maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity
llvm-project/llvm/test/CodeGen/X86/unfold-masked-merge-vector-...

4669 lines
194 KiB
LLVM
Raw Normal View History

[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc -mtriple=x86_64-unknown-linux-gnu -mattr=-sse,-sse2 < %s | FileCheck %s --check-prefixes=CHECK,CHECK-BASELINE
; RUN: llc -mtriple=x86_64-unknown-linux-gnu -mattr=+sse,-sse2 < %s | FileCheck %s --check-prefixes=CHECK,CHECK-SSE,CHECK-SSE1
; RUN: llc -mtriple=x86_64-unknown-linux-gnu -mattr=+sse,+sse2 < %s | FileCheck %s --check-prefixes=CHECK,CHECK-SSE,CHECK-SSE2
; RUN: llc -mtriple=x86_64-unknown-linux-gnu -mattr=+xop < %s | FileCheck %s --check-prefixes=CHECK,CHECK-XOP
; https://bugs.llvm.org/show_bug.cgi?id=37104
; All the advanced stuff (negative tests, commutativity) is handled in the
; scalar version of the test only.
; ============================================================================ ;
; 8-bit vector width
; ============================================================================ ;
define <1 x i8> @out_v1i8(<1 x i8> %x, <1 x i8> %y, <1 x i8> %mask) nounwind {
; CHECK-LABEL: out_v1i8:
; CHECK: # %bb.0:
; CHECK-NEXT: movl %edx, %eax
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-NEXT: andl %edx, %edi
; CHECK-NEXT: notb %al
; CHECK-NEXT: andb %sil, %al
; CHECK-NEXT: orb %dil, %al
; CHECK-NEXT: # kill: def $al killed $al killed $eax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-NEXT: retq
%mx = and <1 x i8> %x, %mask
%notmask = xor <1 x i8> %mask, <i8 -1>
%my = and <1 x i8> %y, %notmask
%r = or <1 x i8> %mx, %my
ret <1 x i8> %r
}
; ============================================================================ ;
; 16-bit vector width
; ============================================================================ ;
define <2 x i8> @out_v2i8(<2 x i8> %x, <2 x i8> %y, <2 x i8> %mask) nounwind {
; CHECK-BASELINE-LABEL: out_v2i8:
; CHECK-BASELINE: # %bb.0:
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movl %r8d, %eax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andl %r9d, %esi
[DAGCombiner] use root SDLoc for all nodes created by logic fold If this is not a valid way to assign an SDLoc, then we get this wrong all over SDAG. I don't know enough about the SDAG to explain this. IIUC, theoretically, debug info is not supposed to affect codegen. But here it has clearly affected 3 different targets, and the x86 change is an actual improvement. llvm-svn: 348552
2018-12-07 08:01:57 +08:00
; CHECK-BASELINE-NEXT: andl %r8d, %edi
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: notb %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: notb %r9b
; CHECK-BASELINE-NEXT: andb %cl, %r9b
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: andb %dl, %al
; CHECK-BASELINE-NEXT: orb %dil, %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: orb %sil, %r9b
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: # kill: def $al killed $al killed $eax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl %r9d, %edx
; CHECK-BASELINE-NEXT: retq
;
; CHECK-SSE1-LABEL: out_v2i8:
; CHECK-SSE1: # %bb.0:
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movl %r8d, %eax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: andl %r9d, %esi
[DAGCombiner] use root SDLoc for all nodes created by logic fold If this is not a valid way to assign an SDLoc, then we get this wrong all over SDAG. I don't know enough about the SDAG to explain this. IIUC, theoretically, debug info is not supposed to affect codegen. But here it has clearly affected 3 different targets, and the x86 change is an actual improvement. llvm-svn: 348552
2018-12-07 08:01:57 +08:00
; CHECK-SSE1-NEXT: andl %r8d, %edi
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: notb %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: notb %r9b
; CHECK-SSE1-NEXT: andb %cl, %r9b
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: andb %dl, %al
; CHECK-SSE1-NEXT: orb %dil, %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: orb %sil, %r9b
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: # kill: def $al killed $al killed $eax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movl %r9d, %edx
; CHECK-SSE1-NEXT: retq
;
; CHECK-SSE2-LABEL: out_v2i8:
; CHECK-SSE2: # %bb.0:
; CHECK-SSE2-NEXT: andps %xmm2, %xmm0
; CHECK-SSE2-NEXT: xorps {{.*}}(%rip), %xmm2
; CHECK-SSE2-NEXT: andps %xmm1, %xmm2
; CHECK-SSE2-NEXT: orps %xmm2, %xmm0
; CHECK-SSE2-NEXT: retq
;
; CHECK-XOP-LABEL: out_v2i8:
; CHECK-XOP: # %bb.0:
; CHECK-XOP-NEXT: vandps %xmm2, %xmm0, %xmm0
; CHECK-XOP-NEXT: vxorps {{.*}}(%rip), %xmm2, %xmm2
; CHECK-XOP-NEXT: vandps %xmm2, %xmm1, %xmm1
; CHECK-XOP-NEXT: vorps %xmm1, %xmm0, %xmm0
; CHECK-XOP-NEXT: retq
%mx = and <2 x i8> %x, %mask
%notmask = xor <2 x i8> %mask, <i8 -1, i8 -1>
%my = and <2 x i8> %y, %notmask
%r = or <2 x i8> %mx, %my
ret <2 x i8> %r
}
define <1 x i16> @out_v1i16(<1 x i16> %x, <1 x i16> %y, <1 x i16> %mask) nounwind {
; CHECK-LABEL: out_v1i16:
; CHECK: # %bb.0:
; CHECK-NEXT: movl %edx, %eax
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-NEXT: andl %edx, %edi
; CHECK-NEXT: notl %eax
; CHECK-NEXT: andl %esi, %eax
; CHECK-NEXT: orl %edi, %eax
; CHECK-NEXT: # kill: def $ax killed $ax killed $eax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-NEXT: retq
%mx = and <1 x i16> %x, %mask
%notmask = xor <1 x i16> %mask, <i16 -1>
%my = and <1 x i16> %y, %notmask
%r = or <1 x i16> %mx, %my
ret <1 x i16> %r
}
; ============================================================================ ;
; 32-bit vector width
; ============================================================================ ;
define <4 x i8> @out_v4i8(<4 x i8> %x, <4 x i8> %y, <4 x i8> %mask) nounwind {
; CHECK-BASELINE-LABEL: out_v4i8:
; CHECK-BASELINE: # %bb.0:
; CHECK-BASELINE-NEXT: pushq %rbx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movq %rdi, %rax
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %dil
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %r10b
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %r11b
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %bl
; CHECK-BASELINE-NEXT: andb %bl, %r8b
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: andb %r11b, %cl
; CHECK-BASELINE-NEXT: andb %r10b, %dl
; CHECK-BASELINE-NEXT: andb %dil, %sil
; CHECK-BASELINE-NEXT: notb %r10b
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: notb %r11b
; CHECK-BASELINE-NEXT: notb %bl
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: notb %dil
; CHECK-BASELINE-NEXT: andb %r9b, %dil
; CHECK-BASELINE-NEXT: orb %sil, %dil
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %bl
; CHECK-BASELINE-NEXT: orb %r8b, %bl
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %r11b
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: orb %cl, %r11b
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %r10b
; CHECK-BASELINE-NEXT: orb %dl, %r10b
; CHECK-BASELINE-NEXT: movb %bl, 3(%rax)
; CHECK-BASELINE-NEXT: movb %r11b, 2(%rax)
; CHECK-BASELINE-NEXT: movb %r10b, 1(%rax)
; CHECK-BASELINE-NEXT: movb %dil, (%rax)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: popq %rbx
; CHECK-BASELINE-NEXT: retq
;
; CHECK-SSE1-LABEL: out_v4i8:
; CHECK-SSE1: # %bb.0:
; CHECK-SSE1-NEXT: pushq %rbx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movq %rdi, %rax
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %dil
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %r10b
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %r11b
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %bl
; CHECK-SSE1-NEXT: andb %bl, %r8b
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: andb %r11b, %cl
; CHECK-SSE1-NEXT: andb %r10b, %dl
; CHECK-SSE1-NEXT: andb %dil, %sil
; CHECK-SSE1-NEXT: notb %r10b
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: notb %r11b
; CHECK-SSE1-NEXT: notb %bl
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: notb %dil
; CHECK-SSE1-NEXT: andb %r9b, %dil
; CHECK-SSE1-NEXT: orb %sil, %dil
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %bl
; CHECK-SSE1-NEXT: orb %r8b, %bl
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %r11b
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: orb %cl, %r11b
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %r10b
; CHECK-SSE1-NEXT: orb %dl, %r10b
; CHECK-SSE1-NEXT: movb %bl, 3(%rax)
; CHECK-SSE1-NEXT: movb %r11b, 2(%rax)
; CHECK-SSE1-NEXT: movb %r10b, 1(%rax)
; CHECK-SSE1-NEXT: movb %dil, (%rax)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: popq %rbx
; CHECK-SSE1-NEXT: retq
;
; CHECK-SSE2-LABEL: out_v4i8:
; CHECK-SSE2: # %bb.0:
; CHECK-SSE2-NEXT: andps %xmm2, %xmm0
; CHECK-SSE2-NEXT: xorps {{.*}}(%rip), %xmm2
; CHECK-SSE2-NEXT: andps %xmm1, %xmm2
; CHECK-SSE2-NEXT: orps %xmm2, %xmm0
; CHECK-SSE2-NEXT: retq
;
; CHECK-XOP-LABEL: out_v4i8:
; CHECK-XOP: # %bb.0:
; CHECK-XOP-NEXT: vandps %xmm2, %xmm0, %xmm0
; CHECK-XOP-NEXT: vxorps {{.*}}(%rip), %xmm2, %xmm2
; CHECK-XOP-NEXT: vandps %xmm2, %xmm1, %xmm1
; CHECK-XOP-NEXT: vorps %xmm1, %xmm0, %xmm0
; CHECK-XOP-NEXT: retq
%mx = and <4 x i8> %x, %mask
%notmask = xor <4 x i8> %mask, <i8 -1, i8 -1, i8 -1, i8 -1>
%my = and <4 x i8> %y, %notmask
%r = or <4 x i8> %mx, %my
ret <4 x i8> %r
}
define <4 x i8> @out_v4i8_undef(<4 x i8> %x, <4 x i8> %y, <4 x i8> %mask) nounwind {
; CHECK-BASELINE-LABEL: out_v4i8_undef:
; CHECK-BASELINE: # %bb.0:
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movq %rdi, %rax
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %dil
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %r10b
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %r11b
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: andb %r11b, %r8b
; CHECK-BASELINE-NEXT: andb %r10b, %dl
; CHECK-BASELINE-NEXT: andb %dil, %sil
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %cl
; CHECK-BASELINE-NEXT: notb %r10b
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: notb %r11b
; CHECK-BASELINE-NEXT: notb %dil
; CHECK-BASELINE-NEXT: andb %r9b, %dil
; CHECK-BASELINE-NEXT: orb %sil, %dil
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %r11b
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: orb %r8b, %r11b
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %r10b
; CHECK-BASELINE-NEXT: orb %dl, %r10b
; CHECK-BASELINE-NEXT: movb %cl, 2(%rax)
; CHECK-BASELINE-NEXT: movb %r11b, 3(%rax)
; CHECK-BASELINE-NEXT: movb %r10b, 1(%rax)
; CHECK-BASELINE-NEXT: movb %dil, (%rax)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: retq
;
; CHECK-SSE1-LABEL: out_v4i8_undef:
; CHECK-SSE1: # %bb.0:
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movq %rdi, %rax
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %dil
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %r10b
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %r11b
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: andb %r11b, %r8b
; CHECK-SSE1-NEXT: andb %r10b, %dl
; CHECK-SSE1-NEXT: andb %dil, %sil
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %cl
; CHECK-SSE1-NEXT: notb %r10b
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: notb %r11b
; CHECK-SSE1-NEXT: notb %dil
; CHECK-SSE1-NEXT: andb %r9b, %dil
; CHECK-SSE1-NEXT: orb %sil, %dil
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %r11b
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: orb %r8b, %r11b
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %r10b
; CHECK-SSE1-NEXT: orb %dl, %r10b
; CHECK-SSE1-NEXT: movb %cl, 2(%rax)
; CHECK-SSE1-NEXT: movb %r11b, 3(%rax)
; CHECK-SSE1-NEXT: movb %r10b, 1(%rax)
; CHECK-SSE1-NEXT: movb %dil, (%rax)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: retq
;
; CHECK-SSE2-LABEL: out_v4i8_undef:
; CHECK-SSE2: # %bb.0:
; CHECK-SSE2-NEXT: andps %xmm2, %xmm0
; CHECK-SSE2-NEXT: xorps {{.*}}(%rip), %xmm2
; CHECK-SSE2-NEXT: andps %xmm1, %xmm2
; CHECK-SSE2-NEXT: orps %xmm2, %xmm0
; CHECK-SSE2-NEXT: retq
;
; CHECK-XOP-LABEL: out_v4i8_undef:
; CHECK-XOP: # %bb.0:
; CHECK-XOP-NEXT: vandps %xmm2, %xmm0, %xmm0
; CHECK-XOP-NEXT: vxorps {{.*}}(%rip), %xmm2, %xmm2
; CHECK-XOP-NEXT: vandps %xmm2, %xmm1, %xmm1
; CHECK-XOP-NEXT: vorps %xmm1, %xmm0, %xmm0
; CHECK-XOP-NEXT: retq
%mx = and <4 x i8> %x, %mask
%notmask = xor <4 x i8> %mask, <i8 -1, i8 -1, i8 undef, i8 -1>
%my = and <4 x i8> %y, %notmask
%r = or <4 x i8> %mx, %my
ret <4 x i8> %r
}
define <2 x i16> @out_v2i16(<2 x i16> %x, <2 x i16> %y, <2 x i16> %mask) nounwind {
; CHECK-BASELINE-LABEL: out_v2i16:
; CHECK-BASELINE: # %bb.0:
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movl %r8d, %eax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andl %r9d, %esi
; CHECK-BASELINE-NEXT: andl %r8d, %edi
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: notl %eax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: notl %r9d
; CHECK-BASELINE-NEXT: andl %ecx, %r9d
; CHECK-BASELINE-NEXT: orl %esi, %r9d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: andl %edx, %eax
; CHECK-BASELINE-NEXT: orl %edi, %eax
; CHECK-BASELINE-NEXT: # kill: def $ax killed $ax killed $eax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl %r9d, %edx
; CHECK-BASELINE-NEXT: retq
;
; CHECK-SSE1-LABEL: out_v2i16:
; CHECK-SSE1: # %bb.0:
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movl %r8d, %eax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: andl %r9d, %esi
; CHECK-SSE1-NEXT: andl %r8d, %edi
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: notl %eax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: notl %r9d
