched: [3:1.00]
+; KNL-NEXT:    pextq %rsi, %rdi, %rax # sched: [3:1.00]
+; KNL-NEXT:    addq %rcx, %rax # sched: [1:0.25]
+; KNL-NEXT:    retq # sched: [2:1.00]
+;
 ; ZNVER1-LABEL: test_pext_i64:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    pextq (%rdx), %rdi, %rcx # sched: [100:?]
 ; ZNVER1-NEXT:    pextq %rsi, %rdi, %rax # sched: [100:?]
 ; ZNVER1-NEXT:    addq %rcx, %rax # sched: [1:0.25]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = load i64, i64 *%a2
   %2 = tail call i64 @llvm.x86.bmi.pext.64(i64 %a0, i64 %1)

 ;
 ; HASWELL-LABEL: test_rorx_i32:
 ; HASWELL:       # BB#0:
 ; HASWELL-NEXT:    rorxl $5, %edi, %ecx # sched: [1:0.50]
 ; HASWELL-NEXT:    rorxl $5, (%rdx), %eax # sched: [1:0.50]
 ; HASWELL-NEXT:    addl %ecx, %eax # sched: [1:0.25]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_rorx_i32:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    rorxl $5, %edi, %ecx # sched: [1:0.50]
+; SKYLAKE-NEXT:    rorxl $5, (%rdx), %eax # sched: [1:0.50]
+; SKYLAKE-NEXT:    addl %ecx, %eax # sched: [1:0.25]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
+; KNL-LABEL: test_rorx_i32:
+; KNL:       # BB#0:
+; KNL-NEXT:    rorxl $5, %edi, %ecx # sched: [1:0.50]
+; KNL-NEXT:    rorxl $5, (%rdx), %eax # sched: [1:0.50]
+; KNL-NEXT:    addl %ecx, %eax # sched: [1:0.25]
+; KNL-NEXT:    retq # sched: [2:1.00]
+;
 ; ZNVER1-LABEL: test_rorx_i32:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    rorxl $5, (%rdx), %eax # sched: [5:0.50]
 ; ZNVER1-NEXT:    rorxl $5, %edi, %ecx # sched: [1:0.25]
 ; ZNVER1-NEXT:    addl %ecx, %eax # sched: [1:0.25]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = load i32, i32 *%a2
   %2 = lshr i32 %a0, 5

 ;
 ; HASWELL-LABEL: test_rorx_i64:
 ; HASWELL:       # BB#0:
 ; HASWELL-NEXT:    rorxq $5, %rdi, %rcx # sched: [1:0.50]
 ; HASWELL-NEXT:    rorxq $5, (%rdx), %rax # sched: [1:0.50]
 ; HASWELL-NEXT:    addq %rcx, %rax # sched: [1:0.25]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_rorx_i64:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    rorxq $5, %rdi, %rcx # sched: [1:0.50]
+; SKYLAKE-NEXT:    rorxq $5, (%rdx), %rax # sched: [1:0.50]
+; SKYLAKE-NEXT:    addq %rcx, %rax # sched: [1:0.25]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
+; KNL-LABEL: test_rorx_i64:
+; KNL:       # BB#0:
+; KNL-NEXT:    rorxq $5, %rdi, %rcx # sched: [1:0.50]
+; KNL-NEXT:    rorxq $5, (%rdx), %rax # sched: [1:0.50]
+; KNL-NEXT:    addq %rcx, %rax # sched: [1:0.25]
+; KNL-NEXT:    retq # sched: [2:1.00]
+;
 ; ZNVER1-LABEL: test_rorx_i64:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    rorxq $5, (%rdx), %rax # sched: [5:0.50]
 ; ZNVER1-NEXT:    rorxq $5, %rdi, %rcx # sched: [1:0.25]
 ; ZNVER1-NEXT:    addq %rcx, %rax # sched: [1:0.25]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = load i64, i64 *%a2
   %2 = lshr i64 %a0, 5

