Index: test/CodeGen/X86/bmi2-schedule.ll =================================================================== --- test/CodeGen/X86/bmi2-schedule.ll +++ test/CodeGen/X86/bmi2-schedule.ll @@ -1,8 +1,8 @@ ; 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=skylake | FileCheck %s --check-prefix=CHECK --check-prefix=SKYLAKE +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=knl | FileCheck %s --check-prefix=CHECK --check-prefix=KNL ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -print-schedule -mcpu=znver1 | FileCheck %s --check-prefix=CHECK --check-prefix=ZNVER1 define i32 @test_bzhi_i32(i32 %a0, i32 %a1, i32 *%a2) { @@ -20,6 +20,20 @@ ; 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] @@ -49,6 +63,20 @@ ; 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] @@ -84,6 +112,24 @@ ; 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] @@ -121,6 +167,20 @@ ; 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:?] @@ -150,6 +210,20 @@ ; 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:?] @@ -179,6 +253,20 @@ ; 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:?] @@ -208,6 +296,20 @@ ; 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:?] @@ -237,6 +339,20 @@ ; 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] @@ -269,6 +385,20 @@ ; 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] @@ -301,6 +431,20 @@ ; 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] @@ -329,6 +473,20 @@ ; 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] @@ -357,6 +515,20 @@ ; 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] @@ -385,6 +557,20 @@ ; 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] @@ -413,6 +599,20 @@ ; 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] @@ -441,6 +641,20 @@ ; 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]