; CHECK-SSE1-NEXT: andl %ecx, %r9d
; CHECK-SSE1-NEXT: orl %esi, %r9d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: andl %edx, %eax
; CHECK-SSE1-NEXT: orl %edi, %eax
; CHECK-SSE1-NEXT: # kill: def $ax killed $ax killed $eax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movl %r9d, %edx
; CHECK-SSE1-NEXT: retq
;
; CHECK-SSE2-LABEL: out_v2i16:
; CHECK-SSE2: # %bb.0:
; CHECK-SSE2-NEXT: andps %xmm2, %xmm0
; CHECK-SSE2-NEXT: xorps {{.*}}(%rip), %xmm2
; CHECK-SSE2-NEXT: andps %xmm1, %xmm2
; CHECK-SSE2-NEXT: orps %xmm2, %xmm0
; CHECK-SSE2-NEXT: retq
;
; CHECK-XOP-LABEL: out_v2i16:
; CHECK-XOP: # %bb.0:
; CHECK-XOP-NEXT: vandps %xmm2, %xmm0, %xmm0
; CHECK-XOP-NEXT: vxorps {{.*}}(%rip), %xmm2, %xmm2
; CHECK-XOP-NEXT: vandps %xmm2, %xmm1, %xmm1
; CHECK-XOP-NEXT: vorps %xmm1, %xmm0, %xmm0
; CHECK-XOP-NEXT: retq
%mx = and <2 x i16> %x, %mask
%notmask = xor <2 x i16> %mask, <i16 -1, i16 -1>
%my = and <2 x i16> %y, %notmask
%r = or <2 x i16> %mx, %my
ret <2 x i16> %r
}
define <1 x i32> @out_v1i32(<1 x i32> %x, <1 x i32> %y, <1 x i32> %mask) nounwind {
; CHECK-LABEL: out_v1i32:
; CHECK: # %bb.0:
; CHECK-NEXT: movl %edx, %eax
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-NEXT: andl %edx, %edi
; CHECK-NEXT: notl %eax
; CHECK-NEXT: andl %esi, %eax
; CHECK-NEXT: orl %edi, %eax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-NEXT: retq
%mx = and <1 x i32> %x, %mask
%notmask = xor <1 x i32> %mask, <i32 -1>
%my = and <1 x i32> %y, %notmask
%r = or <1 x i32> %mx, %my
ret <1 x i32> %r
}
; ============================================================================ ;
; 64-bit vector width
; ============================================================================ ;
define <8 x i8> @out_v8i8(<8 x i8> %x, <8 x i8> %y, <8 x i8> %mask) nounwind {
; CHECK-BASELINE-LABEL: out_v8i8:
; CHECK-BASELINE: # %bb.0:
; CHECK-BASELINE-NEXT: pushq %rbp
; CHECK-BASELINE-NEXT: pushq %r15
; CHECK-BASELINE-NEXT: pushq %r14
; CHECK-BASELINE-NEXT: pushq %r13
; CHECK-BASELINE-NEXT: pushq %r12
; CHECK-BASELINE-NEXT: pushq %rbx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movq %rdi, %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %r10b
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %r14b
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %r12b
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %dil
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %r11b
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %bpl
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %r15b
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %bl
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: andb %bl, %r9b
; CHECK-BASELINE-NEXT: andb %r15b, %r8b
; CHECK-BASELINE-NEXT: andb %bpl, %cl
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl %ecx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-BASELINE-NEXT: andb %r11b, %dl
; CHECK-BASELINE-NEXT: movl %edx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: andb %dil, %sil
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %r13b
; CHECK-BASELINE-NEXT: andb %r12b, %r13b
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %cl
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: andb %r14b, %cl
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %dl
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: andb %r10b, %dl
; CHECK-BASELINE-NEXT: notb %dil
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: notb %r11b
; CHECK-BASELINE-NEXT: notb %bpl
; CHECK-BASELINE-NEXT: notb %r15b
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: notb %bl
; CHECK-BASELINE-NEXT: notb %r10b
; CHECK-BASELINE-NEXT: notb %r14b
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: notb %r12b
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %r12b
; CHECK-BASELINE-NEXT: orb %r13b, %r12b
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %r14b
; CHECK-BASELINE-NEXT: orb %cl, %r14b
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %r10b
; CHECK-BASELINE-NEXT: orb %dl, %r10b
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %bl
; CHECK-BASELINE-NEXT: orb %r9b, %bl
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %r15b
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: orb %r8b, %r15b
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %bpl
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: orb {{[-0-9]+}}(%r{{[sb]}}p), %bpl # 1-byte Folded Reload
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %r11b
; CHECK-BASELINE-NEXT: orb {{[-0-9]+}}(%r{{[sb]}}p), %r11b # 1-byte Folded Reload
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %dil
; CHECK-BASELINE-NEXT: orb %sil, %dil
; CHECK-BASELINE-NEXT: movb %r12b, 7(%rax)
; CHECK-BASELINE-NEXT: movb %r14b, 6(%rax)
; CHECK-BASELINE-NEXT: movb %r10b, 5(%rax)
; CHECK-BASELINE-NEXT: movb %bl, 4(%rax)
; CHECK-BASELINE-NEXT: movb %r15b, 3(%rax)
; CHECK-BASELINE-NEXT: movb %bpl, 2(%rax)
; CHECK-BASELINE-NEXT: movb %r11b, 1(%rax)
; CHECK-BASELINE-NEXT: movb %dil, (%rax)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: popq %rbx
; CHECK-BASELINE-NEXT: popq %r12
; CHECK-BASELINE-NEXT: popq %r13
; CHECK-BASELINE-NEXT: popq %r14
; CHECK-BASELINE-NEXT: popq %r15
; CHECK-BASELINE-NEXT: popq %rbp
; CHECK-BASELINE-NEXT: retq
;
; CHECK-SSE1-LABEL: out_v8i8:
; CHECK-SSE1: # %bb.0:
; CHECK-SSE1-NEXT: pushq %rbp
; CHECK-SSE1-NEXT: pushq %r15
; CHECK-SSE1-NEXT: pushq %r14
; CHECK-SSE1-NEXT: pushq %r13
; CHECK-SSE1-NEXT: pushq %r12
; CHECK-SSE1-NEXT: pushq %rbx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movq %rdi, %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %r10b
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %r14b
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %r12b
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %dil
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %r11b
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %bpl
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %r15b
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %bl
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: andb %bl, %r9b
; CHECK-SSE1-NEXT: andb %r15b, %r8b
; CHECK-SSE1-NEXT: andb %bpl, %cl
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movl %ecx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-SSE1-NEXT: andb %r11b, %dl
; CHECK-SSE1-NEXT: movl %edx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: andb %dil, %sil
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %r13b
; CHECK-SSE1-NEXT: andb %r12b, %r13b
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %cl
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: andb %r14b, %cl
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %dl
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: andb %r10b, %dl
; CHECK-SSE1-NEXT: notb %dil
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: notb %r11b
; CHECK-SSE1-NEXT: notb %bpl
; CHECK-SSE1-NEXT: notb %r15b
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: notb %bl
; CHECK-SSE1-NEXT: notb %r10b
; CHECK-SSE1-NEXT: notb %r14b
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: notb %r12b
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %r12b
; CHECK-SSE1-NEXT: orb %r13b, %r12b
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %r14b
; CHECK-SSE1-NEXT: orb %cl, %r14b
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %r10b
; CHECK-SSE1-NEXT: orb %dl, %r10b
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %bl
; CHECK-SSE1-NEXT: orb %r9b, %bl
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %r15b
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: orb %r8b, %r15b
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %bpl
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: orb {{[-0-9]+}}(%r{{[sb]}}p), %bpl # 1-byte Folded Reload
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %r11b
; CHECK-SSE1-NEXT: orb {{[-0-9]+}}(%r{{[sb]}}p), %r11b # 1-byte Folded Reload
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %dil
; CHECK-SSE1-NEXT: orb %sil, %dil
; CHECK-SSE1-NEXT: movb %r12b, 7(%rax)
; CHECK-SSE1-NEXT: movb %r14b, 6(%rax)
; CHECK-SSE1-NEXT: movb %r10b, 5(%rax)
; CHECK-SSE1-NEXT: movb %bl, 4(%rax)
; CHECK-SSE1-NEXT: movb %r15b, 3(%rax)
; CHECK-SSE1-NEXT: movb %bpl, 2(%rax)
; CHECK-SSE1-NEXT: movb %r11b, 1(%rax)
; CHECK-SSE1-NEXT: movb %dil, (%rax)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: popq %rbx
; CHECK-SSE1-NEXT: popq %r12
; CHECK-SSE1-NEXT: popq %r13
; CHECK-SSE1-NEXT: popq %r14
; CHECK-SSE1-NEXT: popq %r15
; CHECK-SSE1-NEXT: popq %rbp
; CHECK-SSE1-NEXT: retq
;
; CHECK-SSE2-LABEL: out_v8i8:
; CHECK-SSE2: # %bb.0:
; CHECK-SSE2-NEXT: andps %xmm2, %xmm0
; CHECK-SSE2-NEXT: xorps {{.*}}(%rip), %xmm2
; CHECK-SSE2-NEXT: andps %xmm1, %xmm2
; CHECK-SSE2-NEXT: orps %xmm2, %xmm0
; CHECK-SSE2-NEXT: retq
;
; CHECK-XOP-LABEL: out_v8i8:
; CHECK-XOP: # %bb.0:
; CHECK-XOP-NEXT: vandps %xmm2, %xmm0, %xmm0
; CHECK-XOP-NEXT: vxorps {{.*}}(%rip), %xmm2, %xmm2
; CHECK-XOP-NEXT: vandps %xmm2, %xmm1, %xmm1
; CHECK-XOP-NEXT: vorps %xmm1, %xmm0, %xmm0
; CHECK-XOP-NEXT: retq
%mx = and <8 x i8> %x, %mask
%notmask = xor <8 x i8> %mask, <i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1>
%my = and <8 x i8> %y, %notmask
%r = or <8 x i8> %mx, %my
ret <8 x i8> %r
}
define <4 x i16> @out_v4i16(<4 x i16> %x, <4 x i16> %y, <4 x i16> %mask) nounwind {
; CHECK-BASELINE-LABEL: out_v4i16:
; CHECK-BASELINE: # %bb.0:
; CHECK-BASELINE-NEXT: pushq %rbx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movq %rdi, %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl {{[0-9]+}}(%rsp), %r10d
; CHECK-BASELINE-NEXT: movl {{[0-9]+}}(%rsp), %r11d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movl {{[0-9]+}}(%rsp), %edi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl {{[0-9]+}}(%rsp), %ebx
; CHECK-BASELINE-NEXT: andl %ebx, %esi
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: andl %edi, %r8d
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andl %r11d, %ecx
; CHECK-BASELINE-NEXT: andl %r10d, %edx
; CHECK-BASELINE-NEXT: notl %r10d
; CHECK-BASELINE-NEXT: notl %r11d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: notl %edi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: notl %ebx
; CHECK-BASELINE-NEXT: andl %r9d, %ebx
; CHECK-BASELINE-NEXT: orl %esi, %ebx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: andw {{[0-9]+}}(%rsp), %di
; CHECK-BASELINE-NEXT: orl %r8d, %edi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andw {{[0-9]+}}(%rsp), %r11w
; CHECK-BASELINE-NEXT: orl %ecx, %r11d
; CHECK-BASELINE-NEXT: andw {{[0-9]+}}(%rsp), %r10w
; CHECK-BASELINE-NEXT: orl %edx, %r10d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movw %bx, (%rax)
; CHECK-BASELINE-NEXT: movw %di, 6(%rax)
; CHECK-BASELINE-NEXT: movw %r11w, 4(%rax)
; CHECK-BASELINE-NEXT: movw %r10w, 2(%rax)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: popq %rbx
; CHECK-BASELINE-NEXT: retq
;
; CHECK-SSE1-LABEL: out_v4i16:
; CHECK-SSE1: # %bb.0:
; CHECK-SSE1-NEXT: pushq %rbx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movq %rdi, %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movl {{[0-9]+}}(%rsp), %r10d
; CHECK-SSE1-NEXT: movl {{[0-9]+}}(%rsp), %r11d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movl {{[0-9]+}}(%rsp), %edi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movl {{[0-9]+}}(%rsp), %ebx
; CHECK-SSE1-NEXT: andl %ebx, %esi
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: andl %edi, %r8d
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: andl %r11d, %ecx
; CHECK-SSE1-NEXT: andl %r10d, %edx
; CHECK-SSE1-NEXT: notl %r10d
; CHECK-SSE1-NEXT: notl %r11d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: notl %edi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: notl %ebx
; CHECK-SSE1-NEXT: andl %r9d, %ebx
; CHECK-SSE1-NEXT: orl %esi, %ebx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: andw {{[0-9]+}}(%rsp), %di
; CHECK-SSE1-NEXT: orl %r8d, %edi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: andw {{[0-9]+}}(%rsp), %r11w
; CHECK-SSE1-NEXT: orl %ecx, %r11d
; CHECK-SSE1-NEXT: andw {{[0-9]+}}(%rsp), %r10w
; CHECK-SSE1-NEXT: orl %edx, %r10d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movw %bx, (%rax)
; CHECK-SSE1-NEXT: movw %di, 6(%rax)
; CHECK-SSE1-NEXT: movw %r11w, 4(%rax)
; CHECK-SSE1-NEXT: movw %r10w, 2(%rax)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: popq %rbx
; CHECK-SSE1-NEXT: retq
;
; CHECK-SSE2-LABEL: out_v4i16:
; CHECK-SSE2: # %bb.0:
; CHECK-SSE2-NEXT: andps %xmm2, %xmm0
; CHECK-SSE2-NEXT: xorps {{.*}}(%rip), %xmm2
; CHECK-SSE2-NEXT: andps %xmm1, %xmm2
; CHECK-SSE2-NEXT: orps %xmm2, %xmm0
; CHECK-SSE2-NEXT: retq
;
; CHECK-XOP-LABEL: out_v4i16:
; CHECK-XOP: # %bb.0:
; CHECK-XOP-NEXT: vandps %xmm2, %xmm0, %xmm0
; CHECK-XOP-NEXT: vxorps {{.*}}(%rip), %xmm2, %xmm2
; CHECK-XOP-NEXT: vandps %xmm2, %xmm1, %xmm1
; CHECK-XOP-NEXT: vorps %xmm1, %xmm0, %xmm0
; CHECK-XOP-NEXT: retq
%mx = and <4 x i16> %x, %mask
%notmask = xor <4 x i16> %mask, <i16 -1, i16 -1, i16 -1, i16 -1>
%my = and <4 x i16> %y, %notmask
%r = or <4 x i16> %mx, %my
ret <4 x i16> %r
}
define <4 x i16> @out_v4i16_undef(<4 x i16> %x, <4 x i16> %y, <4 x i16> %mask) nounwind {
; CHECK-BASELINE-LABEL: out_v4i16_undef:
; CHECK-BASELINE: # %bb.0:
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movq %rdi, %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl {{[0-9]+}}(%rsp), %r10d
; CHECK-BASELINE-NEXT: movl {{[0-9]+}}(%rsp), %r11d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movl {{[0-9]+}}(%rsp), %edi