 ;
 ; HASWELL-LABEL: test_sarx_i32:
 ; HASWELL:       # BB#0:
 ; HASWELL-NEXT:    sarxl %esi, %edi, %ecx # sched: [1:0.50]
 ; HASWELL-NEXT:    sarxl %esi, (%rdx), %eax # sched: [1:0.50]
 ; HASWELL-NEXT:    addl %ecx, %eax # sched: [1:0.25]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_sarx_i32:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    sarxl %esi, %edi, %ecx # sched: [1:0.50]
+; SKYLAKE-NEXT:    sarxl %esi, (%rdx), %eax # sched: [1:0.50]
+; SKYLAKE-NEXT:    addl %ecx, %eax # sched: [1:0.25]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
+; KNL-LABEL: test_sarx_i32:
+; KNL:       # BB#0:
+; KNL-NEXT:    sarxl %esi, %edi, %ecx # sched: [1:0.50]
+; KNL-NEXT:    sarxl %esi, (%rdx), %eax # sched: [1:0.50]
+; KNL-NEXT:    addl %ecx, %eax # sched: [1:0.25]
+; KNL-NEXT:    retq # sched: [2:1.00]
+;
 ; ZNVER1-LABEL: test_sarx_i32:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    sarxl %esi, (%rdx), %eax # sched: [5:0.50]
 ; ZNVER1-NEXT:    sarxl %esi, %edi, %ecx # sched: [1:0.25]
 ; ZNVER1-NEXT:    addl %ecx, %eax # sched: [1:0.25]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = load i32, i32 *%a2
   %2 = ashr i32 %a0, %a1

 ;
 ; HASWELL-LABEL: test_sarx_i64:
 ; HASWELL:       # BB#0:
 ; HASWELL-NEXT:    sarxq %rsi, %rdi, %rcx # sched: [1:0.50]
 ; HASWELL-NEXT:    sarxq %rsi, (%rdx), %rax # sched: [1:0.50]
 ; HASWELL-NEXT:    addq %rcx, %rax # sched: [1:0.25]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_sarx_i64:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    sarxq %rsi, %rdi, %rcx # sched: [1:0.50]
+; SKYLAKE-NEXT:    sarxq %rsi, (%rdx), %rax # sched: [1:0.50]
+; SKYLAKE-NEXT:    addq %rcx, %rax # sched: [1:0.25]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
+; KNL-LABEL: test_sarx_i64:
+; KNL:       # BB#0:
+; KNL-NEXT:    sarxq %rsi, %rdi, %rcx # sched: [1:0.50]
+; KNL-NEXT:    sarxq %rsi, (%rdx), %rax # sched: [1:0.50]
+; KNL-NEXT:    addq %rcx, %rax # sched: [1:0.25]
+; KNL-NEXT:    retq # sched: [2:1.00]
+;
 ; ZNVER1-LABEL: test_sarx_i64:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    sarxq %rsi, (%rdx), %rax # sched: [5:0.50]
 ; ZNVER1-NEXT:    sarxq %rsi, %rdi, %rcx # sched: [1:0.25]
 ; ZNVER1-NEXT:    addq %rcx, %rax # sched: [1:0.25]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = load i64, i64 *%a2
   %2 = ashr i64 %a0, %a1

 ;
 ; HASWELL-LABEL: test_shlx_i32:
 ; HASWELL:       # BB#0:
 ; HASWELL-NEXT:    shlxl %esi, %edi, %ecx # sched: [1:0.50]
 ; HASWELL-NEXT:    shlxl %esi, (%rdx), %eax # sched: [1:0.50]
 ; HASWELL-NEXT:    addl %ecx, %eax # sched: [1:0.25]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_shlx_i32:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    shlxl %esi, %edi, %ecx # sched: [1:0.50]
+; SKYLAKE-NEXT:    shlxl %esi, (%rdx), %eax # sched: [1:0.50]
+; SKYLAKE-NEXT:    addl %ecx, %eax # sched: [1:0.25]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
+; KNL-LABEL: test_shlx_i32:
+; KNL:       # BB#0:
+; KNL-NEXT:    shlxl %esi, %edi, %ecx # sched: [1:0.50]
+; KNL-NEXT:    shlxl %esi, (%rdx), %eax # sched: [1:0.50]
+; KNL-NEXT:    addl %ecx, %eax # sched: [1:0.25]
+; KNL-NEXT:    retq # sched: [2:1.00]
+;
 ; ZNVER1-LABEL: test_shlx_i32:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    shlxl %esi, (%rdx), %eax # sched: [5:0.50]
 ; ZNVER1-NEXT:    shlxl %esi, %edi, %ecx # sched: [1:0.25]
 ; ZNVER1-NEXT:    addl %ecx, %eax # sched: [1:0.25]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = load i32, i32 *%a2
   %2 = shl i32 %a0, %a1