; CHECK-BASELINE-NEXT: andl %edi, %esi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andl %r11d, %r8d
; CHECK-BASELINE-NEXT: andl %r10d, %edx
; CHECK-BASELINE-NEXT: andw {{[0-9]+}}(%rsp), %cx
; CHECK-BASELINE-NEXT: notl %r10d
; CHECK-BASELINE-NEXT: notl %r11d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: notl %edi
; CHECK-BASELINE-NEXT: andl %r9d, %edi
; CHECK-BASELINE-NEXT: orl %esi, %edi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andw {{[0-9]+}}(%rsp), %r11w
; CHECK-BASELINE-NEXT: orl %r8d, %r11d
; CHECK-BASELINE-NEXT: andw {{[0-9]+}}(%rsp), %r10w
; CHECK-BASELINE-NEXT: orl %edx, %r10d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movw %cx, 4(%rax)
; CHECK-BASELINE-NEXT: movw %di, (%rax)
; CHECK-BASELINE-NEXT: movw %r11w, 6(%rax)
; CHECK-BASELINE-NEXT: movw %r10w, 2(%rax)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: retq
;
; CHECK-SSE1-LABEL: out_v4i16_undef:
; CHECK-SSE1: # %bb.0:
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movq %rdi, %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movl {{[0-9]+}}(%rsp), %r10d
; CHECK-SSE1-NEXT: movl {{[0-9]+}}(%rsp), %r11d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movl {{[0-9]+}}(%rsp), %edi
; CHECK-SSE1-NEXT: andl %edi, %esi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: andl %r11d, %r8d
; CHECK-SSE1-NEXT: andl %r10d, %edx
; CHECK-SSE1-NEXT: andw {{[0-9]+}}(%rsp), %cx
; CHECK-SSE1-NEXT: notl %r10d
; CHECK-SSE1-NEXT: notl %r11d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: notl %edi
; CHECK-SSE1-NEXT: andl %r9d, %edi
; CHECK-SSE1-NEXT: orl %esi, %edi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: andw {{[0-9]+}}(%rsp), %r11w
; CHECK-SSE1-NEXT: orl %r8d, %r11d
; CHECK-SSE1-NEXT: andw {{[0-9]+}}(%rsp), %r10w
; CHECK-SSE1-NEXT: orl %edx, %r10d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movw %cx, 4(%rax)
; CHECK-SSE1-NEXT: movw %di, (%rax)
; CHECK-SSE1-NEXT: movw %r11w, 6(%rax)
; CHECK-SSE1-NEXT: movw %r10w, 2(%rax)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: retq
;
; CHECK-SSE2-LABEL: out_v4i16_undef:
; CHECK-SSE2: # %bb.0:
; CHECK-SSE2-NEXT: andps %xmm2, %xmm0
; CHECK-SSE2-NEXT: xorps {{.*}}(%rip), %xmm2
; CHECK-SSE2-NEXT: andps %xmm1, %xmm2
; CHECK-SSE2-NEXT: orps %xmm2, %xmm0
; CHECK-SSE2-NEXT: retq
;
; CHECK-XOP-LABEL: out_v4i16_undef:
; CHECK-XOP: # %bb.0:
; CHECK-XOP-NEXT: vandps %xmm2, %xmm0, %xmm0
; CHECK-XOP-NEXT: vxorps {{.*}}(%rip), %xmm2, %xmm2
; CHECK-XOP-NEXT: vandps %xmm2, %xmm1, %xmm1
; CHECK-XOP-NEXT: vorps %xmm1, %xmm0, %xmm0
; CHECK-XOP-NEXT: retq
%mx = and <4 x i16> %x, %mask
%notmask = xor <4 x i16> %mask, <i16 -1, i16 -1, i16 undef, i16 -1>
%my = and <4 x i16> %y, %notmask
%r = or <4 x i16> %mx, %my
ret <4 x i16> %r
}
define <2 x i32> @out_v2i32(<2 x i32> %x, <2 x i32> %y, <2 x i32> %mask) nounwind {
; CHECK-BASELINE-LABEL: out_v2i32:
; CHECK-BASELINE: # %bb.0:
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movl %r8d, %eax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andl %r9d, %esi
; CHECK-BASELINE-NEXT: andl %r8d, %edi
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: notl %eax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: notl %r9d
; CHECK-BASELINE-NEXT: andl %ecx, %r9d
; CHECK-BASELINE-NEXT: orl %esi, %r9d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: andl %edx, %eax
; CHECK-BASELINE-NEXT: orl %edi, %eax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl %r9d, %edx
; CHECK-BASELINE-NEXT: retq
;
; CHECK-SSE1-LABEL: out_v2i32:
; CHECK-SSE1: # %bb.0:
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movl %r8d, %eax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: andl %r9d, %esi
; CHECK-SSE1-NEXT: andl %r8d, %edi
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: notl %eax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: notl %r9d
; CHECK-SSE1-NEXT: andl %ecx, %r9d
; CHECK-SSE1-NEXT: orl %esi, %r9d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: andl %edx, %eax
; CHECK-SSE1-NEXT: orl %edi, %eax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movl %r9d, %edx
; CHECK-SSE1-NEXT: retq
;
; CHECK-SSE2-LABEL: out_v2i32:
; CHECK-SSE2: # %bb.0:
; CHECK-SSE2-NEXT: andps %xmm2, %xmm0
; CHECK-SSE2-NEXT: xorps {{.*}}(%rip), %xmm2
; CHECK-SSE2-NEXT: andps %xmm1, %xmm2
; CHECK-SSE2-NEXT: orps %xmm2, %xmm0
; CHECK-SSE2-NEXT: retq
;
; CHECK-XOP-LABEL: out_v2i32:
; CHECK-XOP: # %bb.0:
; CHECK-XOP-NEXT: vandps %xmm2, %xmm0, %xmm0
; CHECK-XOP-NEXT: vxorps {{.*}}(%rip), %xmm2, %xmm2
; CHECK-XOP-NEXT: vandps %xmm2, %xmm1, %xmm1
; CHECK-XOP-NEXT: vorps %xmm1, %xmm0, %xmm0
; CHECK-XOP-NEXT: retq
%mx = and <2 x i32> %x, %mask
%notmask = xor <2 x i32> %mask, <i32 -1, i32 -1>
%my = and <2 x i32> %y, %notmask
%r = or <2 x i32> %mx, %my
ret <2 x i32> %r
}
define <1 x i64> @out_v1i64(<1 x i64> %x, <1 x i64> %y, <1 x i64> %mask) nounwind {
; CHECK-LABEL: out_v1i64:
; CHECK: # %bb.0:
; CHECK-NEXT: movq %rdx, %rax
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-NEXT: andq %rdx, %rdi
; CHECK-NEXT: notq %rax
; CHECK-NEXT: andq %rsi, %rax
; CHECK-NEXT: orq %rdi, %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-NEXT: retq
%mx = and <1 x i64> %x, %mask
%notmask = xor <1 x i64> %mask, <i64 -1>
%my = and <1 x i64> %y, %notmask
%r = or <1 x i64> %mx, %my
ret <1 x i64> %r
}
; ============================================================================ ;
; 128-bit vector width
; ============================================================================ ;
define <16 x i8> @out_v16i8(<16 x i8> %x, <16 x i8> %y, <16 x i8> %mask) nounwind {
; CHECK-BASELINE-LABEL: out_v16i8:
; CHECK-BASELINE: # %bb.0:
; CHECK-BASELINE-NEXT: pushq %rbp
; CHECK-BASELINE-NEXT: pushq %r15
; CHECK-BASELINE-NEXT: pushq %r14
; CHECK-BASELINE-NEXT: pushq %r13
; CHECK-BASELINE-NEXT: pushq %r12
; CHECK-BASELINE-NEXT: pushq %rbx
; CHECK-BASELINE-NEXT: movl %ecx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-BASELINE-NEXT: movl %edx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-BASELINE-NEXT: movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movq %rdi, %rax
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %dil
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %r10b
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %r11b
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %bpl
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %r14b
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %r15b
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %r12b
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %r13b
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %bl
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %dl
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %cl
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %sil
; CHECK-BASELINE-NEXT: andb %cl, %sil
; CHECK-BASELINE-NEXT: notb %cl
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %cl
; CHECK-BASELINE-NEXT: orb %sil, %cl
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %sil
; CHECK-BASELINE-NEXT: andb %dl, %sil
; CHECK-BASELINE-NEXT: notb %dl
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %dl
; CHECK-BASELINE-NEXT: orb %sil, %dl
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %sil
; CHECK-BASELINE-NEXT: andb %bl, %sil
; CHECK-BASELINE-NEXT: notb %bl
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %bl
; CHECK-BASELINE-NEXT: orb %sil, %bl
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %sil
; CHECK-BASELINE-NEXT: andb %r13b, %sil
; CHECK-BASELINE-NEXT: notb %r13b
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %r13b
; CHECK-BASELINE-NEXT: orb %sil, %r13b
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %sil
; CHECK-BASELINE-NEXT: andb %r12b, %sil
; CHECK-BASELINE-NEXT: notb %r12b
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %r12b
; CHECK-BASELINE-NEXT: orb %sil, %r12b
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %sil
; CHECK-BASELINE-NEXT: andb %r15b, %sil
; CHECK-BASELINE-NEXT: notb %r15b
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %r15b
; CHECK-BASELINE-NEXT: orb %sil, %r15b
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %sil
; CHECK-BASELINE-NEXT: andb %r14b, %sil
; CHECK-BASELINE-NEXT: notb %r14b
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %r14b
; CHECK-BASELINE-NEXT: orb %sil, %r14b
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %sil
; CHECK-BASELINE-NEXT: andb %bpl, %sil
; CHECK-BASELINE-NEXT: notb %bpl
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %bpl
; CHECK-BASELINE-NEXT: orb %sil, %bpl
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %sil
; CHECK-BASELINE-NEXT: andb %r11b, %sil
; CHECK-BASELINE-NEXT: notb %r11b
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %r11b
; CHECK-BASELINE-NEXT: orb %sil, %r11b
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %sil
; CHECK-BASELINE-NEXT: andb %r10b, %sil
; CHECK-BASELINE-NEXT: notb %r10b
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %r10b
; CHECK-BASELINE-NEXT: orb %sil, %r10b
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %sil
; CHECK-BASELINE-NEXT: andb %dil, %sil
; CHECK-BASELINE-NEXT: notb %dil
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %dil
; CHECK-BASELINE-NEXT: orb %sil, %dil
; CHECK-BASELINE-NEXT: movb %cl, 15(%rax)
; CHECK-BASELINE-NEXT: movb %dl, 14(%rax)
; CHECK-BASELINE-NEXT: movb %bl, 13(%rax)
; CHECK-BASELINE-NEXT: movb %r13b, 12(%rax)
; CHECK-BASELINE-NEXT: movb %r12b, 11(%rax)
; CHECK-BASELINE-NEXT: movb %r15b, 10(%rax)
; CHECK-BASELINE-NEXT: movb %r14b, 9(%rax)
; CHECK-BASELINE-NEXT: movb %bpl, 8(%rax)
; CHECK-BASELINE-NEXT: movb %r11b, 7(%rax)
; CHECK-BASELINE-NEXT: movb %r10b, 6(%rax)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %cl
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: andb %cl, %r9b
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: notb %cl
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %cl
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: orb %r9b, %cl
; CHECK-BASELINE-NEXT: movb %dil, 5(%rax)
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %dl
; CHECK-BASELINE-NEXT: andb %dl, %r8b
; CHECK-BASELINE-NEXT: notb %dl
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %dl
; CHECK-BASELINE-NEXT: orb %r8b, %dl
; CHECK-BASELINE-NEXT: movb %cl, 4(%rax)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %cl
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %esi # 4-byte Reload
; CHECK-BASELINE-NEXT: andb %cl, %sil
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: notb %cl
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %cl
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: orb %sil, %cl
; CHECK-BASELINE-NEXT: movb %dl, 3(%rax)
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %dl
; CHECK-BASELINE-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %esi # 4-byte Reload
; CHECK-BASELINE-NEXT: andb %dl, %sil
; CHECK-BASELINE-NEXT: notb %dl
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %dl
; CHECK-BASELINE-NEXT: orb %sil, %dl
; CHECK-BASELINE-NEXT: movb %cl, 2(%rax)
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %cl
; CHECK-BASELINE-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %esi # 4-byte Reload
; CHECK-BASELINE-NEXT: andb %cl, %sil
; CHECK-BASELINE-NEXT: notb %cl
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %cl
; CHECK-BASELINE-NEXT: orb %sil, %cl
; CHECK-BASELINE-NEXT: movb %dl, 1(%rax)
; CHECK-BASELINE-NEXT: movb %cl, (%rax)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: popq %rbx
; CHECK-BASELINE-NEXT: popq %r12
; CHECK-BASELINE-NEXT: popq %r13
; CHECK-BASELINE-NEXT: popq %r14
; CHECK-BASELINE-NEXT: popq %r15
; CHECK-BASELINE-NEXT: popq %rbp
; CHECK-BASELINE-NEXT: retq
;
; CHECK-SSE1-LABEL: out_v16i8:
; CHECK-SSE1: # %bb.0:
; CHECK-SSE1-NEXT: pushq %rbp
; CHECK-SSE1-NEXT: pushq %r15
; CHECK-SSE1-NEXT: pushq %r14
; CHECK-SSE1-NEXT: pushq %r13
; CHECK-SSE1-NEXT: pushq %r12
; CHECK-SSE1-NEXT: pushq %rbx
; CHECK-SSE1-NEXT: movl %ecx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-SSE1-NEXT: movl %edx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-SSE1-NEXT: movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movq %rdi, %rax
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %dil
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %r10b
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %r11b
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %bpl
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %r14b
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %r15b
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %r12b
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %r13b
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %bl
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %dl
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %cl
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %sil
; CHECK-SSE1-NEXT: andb %cl, %sil
; CHECK-SSE1-NEXT: notb %cl
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %cl
; CHECK-SSE1-NEXT: orb %sil, %cl
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %sil
; CHECK-SSE1-NEXT: andb %dl, %sil
; CHECK-SSE1-NEXT: notb %dl
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %dl
; CHECK-SSE1-NEXT: orb %sil, %dl
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %sil
; CHECK-SSE1-NEXT: andb %bl, %sil
; CHECK-SSE1-NEXT: notb %bl
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %bl
; CHECK-SSE1-NEXT: orb %sil, %bl
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %sil
; CHECK-SSE1-NEXT: andb %r13b, %sil
; CHECK-SSE1-NEXT: notb %r13b
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %r13b
; CHECK-SSE1-NEXT: orb %sil, %r13b
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %sil
; CHECK-SSE1-NEXT: andb %r12b, %sil
; CHECK-SSE1-NEXT: notb %r12b
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %r12b
; CHECK-SSE1-NEXT: orb %sil, %r12b
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %sil
; CHECK-SSE1-NEXT: andb %r15b, %sil
; CHECK-SSE1-NEXT: notb %r15b
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %r15b
; CHECK-SSE1-NEXT: orb %sil, %r15b
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %sil
; CHECK-SSE1-NEXT: andb %r14b, %sil
; CHECK-SSE1-NEXT: notb %r14b
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %r14b
; CHECK-SSE1-NEXT: orb %sil, %r14b
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %sil
; CHECK-SSE1-NEXT: andb %bpl, %sil
; CHECK-SSE1-NEXT: notb %bpl
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %bpl
; CHECK-SSE1-NEXT: orb %sil, %bpl
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %sil
; CHECK-SSE1-NEXT: andb %r11b, %sil
; CHECK-SSE1-NEXT: notb %r11b
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %r11b
; CHECK-SSE1-NEXT: orb %sil, %r11b
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %sil
; CHECK-SSE1-NEXT: andb %r10b, %sil
; CHECK-SSE1-NEXT: notb %r10b
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %r10b
; CHECK-SSE1-NEXT: orb %sil, %r10b
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %sil
; CHECK-SSE1-NEXT: andb %dil, %sil
; CHECK-SSE1-NEXT: notb %dil
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %dil
; CHECK-SSE1-NEXT: orb %sil, %dil
; CHECK-SSE1-NEXT: movb %cl, 15(%rax)
; CHECK-SSE1-NEXT: movb %dl, 14(%rax)
; CHECK-SSE1-NEXT: movb %bl, 13(%rax)
; CHECK-SSE1-NEXT: movb %r13b, 12(%rax)
; CHECK-SSE1-NEXT: movb %r12b, 11(%rax)
; CHECK-SSE1-NEXT: movb %r15b, 10(%rax)
; CHECK-SSE1-NEXT: movb %r14b, 9(%rax)
; CHECK-SSE1-NEXT: movb %bpl, 8(%rax)
; CHECK-SSE1-NEXT: movb %r11b, 7(%rax)
; CHECK-SSE1-NEXT: movb %r10b, 6(%rax)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %cl
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: andb %cl, %r9b
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: notb %cl
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %cl
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: orb %r9b, %cl
; CHECK-SSE1-NEXT: movb %dil, 5(%rax)
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %dl
; CHECK-SSE1-NEXT: andb %dl, %r8b
; CHECK-SSE1-NEXT: notb %dl
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %dl
; CHECK-SSE1-NEXT: orb %r8b, %dl
; CHECK-SSE1-NEXT: movb %cl, 4(%rax)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %cl
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %esi # 4-byte Reload
; CHECK-SSE1-NEXT: andb %cl, %sil
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: notb %cl
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %cl
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: orb %sil, %cl
; CHECK-SSE1-NEXT: movb %dl, 3(%rax)
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %dl
; CHECK-SSE1-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %esi # 4-byte Reload
; CHECK-SSE1-NEXT: andb %dl, %sil
; CHECK-SSE1-NEXT: notb %dl
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %dl
; CHECK-SSE1-NEXT: orb %sil, %dl
; CHECK-SSE1-NEXT: movb %cl, 2(%rax)
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %cl
; CHECK-SSE1-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %esi # 4-byte Reload
; CHECK-SSE1-NEXT: andb %cl, %sil
; CHECK-SSE1-NEXT: notb %cl
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %cl
; CHECK-SSE1-NEXT: orb %sil, %cl
; CHECK-SSE1-NEXT: movb %dl, 1(%rax)
; CHECK-SSE1-NEXT: movb %cl, (%rax)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: popq %rbx
; CHECK-SSE1-NEXT: popq %r12
; CHECK-SSE1-NEXT: popq %r13
; CHECK-SSE1-NEXT: popq %r14
; CHECK-SSE1-NEXT: popq %r15
; CHECK-SSE1-NEXT: popq %rbp
; CHECK-SSE1-NEXT: retq
;
; CHECK-SSE2-LABEL: out_v16i8:
; CHECK-SSE2: # %bb.0:
; CHECK-SSE2-NEXT: andps %xmm2, %xmm0
; CHECK-SSE2-NEXT: andnps %xmm1, %xmm2
; CHECK-SSE2-NEXT: orps %xmm2, %xmm0
; CHECK-SSE2-NEXT: retq
;
; CHECK-XOP-LABEL: out_v16i8:
; CHECK-XOP: # %bb.0:
; CHECK-XOP-NEXT: vpcmov %xmm2, %xmm1, %xmm0, %xmm0
; CHECK-XOP-NEXT: retq
%mx = and <16 x i8> %x, %mask
%notmask = xor <16 x i8> %mask, <i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1>
%my = and <16 x i8> %y, %notmask
%r = or <16 x i8> %mx, %my
ret <16 x i8> %r
}
define <8 x i16> @out_v8i16(<8 x i16> %x, <8 x i16> %y, <8 x i16> %mask) nounwind {
; CHECK-BASELINE-LABEL: out_v8i16:
; CHECK-BASELINE: # %bb.0:
; CHECK-BASELINE-NEXT: pushq %rbp
; CHECK-BASELINE-NEXT: pushq %r14
; CHECK-BASELINE-NEXT: pushq %rbx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movq %rdi, %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl {{[0-9]+}}(%rsp), %r10d
; CHECK-BASELINE-NEXT: movl {{[0-9]+}}(%rsp), %r11d
; CHECK-BASELINE-NEXT: movl {{[0-9]+}}(%rsp), %r14d
; CHECK-BASELINE-NEXT: movzwl {{[0-9]+}}(%rsp), %ebx
; CHECK-BASELINE-NEXT: andw %r14w, %bx
; CHECK-BASELINE-NEXT: notl %r14d
; CHECK-BASELINE-NEXT: andw {{[0-9]+}}(%rsp), %r14w
; CHECK-BASELINE-NEXT: orl %ebx, %r14d
; CHECK-BASELINE-NEXT: movzwl {{[0-9]+}}(%rsp), %ebx
; CHECK-BASELINE-NEXT: andw %r11w, %bx
; CHECK-BASELINE-NEXT: notl %r11d
; CHECK-BASELINE-NEXT: andw {{[0-9]+}}(%rsp), %r11w
; CHECK-BASELINE-NEXT: orl %ebx, %r11d
; CHECK-BASELINE-NEXT: movzwl {{[0-9]+}}(%rsp), %ebx
; CHECK-BASELINE-NEXT: andw %r10w, %bx
; CHECK-BASELINE-NEXT: notl %r10d
; CHECK-BASELINE-NEXT: andw {{[0-9]+}}(%rsp), %r10w
; CHECK-BASELINE-NEXT: orl %ebx, %r10d
; CHECK-BASELINE-NEXT: movl {{[0-9]+}}(%rsp), %ebx
; CHECK-BASELINE-NEXT: andl %ebx, %r9d
; CHECK-BASELINE-NEXT: notl %ebx
; CHECK-BASELINE-NEXT: andw {{[0-9]+}}(%rsp), %bx
; CHECK-BASELINE-NEXT: orl %r9d, %ebx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movl {{[0-9]+}}(%rsp), %edi
; CHECK-BASELINE-NEXT: andl %edi, %r8d
; CHECK-BASELINE-NEXT: notl %edi
; CHECK-BASELINE-NEXT: andw {{[0-9]+}}(%rsp), %di
; CHECK-BASELINE-NEXT: orl %r8d, %edi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl {{[0-9]+}}(%rsp), %ebp
; CHECK-BASELINE-NEXT: andl %ebp, %ecx
; CHECK-BASELINE-NEXT: notl %ebp
; CHECK-BASELINE-NEXT: andw {{[0-9]+}}(%rsp), %bp
; CHECK-BASELINE-NEXT: orl %ecx, %ebp
; CHECK-BASELINE-NEXT: movl {{[0-9]+}}(%rsp), %ecx
; CHECK-BASELINE-NEXT: andl %ecx, %edx
; CHECK-BASELINE-NEXT: notl %ecx
; CHECK-BASELINE-NEXT: andw {{[0-9]+}}(%rsp), %cx
; CHECK-BASELINE-NEXT: orl %edx, %ecx
; CHECK-BASELINE-NEXT: movl {{[0-9]+}}(%rsp), %edx
; CHECK-BASELINE-NEXT: andl %edx, %esi
; CHECK-BASELINE-NEXT: notl %edx
; CHECK-BASELINE-NEXT: andw {{[0-9]+}}(%rsp), %dx
; CHECK-BASELINE-NEXT: orl %esi, %edx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movw %r14w, 14(%rax)
; CHECK-BASELINE-NEXT: movw %r11w, 12(%rax)
; CHECK-BASELINE-NEXT: movw %r10w, 10(%rax)
; CHECK-BASELINE-NEXT: movw %bx, 8(%rax)
; CHECK-BASELINE-NEXT: movw %di, 6(%rax)
; CHECK-BASELINE-NEXT: movw %bp, 4(%rax)
; CHECK-BASELINE-NEXT: movw %cx, 2(%rax)
; CHECK-BASELINE-NEXT: movw %dx, (%rax)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: popq %rbx
; CHECK-BASELINE-NEXT: popq %r14
; CHECK-BASELINE-NEXT: popq %rbp
; CHECK-BASELINE-NEXT: retq
;
; CHECK-SSE1-LABEL: out_v8i16:
; CHECK-SSE1: # %bb.0:
; CHECK-SSE1-NEXT: pushq %rbp
; CHECK-SSE1-NEXT: pushq %r14
; CHECK-SSE1-NEXT: pushq %rbx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movq %rdi, %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movl {{[0-9]+}}(%rsp), %r10d
; CHECK-SSE1-NEXT: movl {{[0-9]+}}(%rsp), %r11d
; CHECK-SSE1-NEXT: movl {{[0-9]+}}(%rsp), %r14d
; CHECK-SSE1-NEXT: movzwl {{[0-9]+}}(%rsp), %ebx
; CHECK-SSE1-NEXT: andw %r14w, %bx
; CHECK-SSE1-NEXT: notl %r14d
; CHECK-SSE1-NEXT: andw {{[0-9]+}}(%rsp), %r14w
; CHECK-SSE1-NEXT: orl %ebx, %r14d
; CHECK-SSE1-NEXT: movzwl {{[0-9]+}}(%rsp), %ebx
; CHECK-SSE1-NEXT: andw %r11w, %bx
; CHECK-SSE1-NEXT: notl %r11d
; CHECK-SSE1-NEXT: andw {{[0-9]+}}(%rsp), %r11w
; CHECK-SSE1-NEXT: orl %ebx, %r11d
; CHECK-SSE1-NEXT: movzwl {{[0-9]+}}(%rsp), %ebx
; CHECK-SSE1-NEXT: andw %r10w, %bx
; CHECK-SSE1-NEXT: notl %r10d
; CHECK-SSE1-NEXT: andw {{[0-9]+}}(%rsp), %r10w
; CHECK-SSE1-NEXT: orl %ebx, %r10d
; CHECK-SSE1-NEXT: movl {{[0-9]+}}(%rsp), %ebx
; CHECK-SSE1-NEXT: andl %ebx, %r9d
; CHECK-SSE1-NEXT: notl %ebx
; CHECK-SSE1-NEXT: andw {{[0-9]+}}(%rsp), %bx
; CHECK-SSE1-NEXT: orl %r9d, %ebx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movl {{[0-9]+}}(%rsp), %edi
; CHECK-SSE1-NEXT: andl %edi, %r8d
; CHECK-SSE1-NEXT: notl %edi
; CHECK-SSE1-NEXT: andw {{[0-9]+}}(%rsp), %di
; CHECK-SSE1-NEXT: orl %r8d, %edi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movl {{[0-9]+}}(%rsp), %ebp
; CHECK-SSE1-NEXT: andl %ebp, %ecx
; CHECK-SSE1-NEXT: notl %ebp
; CHECK-SSE1-NEXT: andw {{[0-9]+}}(%rsp), %bp
; CHECK-SSE1-NEXT: orl %ecx, %ebp
; CHECK-SSE1-NEXT: movl {{[0-9]+}}(%rsp), %ecx
; CHECK-SSE1-NEXT: andl %ecx, %edx
; CHECK-SSE1-NEXT: notl %ecx
; CHECK-SSE1-NEXT: andw {{[0-9]+}}(%rsp), %cx
; CHECK-SSE1-NEXT: orl %edx, %ecx
; CHECK-SSE1-NEXT: movl {{[0-9]+}}(%rsp), %edx
; CHECK-SSE1-NEXT: andl %edx, %esi
; CHECK-SSE1-NEXT: notl %edx
; CHECK-SSE1-NEXT: andw {{[0-9]+}}(%rsp), %dx
; CHECK-SSE1-NEXT: orl %esi, %edx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movw %r14w, 14(%rax)
; CHECK-SSE1-NEXT: movw %r11w, 12(%rax)
; CHECK-SSE1-NEXT: movw %r10w, 10(%rax)
; CHECK-SSE1-NEXT: movw %bx, 8(%rax)
; CHECK-SSE1-NEXT: movw %di, 6(%rax)
; CHECK-SSE1-NEXT: movw %bp, 4(%rax)
; CHECK-SSE1-NEXT: movw %cx, 2(%rax)
; CHECK-SSE1-NEXT: movw %dx, (%rax)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: popq %rbx
; CHECK-SSE1-NEXT: popq %r14
; CHECK-SSE1-NEXT: popq %rbp
; CHECK-SSE1-NEXT: retq
;
; CHECK-SSE2-LABEL: out_v8i16:
; CHECK-SSE2: # %bb.0:
; CHECK-SSE2-NEXT: andps %xmm2, %xmm0
; CHECK-SSE2-NEXT: andnps %xmm1, %xmm2
; CHECK-SSE2-NEXT: orps %xmm2, %xmm0
; CHECK-SSE2-NEXT: retq
;
; CHECK-XOP-LABEL: out_v8i16:
; CHECK-XOP: # %bb.0:
; CHECK-XOP-NEXT: vpcmov %xmm2, %xmm1, %xmm0, %xmm0
; CHECK-XOP-NEXT: retq
%mx = and <8 x i16> %x, %mask
%notmask = xor <8 x i16> %mask, <i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1>
%my = and <8 x i16> %y, %notmask
%r = or <8 x i16> %mx, %my
ret <8 x i16> %r
}
define <4 x i32> @out_v4i32(<4 x i32> *%px, <4 x i32> *%py, <4 x i32> *%pmask) nounwind {
; CHECK-BASELINE-LABEL: out_v4i32:
; CHECK-BASELINE: # %bb.0:
; CHECK-BASELINE-NEXT: pushq %rbx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movq %rdi, %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl (%rcx), %r8d
; CHECK-BASELINE-NEXT: movl 4(%rcx), %r9d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movl 8(%rcx), %edi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl 12(%rcx), %ecx
; CHECK-BASELINE-NEXT: movl 12(%rsi), %r10d
; CHECK-BASELINE-NEXT: andl %ecx, %r10d
; CHECK-BASELINE-NEXT: movl 8(%rsi), %r11d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: andl %edi, %r11d
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl 4(%rsi), %ebx
; CHECK-BASELINE-NEXT: andl %r9d, %ebx
; CHECK-BASELINE-NEXT: movl (%rsi), %esi
; CHECK-BASELINE-NEXT: andl %r8d, %esi
; CHECK-BASELINE-NEXT: notl %r8d
; CHECK-BASELINE-NEXT: notl %r9d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: notl %edi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: notl %ecx
; CHECK-BASELINE-NEXT: andl 12(%rdx), %ecx
; CHECK-BASELINE-NEXT: orl %r10d, %ecx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: andl 8(%rdx), %edi
; CHECK-BASELINE-NEXT: orl %r11d, %edi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andl 4(%rdx), %r9d
; CHECK-BASELINE-NEXT: orl %ebx, %r9d
; CHECK-BASELINE-NEXT: andl (%rdx), %r8d
; CHECK-BASELINE-NEXT: orl %esi, %r8d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movl %ecx, 12(%rax)
; CHECK-BASELINE-NEXT: movl %edi, 8(%rax)
; CHECK-BASELINE-NEXT: movl %r9d, 4(%rax)
; CHECK-BASELINE-NEXT: movl %r8d, (%rax)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: popq %rbx
; CHECK-BASELINE-NEXT: retq
;
; CHECK-SSE1-LABEL: out_v4i32:
; CHECK-SSE1: # %bb.0:
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movq %rdi, %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movaps (%rcx), %xmm0
; CHECK-SSE1-NEXT: movaps (%rsi), %xmm1
; CHECK-SSE1-NEXT: andps %xmm0, %xmm1
; CHECK-SSE1-NEXT: andnps (%rdx), %xmm0
; CHECK-SSE1-NEXT: orps %xmm1, %xmm0
; CHECK-SSE1-NEXT: movaps %xmm0, (%rdi)
; CHECK-SSE1-NEXT: retq
;
; CHECK-SSE2-LABEL: out_v4i32:
; CHECK-SSE2: # %bb.0:
; CHECK-SSE2-NEXT: movaps (%rdx), %xmm0
; CHECK-SSE2-NEXT: movaps (%rdi), %xmm1
; CHECK-SSE2-NEXT: andps %xmm0, %xmm1
; CHECK-SSE2-NEXT: andnps (%rsi), %xmm0
; CHECK-SSE2-NEXT: orps %xmm1, %xmm0
; CHECK-SSE2-NEXT: retq
;
; CHECK-XOP-LABEL: out_v4i32:
; CHECK-XOP: # %bb.0:
; CHECK-XOP-NEXT: vmovdqa (%rdi), %xmm0
; CHECK-XOP-NEXT: vmovdqa (%rdx), %xmm1
; CHECK-XOP-NEXT: vpcmov %xmm1, (%rsi), %xmm0, %xmm0
; CHECK-XOP-NEXT: retq
%x = load <4 x i32>, <4 x i32> *%px, align 16
%y = load <4 x i32>, <4 x i32> *%py, align 16
%mask = load <4 x i32>, <4 x i32> *%pmask, align 16
%mx = and <4 x i32> %x, %mask
%notmask = xor <4 x i32> %mask, <i32 -1, i32 -1, i32 -1, i32 -1>
%my = and <4 x i32> %y, %notmask
%r = or <4 x i32> %mx, %my
ret <4 x i32> %r
}
define <4 x i32> @out_v4i32_undef(<4 x i32> *%px, <4 x i32> *%py, <4 x i32> *%pmask) nounwind {