 ;
 ; HASWELL-LABEL: test_shlx_i64:
 ; HASWELL:       # BB#0:
 ; HASWELL-NEXT:    shlxq %rsi, %rdi, %rcx # sched: [1:0.50]
 ; HASWELL-NEXT:    shlxq %rsi, (%rdx), %rax # sched: [1:0.50]
 ; HASWELL-NEXT:    addq %rcx, %rax # sched: [1:0.25]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_shlx_i64:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    shlxq %rsi, %rdi, %rcx # sched: [1:0.50]
+; SKYLAKE-NEXT:    shlxq %rsi, (%rdx), %rax # sched: [1:0.50]
+; SKYLAKE-NEXT:    addq %rcx, %rax # sched: [1:0.25]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
+; KNL-LABEL: test_shlx_i64:
+; KNL:       # BB#0:
+; KNL-NEXT:    shlxq %rsi, %rdi, %rcx # sched: [1:0.50]
+; KNL-NEXT:    shlxq %rsi, (%rdx), %rax # sched: [1:0.50]
+; KNL-NEXT:    addq %rcx, %rax # sched: [1:0.25]
+; KNL-NEXT:    retq # sched: [2:1.00]
+;
 ; ZNVER1-LABEL: test_shlx_i64:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    shlxq %rsi, (%rdx), %rax # sched: [5:0.50]
 ; ZNVER1-NEXT:    shlxq %rsi, %rdi, %rcx # sched: [1:0.25]
 ; ZNVER1-NEXT:    addq %rcx, %rax # sched: [1:0.25]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = load i64, i64 *%a2
   %2 = shl i64 %a0, %a1

 ;
 ; HASWELL-LABEL: test_shrx_i32:
 ; HASWELL:       # BB#0:
 ; HASWELL-NEXT:    shrxl %esi, %edi, %ecx # sched: [1:0.50]
 ; HASWELL-NEXT:    shrxl %esi, (%rdx), %eax # sched: [1:0.50]
 ; HASWELL-NEXT:    addl %ecx, %eax # sched: [1:0.25]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_shrx_i32:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    shrxl %esi, %edi, %ecx # sched: [1:0.50]
+; SKYLAKE-NEXT:    shrxl %esi, (%rdx), %eax # sched: [1:0.50]
+; SKYLAKE-NEXT:    addl %ecx, %eax # sched: [1:0.25]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
+; KNL-LABEL: test_shrx_i32:
+; KNL:       # BB#0:
+; KNL-NEXT:    shrxl %esi, %edi, %ecx # sched: [1:0.50]
+; KNL-NEXT:    shrxl %esi, (%rdx), %eax # sched: [1:0.50]
+; KNL-NEXT:    addl %ecx, %eax # sched: [1:0.25]
+; KNL-NEXT:    retq # sched: [2:1.00]
+;
 ; ZNVER1-LABEL: test_shrx_i32:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    shrxl %esi, (%rdx), %eax # sched: [5:0.50]
 ; ZNVER1-NEXT:    shrxl %esi, %edi, %ecx # sched: [1:0.25]
 ; ZNVER1-NEXT:    addl %ecx, %eax # sched: [1:0.25]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = load i32, i32 *%a2
   %2 = lshr i32 %a0, %a1

 ;
 ; HASWELL-LABEL: test_shrx_i64:
 ; HASWELL:       # BB#0:
 ; HASWELL-NEXT:    shrxq %rsi, %rdi, %rcx # sched: [1:0.50]
 ; HASWELL-NEXT:    shrxq %rsi, (%rdx), %rax # sched: [1:0.50]
 ; HASWELL-NEXT:    addq %rcx, %rax # sched: [1:0.25]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
+; SKYLAKE-LABEL: test_shrx_i64:
+; SKYLAKE:       # BB#0:
+; SKYLAKE-NEXT:    shrxq %rsi, %rdi, %rcx # sched: [1:0.50]
+; SKYLAKE-NEXT:    shrxq %rsi, (%rdx), %rax # sched: [1:0.50]
+; SKYLAKE-NEXT:    addq %rcx, %rax # sched: [1:0.25]
+; SKYLAKE-NEXT:    retq # sched: [2:1.00]
+;
+; KNL-LABEL: test_shrx_i64:
+; KNL:       # BB#0:
+; KNL-NEXT:    shrxq %rsi, %rdi, %rcx # sched: [1:0.50]
+; KNL-NEXT:    shrxq %rsi, (%rdx), %rax # sched: [1:0.50]
+; KNL-NEXT:    addq %rcx, %rax # sched: [1:0.25]
+; KNL-NEXT:    retq # sched: [2:1.00]
+;
 ; ZNVER1-LABEL: test_shrx_i64:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    shrxq %rsi, (%rdx), %rax # sched: [5:0.50]
 ; ZNVER1-NEXT:    shrxq %rsi, %rdi, %rcx # sched: [1:0.25]
 ; ZNVER1-NEXT:    addq %rcx, %rax # sched: [1:0.25]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = load i64, i64 *%a2
   %2 = lshr i64 %a0, %a1
   %3 = lshr i64 %1, %a1
   %4 = add i64 %2, %3
   ret i64 %4
 }