; CHECK-BASELINE-LABEL: out_v4i32_undef:
; CHECK-BASELINE: # %bb.0:
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movq %rdi, %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl 8(%rsi), %r8d
; CHECK-BASELINE-NEXT: movl (%rcx), %r9d
; CHECK-BASELINE-NEXT: movl 4(%rcx), %r10d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movl 12(%rcx), %edi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andl 8(%rcx), %r8d
; CHECK-BASELINE-NEXT: movl 12(%rsi), %ecx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: andl %edi, %ecx
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl 4(%rsi), %r11d
; CHECK-BASELINE-NEXT: andl %r10d, %r11d
; CHECK-BASELINE-NEXT: movl (%rsi), %esi
; CHECK-BASELINE-NEXT: andl %r9d, %esi
; CHECK-BASELINE-NEXT: notl %r9d
; CHECK-BASELINE-NEXT: notl %r10d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: notl %edi
; CHECK-BASELINE-NEXT: andl 12(%rdx), %edi
; CHECK-BASELINE-NEXT: orl %ecx, %edi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andl 4(%rdx), %r10d
; CHECK-BASELINE-NEXT: orl %r11d, %r10d
; CHECK-BASELINE-NEXT: andl (%rdx), %r9d
; CHECK-BASELINE-NEXT: orl %esi, %r9d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movl %r8d, 8(%rax)
; CHECK-BASELINE-NEXT: movl %edi, 12(%rax)
; CHECK-BASELINE-NEXT: movl %r10d, 4(%rax)
; CHECK-BASELINE-NEXT: movl %r9d, (%rax)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: retq
;
; CHECK-SSE1-LABEL: out_v4i32_undef:
; CHECK-SSE1: # %bb.0:
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movq %rdi, %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movaps (%rcx), %xmm0
; CHECK-SSE1-NEXT: movaps (%rsi), %xmm1
; CHECK-SSE1-NEXT: andps %xmm0, %xmm1
; CHECK-SSE1-NEXT: andnps (%rdx), %xmm0
; CHECK-SSE1-NEXT: orps %xmm1, %xmm0
; CHECK-SSE1-NEXT: movaps %xmm0, (%rdi)
; CHECK-SSE1-NEXT: retq
;
; CHECK-SSE2-LABEL: out_v4i32_undef:
; CHECK-SSE2: # %bb.0:
; CHECK-SSE2-NEXT: movaps (%rdx), %xmm0
; CHECK-SSE2-NEXT: movaps (%rdi), %xmm1
; CHECK-SSE2-NEXT: andps %xmm0, %xmm1
; CHECK-SSE2-NEXT: andnps (%rsi), %xmm0
; CHECK-SSE2-NEXT: orps %xmm1, %xmm0
; CHECK-SSE2-NEXT: retq
;
; CHECK-XOP-LABEL: out_v4i32_undef:
; CHECK-XOP: # %bb.0:
; CHECK-XOP-NEXT: vmovdqa (%rdi), %xmm0
; CHECK-XOP-NEXT: vmovdqa (%rdx), %xmm1
; CHECK-XOP-NEXT: vpcmov %xmm1, (%rsi), %xmm0, %xmm0
; CHECK-XOP-NEXT: retq
%x = load <4 x i32>, <4 x i32> *%px, align 16
%y = load <4 x i32>, <4 x i32> *%py, align 16
%mask = load <4 x i32>, <4 x i32> *%pmask, align 16
%mx = and <4 x i32> %x, %mask
%notmask = xor <4 x i32> %mask, <i32 -1, i32 -1, i32 undef, i32 -1>
%my = and <4 x i32> %y, %notmask
%r = or <4 x i32> %mx, %my
ret <4 x i32> %r
}
define <2 x i64> @out_v2i64(<2 x i64> %x, <2 x i64> %y, <2 x i64> %mask) nounwind {
; CHECK-BASELINE-LABEL: out_v2i64:
; CHECK-BASELINE: # %bb.0:
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movq %r8, %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andq %r9, %rsi
; CHECK-BASELINE-NEXT: andq %r8, %rdi
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: notq %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: notq %r9
; CHECK-BASELINE-NEXT: andq %rcx, %r9
; CHECK-BASELINE-NEXT: orq %rsi, %r9
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: andq %rdx, %rax
; CHECK-BASELINE-NEXT: orq %rdi, %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movq %r9, %rdx
; CHECK-BASELINE-NEXT: retq
;
; CHECK-SSE1-LABEL: out_v2i64:
; CHECK-SSE1: # %bb.0:
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movq %r8, %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: andq %r9, %rsi
; CHECK-SSE1-NEXT: andq %r8, %rdi
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: notq %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: notq %r9
; CHECK-SSE1-NEXT: andq %rcx, %r9
; CHECK-SSE1-NEXT: orq %rsi, %r9
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: andq %rdx, %rax
; CHECK-SSE1-NEXT: orq %rdi, %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movq %r9, %rdx
; CHECK-SSE1-NEXT: retq
;
; CHECK-SSE2-LABEL: out_v2i64:
; CHECK-SSE2: # %bb.0:
; CHECK-SSE2-NEXT: andps %xmm2, %xmm0
; CHECK-SSE2-NEXT: andnps %xmm1, %xmm2
; CHECK-SSE2-NEXT: orps %xmm2, %xmm0
; CHECK-SSE2-NEXT: retq
;
; CHECK-XOP-LABEL: out_v2i64:
; CHECK-XOP: # %bb.0:
; CHECK-XOP-NEXT: vpcmov %xmm2, %xmm1, %xmm0, %xmm0
; CHECK-XOP-NEXT: retq
%mx = and <2 x i64> %x, %mask
%notmask = xor <2 x i64> %mask, <i64 -1, i64 -1>
%my = and <2 x i64> %y, %notmask
%r = or <2 x i64> %mx, %my
ret <2 x i64> %r
}
; ============================================================================ ;
; 256-bit vector width
; ============================================================================ ;
define <32 x i8> @out_v32i8(<32 x i8> *%px, <32 x i8> *%py, <32 x i8> *%pmask) nounwind {
; CHECK-BASELINE-LABEL: out_v32i8:
; CHECK-BASELINE: # %bb.0:
; CHECK-BASELINE-NEXT: pushq %rbp
; CHECK-BASELINE-NEXT: pushq %r15
; CHECK-BASELINE-NEXT: pushq %r14
; CHECK-BASELINE-NEXT: pushq %r13
; CHECK-BASELINE-NEXT: pushq %r12
; CHECK-BASELINE-NEXT: pushq %rbx
; CHECK-BASELINE-NEXT: movq %rcx, %r15
; CHECK-BASELINE-NEXT: movq %rsi, %r14
; CHECK-BASELINE-NEXT: movq %rdi, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
; CHECK-BASELINE-NEXT: movb 16(%rcx), %al
; CHECK-BASELINE-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 17(%rcx), %al
; CHECK-BASELINE-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 18(%rcx), %al
; CHECK-BASELINE-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 19(%rcx), %al
; CHECK-BASELINE-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 20(%rcx), %al
; CHECK-BASELINE-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 21(%rcx), %r12b
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-BASELINE-NEXT: movb 22(%rcx), %r9b
; CHECK-BASELINE-NEXT: movb 23(%rcx), %r10b
; CHECK-BASELINE-NEXT: movb 24(%rcx), %r11b
; CHECK-BASELINE-NEXT: movb 25(%rcx), %bpl
; CHECK-BASELINE-NEXT: movb 26(%rcx), %r13b
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb 27(%rcx), %r8b
; CHECK-BASELINE-NEXT: movb 28(%rcx), %dil
; CHECK-BASELINE-NEXT: movb 29(%rcx), %sil
; CHECK-BASELINE-NEXT: movb 30(%rcx), %bl
; CHECK-BASELINE-NEXT: movb 31(%rcx), %al
; CHECK-BASELINE-NEXT: movb 31(%r14), %cl
; CHECK-BASELINE-NEXT: andb %al, %cl
; CHECK-BASELINE-NEXT: notb %al
; CHECK-BASELINE-NEXT: andb 31(%rdx), %al
; CHECK-BASELINE-NEXT: orb %cl, %al
; CHECK-BASELINE-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 30(%r14), %al
; CHECK-BASELINE-NEXT: andb %bl, %al
; CHECK-BASELINE-NEXT: notb %bl
; CHECK-BASELINE-NEXT: andb 30(%rdx), %bl
; CHECK-BASELINE-NEXT: orb %al, %bl
; CHECK-BASELINE-NEXT: movb %bl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 29(%r14), %al
; CHECK-BASELINE-NEXT: andb %sil, %al
; CHECK-BASELINE-NEXT: notb %sil
; CHECK-BASELINE-NEXT: andb 29(%rdx), %sil
; CHECK-BASELINE-NEXT: orb %al, %sil
; CHECK-BASELINE-NEXT: movb %sil, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 28(%r14), %al
; CHECK-BASELINE-NEXT: andb %dil, %al
; CHECK-BASELINE-NEXT: notb %dil
; CHECK-BASELINE-NEXT: andb 28(%rdx), %dil
; CHECK-BASELINE-NEXT: orb %al, %dil
; CHECK-BASELINE-NEXT: movb %dil, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 27(%r14), %al
; CHECK-BASELINE-NEXT: andb %r8b, %al
; CHECK-BASELINE-NEXT: notb %r8b
; CHECK-BASELINE-NEXT: andb 27(%rdx), %r8b
; CHECK-BASELINE-NEXT: orb %al, %r8b
; CHECK-BASELINE-NEXT: movb %r8b, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 26(%r14), %al
; CHECK-BASELINE-NEXT: andb %r13b, %al
; CHECK-BASELINE-NEXT: notb %r13b
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-BASELINE-NEXT: andb 26(%rdx), %r13b
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: orb %al, %r13b
; CHECK-BASELINE-NEXT: movb %r13b, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-BASELINE-NEXT: movb 25(%r14), %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andb %bpl, %al
; CHECK-BASELINE-NEXT: notb %bpl
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-BASELINE-NEXT: andb 25(%rdx), %bpl
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: orb %al, %bpl
; CHECK-BASELINE-NEXT: movb %bpl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-BASELINE-NEXT: movb 24(%r14), %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andb %r11b, %al
; CHECK-BASELINE-NEXT: notb %r11b
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-BASELINE-NEXT: andb 24(%rdx), %r11b
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: orb %al, %r11b
; CHECK-BASELINE-NEXT: movb %r11b, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-BASELINE-NEXT: movb 23(%r14), %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andb %r10b, %al
; CHECK-BASELINE-NEXT: notb %r10b
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-BASELINE-NEXT: andb 23(%rdx), %r10b
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: orb %al, %r10b
; CHECK-BASELINE-NEXT: movb %r10b, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-BASELINE-NEXT: movb 22(%r14), %al
; CHECK-BASELINE-NEXT: andb %r9b, %al
; CHECK-BASELINE-NEXT: notb %r9b
; CHECK-BASELINE-NEXT: andb 22(%rdx), %r9b
; CHECK-BASELINE-NEXT: orb %al, %r9b
; CHECK-BASELINE-NEXT: movb %r9b, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb 21(%r14), %al
; CHECK-BASELINE-NEXT: andb %r12b, %al
; CHECK-BASELINE-NEXT: notb %r12b
; CHECK-BASELINE-NEXT: andb 21(%rdx), %r12b
; CHECK-BASELINE-NEXT: orb %al, %r12b
; CHECK-BASELINE-NEXT: movb %r12b, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 20(%r14), %al
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %cl # 1-byte Reload
; CHECK-BASELINE-NEXT: andb %cl, %al
; CHECK-BASELINE-NEXT: notb %cl
; CHECK-BASELINE-NEXT: andb 20(%rdx), %cl
; CHECK-BASELINE-NEXT: orb %al, %cl
; CHECK-BASELINE-NEXT: movb %cl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 19(%r14), %al
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %cl # 1-byte Reload
; CHECK-BASELINE-NEXT: andb %cl, %al
; CHECK-BASELINE-NEXT: notb %cl
; CHECK-BASELINE-NEXT: andb 19(%rdx), %cl
; CHECK-BASELINE-NEXT: orb %al, %cl
; CHECK-BASELINE-NEXT: movb %cl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 18(%r14), %al
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %cl # 1-byte Reload
; CHECK-BASELINE-NEXT: andb %cl, %al
; CHECK-BASELINE-NEXT: notb %cl
; CHECK-BASELINE-NEXT: andb 18(%rdx), %cl
; CHECK-BASELINE-NEXT: orb %al, %cl
; CHECK-BASELINE-NEXT: movb %cl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 17(%r14), %al
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %cl # 1-byte Reload
; CHECK-BASELINE-NEXT: andb %cl, %al
; CHECK-BASELINE-NEXT: notb %cl
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-BASELINE-NEXT: movq %rdx, %rbx
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andb 17(%rdx), %cl
; CHECK-BASELINE-NEXT: orb %al, %cl
; CHECK-BASELINE-NEXT: movb %cl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 16(%r14), %al
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %cl # 1-byte Reload
; CHECK-BASELINE-NEXT: andb %cl, %al
; CHECK-BASELINE-NEXT: notb %cl
; CHECK-BASELINE-NEXT: andb 16(%rdx), %cl
; CHECK-BASELINE-NEXT: orb %al, %cl
; CHECK-BASELINE-NEXT: movb %cl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-BASELINE-NEXT: movb 15(%r15), %cl
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb 15(%r14), %al
; CHECK-BASELINE-NEXT: andb %cl, %al
; CHECK-BASELINE-NEXT: notb %cl
; CHECK-BASELINE-NEXT: andb 15(%rdx), %cl
; CHECK-BASELINE-NEXT: orb %al, %cl
; CHECK-BASELINE-NEXT: movb %cl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 14(%r15), %cl
; CHECK-BASELINE-NEXT: movb 14(%r14), %al
; CHECK-BASELINE-NEXT: andb %cl, %al
; CHECK-BASELINE-NEXT: notb %cl
; CHECK-BASELINE-NEXT: andb 14(%rdx), %cl
; CHECK-BASELINE-NEXT: orb %al, %cl
; CHECK-BASELINE-NEXT: movb %cl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 13(%r15), %cl
; CHECK-BASELINE-NEXT: movb 13(%r14), %al
; CHECK-BASELINE-NEXT: andb %cl, %al
; CHECK-BASELINE-NEXT: notb %cl
; CHECK-BASELINE-NEXT: andb 13(%rdx), %cl
; CHECK-BASELINE-NEXT: orb %al, %cl
; CHECK-BASELINE-NEXT: movb %cl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 12(%r15), %cl
; CHECK-BASELINE-NEXT: movb 12(%r14), %al
; CHECK-BASELINE-NEXT: andb %cl, %al
; CHECK-BASELINE-NEXT: notb %cl
; CHECK-BASELINE-NEXT: andb 12(%rdx), %cl
; CHECK-BASELINE-NEXT: orb %al, %cl
; CHECK-BASELINE-NEXT: movb %cl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 11(%r15), %r13b
; CHECK-BASELINE-NEXT: movb 11(%r14), %al
; CHECK-BASELINE-NEXT: andb %r13b, %al
; CHECK-BASELINE-NEXT: notb %r13b
; CHECK-BASELINE-NEXT: andb 11(%rdx), %r13b
; CHECK-BASELINE-NEXT: orb %al, %r13b
; CHECK-BASELINE-NEXT: movb 10(%r15), %r12b
; CHECK-BASELINE-NEXT: movb 10(%r14), %al
; CHECK-BASELINE-NEXT: andb %r12b, %al
; CHECK-BASELINE-NEXT: notb %r12b
; CHECK-BASELINE-NEXT: andb 10(%rdx), %r12b
; CHECK-BASELINE-NEXT: orb %al, %r12b
; CHECK-BASELINE-NEXT: movb 9(%r15), %bpl
; CHECK-BASELINE-NEXT: movb 9(%r14), %al
; CHECK-BASELINE-NEXT: andb %bpl, %al
; CHECK-BASELINE-NEXT: notb %bpl
; CHECK-BASELINE-NEXT: andb 9(%rdx), %bpl
; CHECK-BASELINE-NEXT: orb %al, %bpl
; CHECK-BASELINE-NEXT: movb 8(%r15), %r11b
; CHECK-BASELINE-NEXT: movb 8(%r14), %al
; CHECK-BASELINE-NEXT: andb %r11b, %al
; CHECK-BASELINE-NEXT: notb %r11b
; CHECK-BASELINE-NEXT: andb 8(%rdx), %r11b
; CHECK-BASELINE-NEXT: orb %al, %r11b
; CHECK-BASELINE-NEXT: movb 7(%r15), %r10b
; CHECK-BASELINE-NEXT: movb 7(%r14), %al
; CHECK-BASELINE-NEXT: andb %r10b, %al
; CHECK-BASELINE-NEXT: notb %r10b
; CHECK-BASELINE-NEXT: andb 7(%rdx), %r10b
; CHECK-BASELINE-NEXT: orb %al, %r10b
; CHECK-BASELINE-NEXT: movb 6(%r15), %r9b
; CHECK-BASELINE-NEXT: movb 6(%r14), %al
; CHECK-BASELINE-NEXT: andb %r9b, %al
; CHECK-BASELINE-NEXT: notb %r9b
; CHECK-BASELINE-NEXT: andb 6(%rdx), %r9b
; CHECK-BASELINE-NEXT: orb %al, %r9b
; CHECK-BASELINE-NEXT: movb 5(%r15), %r8b
; CHECK-BASELINE-NEXT: movb 5(%r14), %al
; CHECK-BASELINE-NEXT: andb %r8b, %al
; CHECK-BASELINE-NEXT: notb %r8b
; CHECK-BASELINE-NEXT: andb 5(%rdx), %r8b
; CHECK-BASELINE-NEXT: orb %al, %r8b
; CHECK-BASELINE-NEXT: movb 4(%r15), %dil
; CHECK-BASELINE-NEXT: movb 4(%r14), %al
; CHECK-BASELINE-NEXT: andb %dil, %al
; CHECK-BASELINE-NEXT: notb %dil
; CHECK-BASELINE-NEXT: andb 4(%rdx), %dil
; CHECK-BASELINE-NEXT: orb %al, %dil
; CHECK-BASELINE-NEXT: movb 3(%r15), %sil
; CHECK-BASELINE-NEXT: movb 3(%r14), %al
; CHECK-BASELINE-NEXT: andb %sil, %al
; CHECK-BASELINE-NEXT: notb %sil
; CHECK-BASELINE-NEXT: andb 3(%rdx), %sil
; CHECK-BASELINE-NEXT: orb %al, %sil
; CHECK-BASELINE-NEXT: movb 2(%r15), %dl
; CHECK-BASELINE-NEXT: movb 2(%r14), %al
; CHECK-BASELINE-NEXT: andb %dl, %al
; CHECK-BASELINE-NEXT: notb %dl
; CHECK-BASELINE-NEXT: andb 2(%rbx), %dl
; CHECK-BASELINE-NEXT: orb %al, %dl
; CHECK-BASELINE-NEXT: movb 1(%r15), %al
; CHECK-BASELINE-NEXT: movb 1(%r14), %cl
; CHECK-BASELINE-NEXT: andb %al, %cl
; CHECK-BASELINE-NEXT: notb %al
; CHECK-BASELINE-NEXT: andb 1(%rbx), %al
; CHECK-BASELINE-NEXT: orb %cl, %al
; CHECK-BASELINE-NEXT: movb (%r15), %r15b
; CHECK-BASELINE-NEXT: movb (%r14), %r14b
; CHECK-BASELINE-NEXT: andb %r15b, %r14b
; CHECK-BASELINE-NEXT: notb %r15b
; CHECK-BASELINE-NEXT: andb (%rbx), %r15b
; CHECK-BASELINE-NEXT: orb %r14b, %r15b
; CHECK-BASELINE-NEXT: movq {{[-0-9]+}}(%r{{[sb]}}p), %rcx # 8-byte Reload
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %bl, 31(%rcx)
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %bl, 30(%rcx)
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %bl, 29(%rcx)
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %bl, 28(%rcx)
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %bl, 27(%rcx)
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %bl, 26(%rcx)
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %bl, 25(%rcx)
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %bl, 24(%rcx)
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %bl, 23(%rcx)
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %bl, 22(%rcx)
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %bl, 21(%rcx)
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %bl, 20(%rcx)
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %bl, 19(%rcx)
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %bl, 18(%rcx)
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %bl, 17(%rcx)
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %bl, 16(%rcx)
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %bl, 15(%rcx)
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %bl, 14(%rcx)
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %bl, 13(%rcx)
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %bl, 12(%rcx)
; CHECK-BASELINE-NEXT: movb %r13b, 11(%rcx)
; CHECK-BASELINE-NEXT: movb %r12b, 10(%rcx)
; CHECK-BASELINE-NEXT: movb %bpl, 9(%rcx)
; CHECK-BASELINE-NEXT: movb %r11b, 8(%rcx)
; CHECK-BASELINE-NEXT: movb %r10b, 7(%rcx)
; CHECK-BASELINE-NEXT: movb %r9b, 6(%rcx)
; CHECK-BASELINE-NEXT: movb %r8b, 5(%rcx)
; CHECK-BASELINE-NEXT: movb %dil, 4(%rcx)
; CHECK-BASELINE-NEXT: movb %sil, 3(%rcx)
; CHECK-BASELINE-NEXT: movb %dl, 2(%rcx)
; CHECK-BASELINE-NEXT: movb %al, 1(%rcx)
; CHECK-BASELINE-NEXT: movb %r15b, (%rcx)
; CHECK-BASELINE-NEXT: movq %rcx, %rax
; CHECK-BASELINE-NEXT: popq %rbx
; CHECK-BASELINE-NEXT: popq %r12
; CHECK-BASELINE-NEXT: popq %r13
; CHECK-BASELINE-NEXT: popq %r14
; CHECK-BASELINE-NEXT: popq %r15
; CHECK-BASELINE-NEXT: popq %rbp
; CHECK-BASELINE-NEXT: retq
;
; CHECK-SSE1-LABEL: out_v32i8:
; CHECK-SSE1: # %bb.0:
; CHECK-SSE1-NEXT: pushq %rbp
; CHECK-SSE1-NEXT: pushq %r15
; CHECK-SSE1-NEXT: pushq %r14
; CHECK-SSE1-NEXT: pushq %r13
; CHECK-SSE1-NEXT: pushq %r12
; CHECK-SSE1-NEXT: pushq %rbx
; CHECK-SSE1-NEXT: movq %rcx, %r15
; CHECK-SSE1-NEXT: movq %rsi, %r14
; CHECK-SSE1-NEXT: movq %rdi, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
; CHECK-SSE1-NEXT: movb 16(%rcx), %al
; CHECK-SSE1-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 17(%rcx), %al
; CHECK-SSE1-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 18(%rcx), %al
; CHECK-SSE1-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 19(%rcx), %al
; CHECK-SSE1-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 20(%rcx), %al
; CHECK-SSE1-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 21(%rcx), %r12b
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-SSE1-NEXT: movb 22(%rcx), %r9b
; CHECK-SSE1-NEXT: movb 23(%rcx), %r10b
; CHECK-SSE1-NEXT: movb 24(%rcx), %r11b
; CHECK-SSE1-NEXT: movb 25(%rcx), %bpl
; CHECK-SSE1-NEXT: movb 26(%rcx), %r13b
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb 27(%rcx), %r8b
; CHECK-SSE1-NEXT: movb 28(%rcx), %dil
; CHECK-SSE1-NEXT: movb 29(%rcx), %sil
; CHECK-SSE1-NEXT: movb 30(%rcx), %bl
; CHECK-SSE1-NEXT: movb 31(%rcx), %al
; CHECK-SSE1-NEXT: movb 31(%r14), %cl
; CHECK-SSE1-NEXT: andb %al, %cl
; CHECK-SSE1-NEXT: notb %al
; CHECK-SSE1-NEXT: andb 31(%rdx), %al
; CHECK-SSE1-NEXT: orb %cl, %al
; CHECK-SSE1-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 30(%r14), %al
; CHECK-SSE1-NEXT: andb %bl, %al
; CHECK-SSE1-NEXT: notb %bl
; CHECK-SSE1-NEXT: andb 30(%rdx), %bl
; CHECK-SSE1-NEXT: orb %al, %bl
; CHECK-SSE1-NEXT: movb %bl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 29(%r14), %al
; CHECK-SSE1-NEXT: andb %sil, %al
; CHECK-SSE1-NEXT: notb %sil
; CHECK-SSE1-NEXT: andb 29(%rdx), %sil
; CHECK-SSE1-NEXT: orb %al, %sil
; CHECK-SSE1-NEXT: movb %sil, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 28(%r14), %al
; CHECK-SSE1-NEXT: andb %dil, %al
; CHECK-SSE1-NEXT: notb %dil
; CHECK-SSE1-NEXT: andb 28(%rdx), %dil
; CHECK-SSE1-NEXT: orb %al, %dil
; CHECK-SSE1-NEXT: movb %dil, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 27(%r14), %al
; CHECK-SSE1-NEXT: andb %r8b, %al
; CHECK-SSE1-NEXT: notb %r8b
; CHECK-SSE1-NEXT: andb 27(%rdx), %r8b
; CHECK-SSE1-NEXT: orb %al, %r8b
; CHECK-SSE1-NEXT: movb %r8b, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 26(%r14), %al
; CHECK-SSE1-NEXT: andb %r13b, %al
; CHECK-SSE1-NEXT: notb %r13b
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-SSE1-NEXT: andb 26(%rdx), %r13b
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: orb %al, %r13b
; CHECK-SSE1-NEXT: movb %r13b, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-SSE1-NEXT: movb 25(%r14), %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: andb %bpl, %al
; CHECK-SSE1-NEXT: notb %bpl
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-SSE1-NEXT: andb 25(%rdx), %bpl
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: orb %al, %bpl
; CHECK-SSE1-NEXT: movb %bpl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-SSE1-NEXT: movb 24(%r14), %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: andb %r11b, %al
; CHECK-SSE1-NEXT: notb %r11b
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-SSE1-NEXT: andb 24(%rdx), %r11b
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: orb %al, %r11b
; CHECK-SSE1-NEXT: movb %r11b, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-SSE1-NEXT: movb 23(%r14), %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: andb %r10b, %al
; CHECK-SSE1-NEXT: notb %r10b
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-SSE1-NEXT: andb 23(%rdx), %r10b
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: orb %al, %r10b
; CHECK-SSE1-NEXT: movb %r10b, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-SSE1-NEXT: movb 22(%r14), %al
; CHECK-SSE1-NEXT: andb %r9b, %al
; CHECK-SSE1-NEXT: notb %r9b
; CHECK-SSE1-NEXT: andb 22(%rdx), %r9b
; CHECK-SSE1-NEXT: orb %al, %r9b
; CHECK-SSE1-NEXT: movb %r9b, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb 21(%r14), %al
; CHECK-SSE1-NEXT: andb %r12b, %al
; CHECK-SSE1-NEXT: notb %r12b
; CHECK-SSE1-NEXT: andb 21(%rdx), %r12b
; CHECK-SSE1-NEXT: orb %al, %r12b
; CHECK-SSE1-NEXT: movb %r12b, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 20(%r14), %al
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %cl # 1-byte Reload
; CHECK-SSE1-NEXT: andb %cl, %al
; CHECK-SSE1-NEXT: notb %cl
; CHECK-SSE1-NEXT: andb 20(%rdx), %cl
; CHECK-SSE1-NEXT: orb %al, %cl
; CHECK-SSE1-NEXT: movb %cl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 19(%r14), %al
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %cl # 1-byte Reload
; CHECK-SSE1-NEXT: andb %cl, %al
; CHECK-SSE1-NEXT: notb %cl
; CHECK-SSE1-NEXT: andb 19(%rdx), %cl
; CHECK-SSE1-NEXT: orb %al, %cl
; CHECK-SSE1-NEXT: movb %cl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 18(%r14), %al
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %cl # 1-byte Reload
; CHECK-SSE1-NEXT: andb %cl, %al
; CHECK-SSE1-NEXT: notb %cl
; CHECK-SSE1-NEXT: andb 18(%rdx), %cl
; CHECK-SSE1-NEXT: orb %al, %cl
; CHECK-SSE1-NEXT: movb %cl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 17(%r14), %al
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %cl # 1-byte Reload
; CHECK-SSE1-NEXT: andb %cl, %al
; CHECK-SSE1-NEXT: notb %cl
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-SSE1-NEXT: movq %rdx, %rbx
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: andb 17(%rdx), %cl
; CHECK-SSE1-NEXT: orb %al, %cl
; CHECK-SSE1-NEXT: movb %cl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 16(%r14), %al
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %cl # 1-byte Reload
; CHECK-SSE1-NEXT: andb %cl, %al
; CHECK-SSE1-NEXT: notb %cl
; CHECK-SSE1-NEXT: andb 16(%rdx), %cl
; CHECK-SSE1-NEXT: orb %al, %cl
; CHECK-SSE1-NEXT: movb %cl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-SSE1-NEXT: movb 15(%r15), %cl
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb 15(%r14), %al
; CHECK-SSE1-NEXT: andb %cl, %al
; CHECK-SSE1-NEXT: notb %cl
; CHECK-SSE1-NEXT: andb 15(%rdx), %cl
; CHECK-SSE1-NEXT: orb %al, %cl
; CHECK-SSE1-NEXT: movb %cl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 14(%r15), %cl
; CHECK-SSE1-NEXT: movb 14(%r14), %al
; CHECK-SSE1-NEXT: andb %cl, %al
; CHECK-SSE1-NEXT: notb %cl
; CHECK-SSE1-NEXT: andb 14(%rdx), %cl
; CHECK-SSE1-NEXT: orb %al, %cl
; CHECK-SSE1-NEXT: movb %cl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 13(%r15), %cl
; CHECK-SSE1-NEXT: movb 13(%r14), %al
; CHECK-SSE1-NEXT: andb %cl, %al
; CHECK-SSE1-NEXT: notb %cl
; CHECK-SSE1-NEXT: andb 13(%rdx), %cl
; CHECK-SSE1-NEXT: orb %al, %cl
; CHECK-SSE1-NEXT: movb %cl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 12(%r15), %cl
; CHECK-SSE1-NEXT: movb 12(%r14), %al
; CHECK-SSE1-NEXT: andb %cl, %al
; CHECK-SSE1-NEXT: notb %cl
; CHECK-SSE1-NEXT: andb 12(%rdx), %cl
; CHECK-SSE1-NEXT: orb %al, %cl
; CHECK-SSE1-NEXT: movb %cl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 11(%r15), %r13b
; CHECK-SSE1-NEXT: movb 11(%r14), %al
; CHECK-SSE1-NEXT: andb %r13b, %al
; CHECK-SSE1-NEXT: notb %r13b
; CHECK-SSE1-NEXT: andb 11(%rdx), %r13b
; CHECK-SSE1-NEXT: orb %al, %r13b
; CHECK-SSE1-NEXT: movb 10(%r15), %r12b
; CHECK-SSE1-NEXT: movb 10(%r14), %al
; CHECK-SSE1-NEXT: andb %r12b, %al
; CHECK-SSE1-NEXT: notb %r12b
; CHECK-SSE1-NEXT: andb 10(%rdx), %r12b
; CHECK-SSE1-NEXT: orb %al, %r12b
; CHECK-SSE1-NEXT: movb 9(%r15), %bpl
; CHECK-SSE1-NEXT: movb 9(%r14), %al
; CHECK-SSE1-NEXT: andb %bpl, %al
; CHECK-SSE1-NEXT: notb %bpl
; CHECK-SSE1-NEXT: andb 9(%rdx), %bpl
; CHECK-SSE1-NEXT: orb %al, %bpl
; CHECK-SSE1-NEXT: movb 8(%r15), %r11b
; CHECK-SSE1-NEXT: movb 8(%r14), %al
; CHECK-SSE1-NEXT: andb %r11b, %al
; CHECK-SSE1-NEXT: notb %r11b
; CHECK-SSE1-NEXT: andb 8(%rdx), %r11b
; CHECK-SSE1-NEXT: orb %al, %r11b
; CHECK-SSE1-NEXT: movb 7(%r15), %r10b
; CHECK-SSE1-NEXT: movb 7(%r14), %al
; CHECK-SSE1-NEXT: andb %r10b, %al
; CHECK-SSE1-NEXT: notb %r10b
; CHECK-SSE1-NEXT: andb 7(%rdx), %r10b
; CHECK-SSE1-NEXT: orb %al, %r10b
; CHECK-SSE1-NEXT: movb 6(%r15), %r9b
; CHECK-SSE1-NEXT: movb 6(%r14), %al
; CHECK-SSE1-NEXT: andb %r9b, %al
; CHECK-SSE1-NEXT: notb %r9b
; CHECK-SSE1-NEXT: andb 6(%rdx), %r9b
; CHECK-SSE1-NEXT: orb %al, %r9b
; CHECK-SSE1-NEXT: movb 5(%r15), %r8b
; CHECK-SSE1-NEXT: movb 5(%r14), %al
; CHECK-SSE1-NEXT: andb %r8b, %al
; CHECK-SSE1-NEXT: notb %r8b
; CHECK-SSE1-NEXT: andb 5(%rdx), %r8b
; CHECK-SSE1-NEXT: orb %al, %r8b
; CHECK-SSE1-NEXT: movb 4(%r15), %dil
; CHECK-SSE1-NEXT: movb 4(%r14), %al
; CHECK-SSE1-NEXT: andb %dil, %al
; CHECK-SSE1-NEXT: notb %dil
; CHECK-SSE1-NEXT: andb 4(%rdx), %dil
; CHECK-SSE1-NEXT: orb %al, %dil
; CHECK-SSE1-NEXT: movb 3(%r15), %sil
; CHECK-SSE1-NEXT: movb 3(%r14), %al
; CHECK-SSE1-NEXT: andb %sil, %al
; CHECK-SSE1-NEXT: notb %sil
; CHECK-SSE1-NEXT: andb 3(%rdx), %sil
; CHECK-SSE1-NEXT: orb %al, %sil
; CHECK-SSE1-NEXT: movb 2(%r15), %dl
; CHECK-SSE1-NEXT: movb 2(%r14), %al
; CHECK-SSE1-NEXT: andb %dl, %al
; CHECK-SSE1-NEXT: notb %dl
; CHECK-SSE1-NEXT: andb 2(%rbx), %dl
; CHECK-SSE1-NEXT: orb %al, %dl
; CHECK-SSE1-NEXT: movb 1(%r15), %al
; CHECK-SSE1-NEXT: movb 1(%r14), %cl
; CHECK-SSE1-NEXT: andb %al, %cl
; CHECK-SSE1-NEXT: notb %al
; CHECK-SSE1-NEXT: andb 1(%rbx), %al
; CHECK-SSE1-NEXT: orb %cl, %al
; CHECK-SSE1-NEXT: movb (%r15), %r15b
; CHECK-SSE1-NEXT: movb (%r14), %r14b
; CHECK-SSE1-NEXT: andb %r15b, %r14b
; CHECK-SSE1-NEXT: notb %r15b
; CHECK-SSE1-NEXT: andb (%rbx), %r15b
; CHECK-SSE1-NEXT: orb %r14b, %r15b
; CHECK-SSE1-NEXT: movq {{[-0-9]+}}(%r{{[sb]}}p), %rcx # 8-byte Reload
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-SSE1-NEXT: movb %bl, 31(%rcx)
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-SSE1-NEXT: movb %bl, 30(%rcx)
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-SSE1-NEXT: movb %bl, 29(%rcx)
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-SSE1-NEXT: movb %bl, 28(%rcx)
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-SSE1-NEXT: movb %bl, 27(%rcx)
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-SSE1-NEXT: movb %bl, 26(%rcx)
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-SSE1-NEXT: movb %bl, 25(%rcx)
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-SSE1-NEXT: movb %bl, 24(%rcx)
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-SSE1-NEXT: movb %bl, 23(%rcx)
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-SSE1-NEXT: movb %bl, 22(%rcx)
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-SSE1-NEXT: movb %bl, 21(%rcx)
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-SSE1-NEXT: movb %bl, 20(%rcx)
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-SSE1-NEXT: movb %bl, 19(%rcx)
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-SSE1-NEXT: movb %bl, 18(%rcx)
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-SSE1-NEXT: movb %bl, 17(%rcx)
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-SSE1-NEXT: movb %bl, 16(%rcx)
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-SSE1-NEXT: movb %bl, 15(%rcx)
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-SSE1-NEXT: movb %bl, 14(%rcx)
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-SSE1-NEXT: movb %bl, 13(%rcx)
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %bl # 1-byte Reload
; CHECK-SSE1-NEXT: movb %bl, 12(%rcx)
; CHECK-SSE1-NEXT: movb %r13b, 11(%rcx)
; CHECK-SSE1-NEXT: movb %r12b, 10(%rcx)
; CHECK-SSE1-NEXT: movb %bpl, 9(%rcx)
; CHECK-SSE1-NEXT: movb %r11b, 8(%rcx)
; CHECK-SSE1-NEXT: movb %r10b, 7(%rcx)
; CHECK-SSE1-NEXT: movb %r9b, 6(%rcx)
; CHECK-SSE1-NEXT: movb %r8b, 5(%rcx)
; CHECK-SSE1-NEXT: movb %dil, 4(%rcx)
; CHECK-SSE1-NEXT: movb %sil, 3(%rcx)
; CHECK-SSE1-NEXT: movb %dl, 2(%rcx)
; CHECK-SSE1-NEXT: movb %al, 1(%rcx)
; CHECK-SSE1-NEXT: movb %r15b, (%rcx)
; CHECK-SSE1-NEXT: movq %rcx, %rax
; CHECK-SSE1-NEXT: popq %rbx
; CHECK-SSE1-NEXT: popq %r12
; CHECK-SSE1-NEXT: popq %r13
; CHECK-SSE1-NEXT: popq %r14
; CHECK-SSE1-NEXT: popq %r15
; CHECK-SSE1-NEXT: popq %rbp
; CHECK-SSE1-NEXT: retq
;
; CHECK-SSE2-LABEL: out_v32i8:
; CHECK-SSE2: # %bb.0:
; CHECK-SSE2-NEXT: movaps (%rdx), %xmm0
; CHECK-SSE2-NEXT: movaps 16(%rdx), %xmm1
; CHECK-SSE2-NEXT: movaps 16(%rdi), %xmm2
; CHECK-SSE2-NEXT: andps %xmm1, %xmm2
; CHECK-SSE2-NEXT: movaps (%rdi), %xmm3
; CHECK-SSE2-NEXT: andps %xmm0, %xmm3
; CHECK-SSE2-NEXT: andnps 16(%rsi), %xmm1
; CHECK-SSE2-NEXT: orps %xmm2, %xmm1
; CHECK-SSE2-NEXT: andnps (%rsi), %xmm0
; CHECK-SSE2-NEXT: orps %xmm3, %xmm0
; CHECK-SSE2-NEXT: retq
;
; CHECK-XOP-LABEL: out_v32i8:
; CHECK-XOP: # %bb.0:
; CHECK-XOP-NEXT: vmovdqa (%rdi), %ymm0
; CHECK-XOP-NEXT: vmovdqa (%rdx), %ymm1
; CHECK-XOP-NEXT: vpcmov %ymm1, (%rsi), %ymm0, %ymm0
; CHECK-XOP-NEXT: retq
%x = load <32 x i8>, <32 x i8> *%px, align 32
%y = load <32 x i8>, <32 x i8> *%py, align 32
%mask = load <32 x i8>, <32 x i8> *%pmask, align 32
%mx = and <32 x i8> %x, %mask
%notmask = xor <32 x i8> %mask, <i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1>
%my = and <32 x i8> %y, %notmask
%r = or <32 x i8> %mx, %my
ret <32 x i8> %r
}
define <16 x i16> @out_v16i16(<16 x i16> *%px, <16 x i16> *%py, <16 x i16> *%pmask) nounwind {
; CHECK-BASELINE-LABEL: out_v16i16:
; CHECK-BASELINE: # %bb.0:
; CHECK-BASELINE-NEXT: pushq %rbp
; CHECK-BASELINE-NEXT: pushq %r15
; CHECK-BASELINE-NEXT: pushq %r14
; CHECK-BASELINE-NEXT: pushq %r13
; CHECK-BASELINE-NEXT: pushq %r12
; CHECK-BASELINE-NEXT: pushq %rbx
; CHECK-BASELINE-NEXT: movq %rcx, %r9
; CHECK-BASELINE-NEXT: movq %rdx, %r10
; CHECK-BASELINE-NEXT: movq %rsi, %r8
; CHECK-BASELINE-NEXT: movq %rdi, %r11
; CHECK-BASELINE-NEXT: movl 12(%rcx), %eax
; CHECK-BASELINE-NEXT: movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-BASELINE-NEXT: movzwl 14(%rcx), %edx
; CHECK-BASELINE-NEXT: movl 16(%rcx), %esi
; CHECK-BASELINE-NEXT: movzwl 18(%rcx), %edi
; CHECK-BASELINE-NEXT: movl 20(%rcx), %ecx
; CHECK-BASELINE-NEXT: movzwl 22(%r9), %r15d
; CHECK-BASELINE-NEXT: movl 24(%r9), %r12d
; CHECK-BASELINE-NEXT: movzwl 26(%r9), %r14d
; CHECK-BASELINE-NEXT: movl 28(%r9), %ebx
; CHECK-BASELINE-NEXT: movzwl 30(%r9), %ebp
; CHECK-BASELINE-NEXT: movzwl 30(%r8), %r13d
; CHECK-BASELINE-NEXT: andw %bp, %r13w
; CHECK-BASELINE-NEXT: notl %ebp
; CHECK-BASELINE-NEXT: andw 30(%r10), %bp
; CHECK-BASELINE-NEXT: orl %r13d, %ebp
; CHECK-BASELINE-NEXT: movzwl 28(%r8), %eax
; CHECK-BASELINE-NEXT: andw %bx, %ax
; CHECK-BASELINE-NEXT: notl %ebx
; CHECK-BASELINE-NEXT: andw 28(%r10), %bx
; CHECK-BASELINE-NEXT: orl %eax, %ebx
; CHECK-BASELINE-NEXT: movzwl 26(%r8), %eax
; CHECK-BASELINE-NEXT: andw %r14w, %ax
; CHECK-BASELINE-NEXT: notl %r14d
; CHECK-BASELINE-NEXT: andw 26(%r10), %r14w
; CHECK-BASELINE-NEXT: orl %eax, %r14d
; CHECK-BASELINE-NEXT: movzwl 24(%r8), %eax
; CHECK-BASELINE-NEXT: andw %r12w, %ax
; CHECK-BASELINE-NEXT: notl %r12d
; CHECK-BASELINE-NEXT: andw 24(%r10), %r12w
; CHECK-BASELINE-NEXT: orl %eax, %r12d
; CHECK-BASELINE-NEXT: movzwl 22(%r8), %eax
; CHECK-BASELINE-NEXT: andw %r15w, %ax
; CHECK-BASELINE-NEXT: notl %r15d
; CHECK-BASELINE-NEXT: andw 22(%r10), %r15w
; CHECK-BASELINE-NEXT: orl %eax, %r15d
; CHECK-BASELINE-NEXT: movzwl 20(%r8), %eax
; CHECK-BASELINE-NEXT: andw %cx, %ax
; CHECK-BASELINE-NEXT: notl %ecx
; CHECK-BASELINE-NEXT: andw 20(%r10), %cx
; CHECK-BASELINE-NEXT: orl %eax, %ecx
; CHECK-BASELINE-NEXT: movl %ecx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-BASELINE-NEXT: movzwl 18(%r8), %eax
; CHECK-BASELINE-NEXT: andw %di, %ax
; CHECK-BASELINE-NEXT: notl %edi
; CHECK-BASELINE-NEXT: andw 18(%r10), %di
; CHECK-BASELINE-NEXT: orl %eax, %edi
; CHECK-BASELINE-NEXT: movl %edi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-BASELINE-NEXT: movzwl 16(%r8), %eax
; CHECK-BASELINE-NEXT: andw %si, %ax
; CHECK-BASELINE-NEXT: notl %esi
; CHECK-BASELINE-NEXT: andw 16(%r10), %si
; CHECK-BASELINE-NEXT: orl %eax, %esi
; CHECK-BASELINE-NEXT: movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-BASELINE-NEXT: movzwl 14(%r8), %eax
; CHECK-BASELINE-NEXT: andw %dx, %ax
; CHECK-BASELINE-NEXT: notl %edx
; CHECK-BASELINE-NEXT: andw 14(%r10), %dx
; CHECK-BASELINE-NEXT: orl %eax, %edx
; CHECK-BASELINE-NEXT: movl %edx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-BASELINE-NEXT: movzwl 12(%r8), %eax
; CHECK-BASELINE-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %ecx # 4-byte Reload
; CHECK-BASELINE-NEXT: andw %cx, %ax
; CHECK-BASELINE-NEXT: notl %ecx
; CHECK-BASELINE-NEXT: andw 12(%r10), %cx
; CHECK-BASELINE-NEXT: orl %eax, %ecx
; CHECK-BASELINE-NEXT: movl %ecx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-BASELINE-NEXT: movzwl 10(%r9), %r13d
; CHECK-BASELINE-NEXT: movzwl 10(%r8), %eax
; CHECK-BASELINE-NEXT: andw %r13w, %ax
; CHECK-BASELINE-NEXT: notl %r13d
; CHECK-BASELINE-NEXT: andw 10(%r10), %r13w
; CHECK-BASELINE-NEXT: orl %eax, %r13d
; CHECK-BASELINE-NEXT: movl 8(%r9), %edi
; CHECK-BASELINE-NEXT: movzwl 8(%r8), %eax
; CHECK-BASELINE-NEXT: andw %di, %ax
; CHECK-BASELINE-NEXT: notl %edi
; CHECK-BASELINE-NEXT: andw 8(%r10), %di
; CHECK-BASELINE-NEXT: orl %eax, %edi
; CHECK-BASELINE-NEXT: movzwl 6(%r9), %esi
; CHECK-BASELINE-NEXT: movzwl 6(%r8), %eax
; CHECK-BASELINE-NEXT: andw %si, %ax
; CHECK-BASELINE-NEXT: notl %esi
; CHECK-BASELINE-NEXT: andw 6(%r10), %si
; CHECK-BASELINE-NEXT: orl %eax, %esi
; CHECK-BASELINE-NEXT: movl 4(%r9), %edx
; CHECK-BASELINE-NEXT: movzwl 4(%r8), %eax
; CHECK-BASELINE-NEXT: andw %dx, %ax
; CHECK-BASELINE-NEXT: notl %edx
; CHECK-BASELINE-NEXT: andw 4(%r10), %dx
; CHECK-BASELINE-NEXT: orl %eax, %edx
; CHECK-BASELINE-NEXT: movzwl 2(%r9), %eax
; CHECK-BASELINE-NEXT: movzwl 2(%r8), %ecx
; CHECK-BASELINE-NEXT: andw %ax, %cx
; CHECK-BASELINE-NEXT: notl %eax
; CHECK-BASELINE-NEXT: andw 2(%r10), %ax
; CHECK-BASELINE-NEXT: orl %ecx, %eax
; CHECK-BASELINE-NEXT: movl (%r9), %r9d
; CHECK-BASELINE-NEXT: movzwl (%r8), %ecx
; CHECK-BASELINE-NEXT: andw %r9w, %cx
; CHECK-BASELINE-NEXT: notl %r9d
; CHECK-BASELINE-NEXT: andw (%r10), %r9w
; CHECK-BASELINE-NEXT: orl %ecx, %r9d
; CHECK-BASELINE-NEXT: movw %bp, 30(%r11)
; CHECK-BASELINE-NEXT: movw %bx, 28(%r11)
; CHECK-BASELINE-NEXT: movw %r14w, 26(%r11)
; CHECK-BASELINE-NEXT: movw %r12w, 24(%r11)
; CHECK-BASELINE-NEXT: movw %r15w, 22(%r11)
; CHECK-BASELINE-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %ecx # 4-byte Reload
; CHECK-BASELINE-NEXT: movw %cx, 20(%r11)
; CHECK-BASELINE-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %ecx # 4-byte Reload
; CHECK-BASELINE-NEXT: movw %cx, 18(%r11)
; CHECK-BASELINE-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %ecx # 4-byte Reload
; CHECK-BASELINE-NEXT: movw %cx, 16(%r11)
; CHECK-BASELINE-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %ecx # 4-byte Reload
; CHECK-BASELINE-NEXT: movw %cx, 14(%r11)
; CHECK-BASELINE-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %ecx # 4-byte Reload
; CHECK-BASELINE-NEXT: movw %cx, 12(%r11)
; CHECK-BASELINE-NEXT: movw %r13w, 10(%r11)
; CHECK-BASELINE-NEXT: movw %di, 8(%r11)
; CHECK-BASELINE-NEXT: movw %si, 6(%r11)
; CHECK-BASELINE-NEXT: movw %dx, 4(%r11)
; CHECK-BASELINE-NEXT: movw %ax, 2(%r11)
; CHECK-BASELINE-NEXT: movw %r9w, (%r11)
; CHECK-BASELINE-NEXT: movq %r11, %rax
; CHECK-BASELINE-NEXT: popq %rbx
; CHECK-BASELINE-NEXT: popq %r12
; CHECK-BASELINE-NEXT: popq %r13
; CHECK-BASELINE-NEXT: popq %r14
; CHECK-BASELINE-NEXT: popq %r15
; CHECK-BASELINE-NEXT: popq %rbp
; CHECK-BASELINE-NEXT: retq
;
; CHECK-SSE1-LABEL: out_v16i16:
; CHECK-SSE1: # %bb.0:
; CHECK-SSE1-NEXT: pushq %rbp
; CHECK-SSE1-NEXT: pushq %r15
; CHECK-SSE1-NEXT: pushq %r14
; CHECK-SSE1-NEXT: pushq %r13
; CHECK-SSE1-NEXT: pushq %r12
; CHECK-SSE1-NEXT: pushq %rbx
; CHECK-SSE1-NEXT: movq %rcx, %r9
; CHECK-SSE1-NEXT: movq %rdx, %r10
; CHECK-SSE1-NEXT: movq %rsi, %r8
; CHECK-SSE1-NEXT: movq %rdi, %r11
; CHECK-SSE1-NEXT: movl 12(%rcx), %eax
; CHECK-SSE1-NEXT: movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-SSE1-NEXT: movzwl 14(%rcx), %edx
; CHECK-SSE1-NEXT: movl 16(%rcx), %esi
; CHECK-SSE1-NEXT: movzwl 18(%rcx), %edi
; CHECK-SSE1-NEXT: movl 20(%rcx), %ecx
; CHECK-SSE1-NEXT: movzwl 22(%r9), %r15d
; CHECK-SSE1-NEXT: movl 24(%r9), %r12d
; CHECK-SSE1-NEXT: movzwl 26(%r9), %r14d
; CHECK-SSE1-NEXT: movl 28(%r9), %ebx
; CHECK-SSE1-NEXT: movzwl 30(%r9), %ebp
; CHECK-SSE1-NEXT: movzwl 30(%r8), %r13d
; CHECK-SSE1-NEXT: andw %bp, %r13w
; CHECK-SSE1-NEXT: notl %ebp
; CHECK-SSE1-NEXT: andw 30(%r10), %bp
; CHECK-SSE1-NEXT: orl %r13d, %ebp
; CHECK-SSE1-NEXT: movzwl 28(%r8), %eax
; CHECK-SSE1-NEXT: andw %bx, %ax
; CHECK-SSE1-NEXT: notl %ebx
; CHECK-SSE1-NEXT: andw 28(%r10), %bx
; CHECK-SSE1-NEXT: orl %eax, %ebx
; CHECK-SSE1-NEXT: movzwl 26(%r8), %eax
; CHECK-SSE1-NEXT: andw %r14w, %ax
; CHECK-SSE1-NEXT: notl %r14d
; CHECK-SSE1-NEXT: andw 26(%r10), %r14w
; CHECK-SSE1-NEXT: orl %eax, %r14d
; CHECK-SSE1-NEXT: movzwl 24(%r8), %eax
; CHECK-SSE1-NEXT: andw %r12w, %ax
; CHECK-SSE1-NEXT: notl %r12d
; CHECK-SSE1-NEXT: andw 24(%r10), %r12w
; CHECK-SSE1-NEXT: orl %eax, %r12d
; CHECK-SSE1-NEXT: movzwl 22(%r8), %eax
; CHECK-SSE1-NEXT: andw %r15w, %ax
; CHECK-SSE1-NEXT: notl %r15d
; CHECK-SSE1-NEXT: andw 22(%r10), %r15w
; CHECK-SSE1-NEXT: orl %eax, %r15d
; CHECK-SSE1-NEXT: movzwl 20(%r8), %eax
; CHECK-SSE1-NEXT: andw %cx, %ax
; CHECK-SSE1-NEXT: notl %ecx
; CHECK-SSE1-NEXT: andw 20(%r10), %cx
; CHECK-SSE1-NEXT: orl %eax, %ecx
; CHECK-SSE1-NEXT: movl %ecx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-SSE1-NEXT: movzwl 18(%r8), %eax
; CHECK-SSE1-NEXT: andw %di, %ax
; CHECK-SSE1-NEXT: notl %edi
; CHECK-SSE1-NEXT: andw 18(%r10), %di
; CHECK-SSE1-NEXT: orl %eax, %edi
; CHECK-SSE1-NEXT: movl %edi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-SSE1-NEXT: movzwl 16(%r8), %eax
; CHECK-SSE1-NEXT: andw %si, %ax
; CHECK-SSE1-NEXT: notl %esi
; CHECK-SSE1-NEXT: andw 16(%r10), %si
; CHECK-SSE1-NEXT: orl %eax, %esi
; CHECK-SSE1-NEXT: movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-SSE1-NEXT: movzwl 14(%r8), %eax
; CHECK-SSE1-NEXT: andw %dx, %ax
; CHECK-SSE1-NEXT: notl %edx
; CHECK-SSE1-NEXT: andw 14(%r10), %dx
; CHECK-SSE1-NEXT: orl %eax, %edx
; CHECK-SSE1-NEXT: movl %edx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-SSE1-NEXT: movzwl 12(%r8), %eax
; CHECK-SSE1-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %ecx # 4-byte Reload
; CHECK-SSE1-NEXT: andw %cx, %ax
; CHECK-SSE1-NEXT: notl %ecx
; CHECK-SSE1-NEXT: andw 12(%r10), %cx
; CHECK-SSE1-NEXT: orl %eax, %ecx
; CHECK-SSE1-NEXT: movl %ecx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-SSE1-NEXT: movzwl 10(%r9), %r13d
; CHECK-SSE1-NEXT: movzwl 10(%r8), %eax
; CHECK-SSE1-NEXT: andw %r13w, %ax
; CHECK-SSE1-NEXT: notl %r13d
; CHECK-SSE1-NEXT: andw 10(%r10), %r13w
; CHECK-SSE1-NEXT: orl %eax, %r13d
; CHECK-SSE1-NEXT: movl 8(%r9), %edi
; CHECK-SSE1-NEXT: movzwl 8(%r8), %eax
; CHECK-SSE1-NEXT: andw %di, %ax
; CHECK-SSE1-NEXT: notl %edi
; CHECK-SSE1-NEXT: andw 8(%r10), %di
; CHECK-SSE1-NEXT: orl %eax, %edi
; CHECK-SSE1-NEXT: movzwl 6(%r9), %esi
; CHECK-SSE1-NEXT: movzwl 6(%r8), %eax
; CHECK-SSE1-NEXT: andw %si, %ax
; CHECK-SSE1-NEXT: notl %esi
; CHECK-SSE1-NEXT: andw 6(%r10), %si
; CHECK-SSE1-NEXT: orl %eax, %esi
; CHECK-SSE1-NEXT: movl 4(%r9), %edx
; CHECK-SSE1-NEXT: movzwl 4(%r8), %eax
; CHECK-SSE1-NEXT: andw %dx, %ax
; CHECK-SSE1-NEXT: notl %edx
; CHECK-SSE1-NEXT: andw 4(%r10), %dx
; CHECK-SSE1-NEXT: orl %eax, %edx
; CHECK-SSE1-NEXT: movzwl 2(%r9), %eax
; CHECK-SSE1-NEXT: movzwl 2(%r8), %ecx
; CHECK-SSE1-NEXT: andw %ax, %cx
; CHECK-SSE1-NEXT: notl %eax
; CHECK-SSE1-NEXT: andw 2(%r10), %ax
; CHECK-SSE1-NEXT: orl %ecx, %eax
; CHECK-SSE1-NEXT: movl (%r9), %r9d
; CHECK-SSE1-NEXT: movzwl (%r8), %ecx
; CHECK-SSE1-NEXT: andw %r9w, %cx
; CHECK-SSE1-NEXT: notl %r9d
; CHECK-SSE1-NEXT: andw (%r10), %r9w
; CHECK-SSE1-NEXT: orl %ecx, %r9d
; CHECK-SSE1-NEXT: movw %bp, 30(%r11)
; CHECK-SSE1-NEXT: movw %bx, 28(%r11)
; CHECK-SSE1-NEXT: movw %r14w, 26(%r11)
; CHECK-SSE1-NEXT: movw %r12w, 24(%r11)
; CHECK-SSE1-NEXT: movw %r15w, 22(%r11)
; CHECK-SSE1-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %ecx # 4-byte Reload
; CHECK-SSE1-NEXT: movw %cx, 20(%r11)
; CHECK-SSE1-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %ecx # 4-byte Reload
; CHECK-SSE1-NEXT: movw %cx, 18(%r11)
; CHECK-SSE1-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %ecx # 4-byte Reload
; CHECK-SSE1-NEXT: movw %cx, 16(%r11)
; CHECK-SSE1-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %ecx # 4-byte Reload
; CHECK-SSE1-NEXT: movw %cx, 14(%r11)
; CHECK-SSE1-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %ecx # 4-byte Reload
; CHECK-SSE1-NEXT: movw %cx, 12(%r11)
; CHECK-SSE1-NEXT: movw %r13w, 10(%r11)
; CHECK-SSE1-NEXT: movw %di, 8(%r11)
; CHECK-SSE1-NEXT: movw %si, 6(%r11)
; CHECK-SSE1-NEXT: movw %dx, 4(%r11)
; CHECK-SSE1-NEXT: movw %ax, 2(%r11)
; CHECK-SSE1-NEXT: movw %r9w, (%r11)
; CHECK-SSE1-NEXT: movq %r11, %rax
; CHECK-SSE1-NEXT: popq %rbx
; CHECK-SSE1-NEXT: popq %r12
; CHECK-SSE1-NEXT: popq %r13
; CHECK-SSE1-NEXT: popq %r14
; CHECK-SSE1-NEXT: popq %r15
; CHECK-SSE1-NEXT: popq %rbp
; CHECK-SSE1-NEXT: retq
;
; CHECK-SSE2-LABEL: out_v16i16:
; CHECK-SSE2: # %bb.0:
; CHECK-SSE2-NEXT: movaps (%rdx), %xmm0
; CHECK-SSE2-NEXT: movaps 16(%rdx), %xmm1
; CHECK-SSE2-NEXT: movaps 16(%rdi), %xmm2
; CHECK-SSE2-NEXT: andps %xmm1, %xmm2
; CHECK-SSE2-NEXT: movaps (%rdi), %xmm3
; CHECK-SSE2-NEXT: andps %xmm0, %xmm3
; CHECK-SSE2-NEXT: andnps 16(%rsi), %xmm1
; CHECK-SSE2-NEXT: orps %xmm2, %xmm1
; CHECK-SSE2-NEXT: andnps (%rsi), %xmm0
; CHECK-SSE2-NEXT: orps %xmm3, %xmm0
; CHECK-SSE2-NEXT: retq
;
; CHECK-XOP-LABEL: out_v16i16:
; CHECK-XOP: # %bb.0:
; CHECK-XOP-NEXT: vmovdqa (%rdi), %ymm0
; CHECK-XOP-NEXT: vmovdqa (%rdx), %ymm1
; CHECK-XOP-NEXT: vpcmov %ymm1, (%rsi), %ymm0, %ymm0
; CHECK-XOP-NEXT: retq
%x = load <16 x i16>, <16 x i16> *%px, align 32
%y = load <16 x i16>, <16 x i16> *%py, align 32
%mask = load <16 x i16>, <16 x i16> *%pmask, align 32
%mx = and <16 x i16> %x, %mask
%notmask = xor <16 x i16> %mask, <i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1>
%my = and <16 x i16> %y, %notmask
%r = or <16 x i16> %mx, %my
ret <16 x i16> %r
}
define <8 x i32> @out_v8i32(<8 x i32> *%px, <8 x i32> *%py, <8 x i32> *%pmask) nounwind {
; CHECK-BASELINE-LABEL: out_v8i32:
; CHECK-BASELINE: # %bb.0:
; CHECK-BASELINE-NEXT: pushq %rbp
; CHECK-BASELINE-NEXT: pushq %r15
; CHECK-BASELINE-NEXT: pushq %r14
; CHECK-BASELINE-NEXT: pushq %rbx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movq %rdi, %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl 4(%rcx), %r8d
; CHECK-BASELINE-NEXT: movl 8(%rcx), %r9d
; CHECK-BASELINE-NEXT: movl 12(%rcx), %r10d
; CHECK-BASELINE-NEXT: movl 16(%rcx), %r11d
; CHECK-BASELINE-NEXT: movl 20(%rcx), %r15d
; CHECK-BASELINE-NEXT: movl 24(%rcx), %ebx
; CHECK-BASELINE-NEXT: movl 28(%rcx), %ebp
; CHECK-BASELINE-NEXT: movl 28(%rsi), %r14d
; CHECK-BASELINE-NEXT: andl %ebp, %r14d
; CHECK-BASELINE-NEXT: notl %ebp
; CHECK-BASELINE-NEXT: andl 28(%rdx), %ebp
; CHECK-BASELINE-NEXT: orl %r14d, %ebp
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movl 24(%rsi), %edi
; CHECK-BASELINE-NEXT: andl %ebx, %edi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: notl %ebx
; CHECK-BASELINE-NEXT: andl 24(%rdx), %ebx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: orl %edi, %ebx
; CHECK-BASELINE-NEXT: movl 20(%rsi), %edi
; CHECK-BASELINE-NEXT: andl %r15d, %edi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: notl %r15d
; CHECK-BASELINE-NEXT: andl 20(%rdx), %r15d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: orl %edi, %r15d
; CHECK-BASELINE-NEXT: movl 16(%rsi), %edi
; CHECK-BASELINE-NEXT: andl %r11d, %edi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: notl %r11d
; CHECK-BASELINE-NEXT: andl 16(%rdx), %r11d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: orl %edi, %r11d
; CHECK-BASELINE-NEXT: movl 12(%rsi), %edi
; CHECK-BASELINE-NEXT: andl %r10d, %edi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: notl %r10d
; CHECK-BASELINE-NEXT: andl 12(%rdx), %r10d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: orl %edi, %r10d
; CHECK-BASELINE-NEXT: movl 8(%rsi), %edi
; CHECK-BASELINE-NEXT: andl %r9d, %edi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: notl %r9d
; CHECK-BASELINE-NEXT: andl 8(%rdx), %r9d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: orl %edi, %r9d
; CHECK-BASELINE-NEXT: movl 4(%rsi), %edi
; CHECK-BASELINE-NEXT: andl %r8d, %edi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: notl %r8d
; CHECK-BASELINE-NEXT: andl 4(%rdx), %r8d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: orl %edi, %r8d
; CHECK-BASELINE-NEXT: movl (%rcx), %ecx
; CHECK-BASELINE-NEXT: movl (%rsi), %esi
; CHECK-BASELINE-NEXT: andl %ecx, %esi
; CHECK-BASELINE-NEXT: notl %ecx
; CHECK-BASELINE-NEXT: andl (%rdx), %ecx
; CHECK-BASELINE-NEXT: orl %esi, %ecx
; CHECK-BASELINE-NEXT: movl %ebp, 28(%rax)
; CHECK-BASELINE-NEXT: movl %ebx, 24(%rax)
; CHECK-BASELINE-NEXT: movl %r15d, 20(%rax)
; CHECK-BASELINE-NEXT: movl %r11d, 16(%rax)
; CHECK-BASELINE-NEXT: movl %r10d, 12(%rax)
; CHECK-BASELINE-NEXT: movl %r9d, 8(%rax)
; CHECK-BASELINE-NEXT: movl %r8d, 4(%rax)
; CHECK-BASELINE-NEXT: movl %ecx, (%rax)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: popq %rbx
; CHECK-BASELINE-NEXT: popq %r14
; CHECK-BASELINE-NEXT: popq %r15
; CHECK-BASELINE-NEXT: popq %rbp
; CHECK-BASELINE-NEXT: retq
;
; CHECK-SSE1-LABEL: out_v8i32:
; CHECK-SSE1: # %bb.0:
; CHECK-SSE1-NEXT: pushq %rbp
; CHECK-SSE1-NEXT: pushq %r15
; CHECK-SSE1-NEXT: pushq %r14
; CHECK-SSE1-NEXT: pushq %rbx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movq %rdi, %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movl 4(%rcx), %r8d
; CHECK-SSE1-NEXT: movl 8(%rcx), %r9d
; CHECK-SSE1-NEXT: movl 12(%rcx), %r10d
; CHECK-SSE1-NEXT: movl 16(%rcx), %r11d
; CHECK-SSE1-NEXT: movl 20(%rcx), %r15d
; CHECK-SSE1-NEXT: movl 24(%rcx), %ebx
; CHECK-SSE1-NEXT: movl 28(%rcx), %ebp
; CHECK-SSE1-NEXT: movl 28(%rsi), %r14d
; CHECK-SSE1-NEXT: andl %ebp, %r14d
; CHECK-SSE1-NEXT: notl %ebp
; CHECK-SSE1-NEXT: andl 28(%rdx), %ebp
; CHECK-SSE1-NEXT: orl %r14d, %ebp
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movl 24(%rsi), %edi
; CHECK-SSE1-NEXT: andl %ebx, %edi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: notl %ebx
; CHECK-SSE1-NEXT: andl 24(%rdx), %ebx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: orl %edi, %ebx
; CHECK-SSE1-NEXT: movl 20(%rsi), %edi
; CHECK-SSE1-NEXT: andl %r15d, %edi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: notl %r15d
; CHECK-SSE1-NEXT: andl 20(%rdx), %r15d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: orl %edi, %r15d
; CHECK-SSE1-NEXT: movl 16(%rsi), %edi
; CHECK-SSE1-NEXT: andl %r11d, %edi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: notl %r11d
; CHECK-SSE1-NEXT: andl 16(%rdx), %r11d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: orl %edi, %r11d
; CHECK-SSE1-NEXT: movl 12(%rsi), %edi
; CHECK-SSE1-NEXT: andl %r10d, %edi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: notl %r10d
; CHECK-SSE1-NEXT: andl 12(%rdx), %r10d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: orl %edi, %r10d
; CHECK-SSE1-NEXT: movl 8(%rsi), %edi
; CHECK-SSE1-NEXT: andl %r9d, %edi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: notl %r9d
; CHECK-SSE1-NEXT: andl 8(%rdx), %r9d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: orl %edi, %r9d
; CHECK-SSE1-NEXT: movl 4(%rsi), %edi
; CHECK-SSE1-NEXT: andl %r8d, %edi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: notl %r8d
; CHECK-SSE1-NEXT: andl 4(%rdx), %r8d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: orl %edi, %r8d
; CHECK-SSE1-NEXT: movl (%rcx), %ecx
; CHECK-SSE1-NEXT: movl (%rsi), %esi
; CHECK-SSE1-NEXT: andl %ecx, %esi
; CHECK-SSE1-NEXT: notl %ecx
; CHECK-SSE1-NEXT: andl (%rdx), %ecx
; CHECK-SSE1-NEXT: orl %esi, %ecx
; CHECK-SSE1-NEXT: movl %ebp, 28(%rax)
; CHECK-SSE1-NEXT: movl %ebx, 24(%rax)
; CHECK-SSE1-NEXT: movl %r15d, 20(%rax)
; CHECK-SSE1-NEXT: movl %r11d, 16(%rax)
; CHECK-SSE1-NEXT: movl %r10d, 12(%rax)
; CHECK-SSE1-NEXT: movl %r9d, 8(%rax)
; CHECK-SSE1-NEXT: movl %r8d, 4(%rax)
; CHECK-SSE1-NEXT: movl %ecx, (%rax)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: popq %rbx
; CHECK-SSE1-NEXT: popq %r14
; CHECK-SSE1-NEXT: popq %r15
; CHECK-SSE1-NEXT: popq %rbp
; CHECK-SSE1-NEXT: retq
;
; CHECK-SSE2-LABEL: out_v8i32:
; CHECK-SSE2: # %bb.0:
; CHECK-SSE2-NEXT: movaps (%rdx), %xmm0
; CHECK-SSE2-NEXT: movaps 16(%rdx), %xmm1
; CHECK-SSE2-NEXT: movaps 16(%rdi), %xmm2
; CHECK-SSE2-NEXT: andps %xmm1, %xmm2
; CHECK-SSE2-NEXT: movaps (%rdi), %xmm3
; CHECK-SSE2-NEXT: andps %xmm0, %xmm3
; CHECK-SSE2-NEXT: andnps 16(%rsi), %xmm1
; CHECK-SSE2-NEXT: orps %xmm2, %xmm1
; CHECK-SSE2-NEXT: andnps (%rsi), %xmm0
; CHECK-SSE2-NEXT: orps %xmm3, %xmm0
; CHECK-SSE2-NEXT: retq
;
; CHECK-XOP-LABEL: out_v8i32:
; CHECK-XOP: # %bb.0:
; CHECK-XOP-NEXT: vmovdqa (%rdi), %ymm0
; CHECK-XOP-NEXT: vmovdqa (%rdx), %ymm1
; CHECK-XOP-NEXT: vpcmov %ymm1, (%rsi), %ymm0, %ymm0
; CHECK-XOP-NEXT: retq
%x = load <8 x i32>, <8 x i32> *%px, align 32
%y = load <8 x i32>, <8 x i32> *%py, align 32
%mask = load <8 x i32>, <8 x i32> *%pmask, align 32
%mx = and <8 x i32> %x, %mask
%notmask = xor <8 x i32> %mask, <i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1>
%my = and <8 x i32> %y, %notmask
%r = or <8 x i32> %mx, %my
ret <8 x i32> %r
}
define <4 x i64> @out_v4i64(<4 x i64> *%px, <4 x i64> *%py, <4 x i64> *%pmask) nounwind {
; CHECK-BASELINE-LABEL: out_v4i64:
; CHECK-BASELINE: # %bb.0:
; CHECK-BASELINE-NEXT: pushq %rbx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movq %rdi, %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movq (%rcx), %r8
; CHECK-BASELINE-NEXT: movq 8(%rcx), %r9
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movq 16(%rcx), %rdi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movq 24(%rcx), %rcx
; CHECK-BASELINE-NEXT: movq 24(%rsi), %r10
; CHECK-BASELINE-NEXT: andq %rcx, %r10
; CHECK-BASELINE-NEXT: movq 16(%rsi), %r11
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: andq %rdi, %r11
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movq 8(%rsi), %rbx
; CHECK-BASELINE-NEXT: andq %r9, %rbx
; CHECK-BASELINE-NEXT: movq (%rsi), %rsi
; CHECK-BASELINE-NEXT: andq %r8, %rsi
; CHECK-BASELINE-NEXT: notq %r8
; CHECK-BASELINE-NEXT: notq %r9
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: notq %rdi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: notq %rcx
; CHECK-BASELINE-NEXT: andq 24(%rdx), %rcx
; CHECK-BASELINE-NEXT: orq %r10, %rcx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: andq 16(%rdx), %rdi
; CHECK-BASELINE-NEXT: orq %r11, %rdi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andq 8(%rdx), %r9
; CHECK-BASELINE-NEXT: orq %rbx, %r9
; CHECK-BASELINE-NEXT: andq (%rdx), %r8
; CHECK-BASELINE-NEXT: orq %rsi, %r8
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movq %rcx, 24(%rax)
; CHECK-BASELINE-NEXT: movq %rdi, 16(%rax)
; CHECK-BASELINE-NEXT: movq %r9, 8(%rax)
; CHECK-BASELINE-NEXT: movq %r8, (%rax)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: popq %rbx
; CHECK-BASELINE-NEXT: retq
;
; CHECK-SSE1-LABEL: out_v4i64:
; CHECK-SSE1: # %bb.0:
; CHECK-SSE1-NEXT: pushq %rbx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movq %rdi, %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movq (%rcx), %r8
; CHECK-SSE1-NEXT: movq 8(%rcx), %r9
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movq 16(%rcx), %rdi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movq 24(%rcx), %rcx
; CHECK-SSE1-NEXT: movq 24(%rsi), %r10
; CHECK-SSE1-NEXT: andq %rcx, %r10
; CHECK-SSE1-NEXT: movq 16(%rsi), %r11
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: andq %rdi, %r11
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movq 8(%rsi), %rbx
; CHECK-SSE1-NEXT: andq %r9, %rbx
; CHECK-SSE1-NEXT: movq (%rsi), %rsi
; CHECK-SSE1-NEXT: andq %r8, %rsi
; CHECK-SSE1-NEXT: notq %r8
; CHECK-SSE1-NEXT: notq %r9
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: notq %rdi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: notq %rcx
; CHECK-SSE1-NEXT: andq 24(%rdx), %rcx
; CHECK-SSE1-NEXT: orq %r10, %rcx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: andq 16(%rdx), %rdi
; CHECK-SSE1-NEXT: orq %r11, %rdi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: andq 8(%rdx), %r9
; CHECK-SSE1-NEXT: orq %rbx, %r9
; CHECK-SSE1-NEXT: andq (%rdx), %r8
; CHECK-SSE1-NEXT: orq %rsi, %r8
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movq %rcx, 24(%rax)
; CHECK-SSE1-NEXT: movq %rdi, 16(%rax)
; CHECK-SSE1-NEXT: movq %r9, 8(%rax)
; CHECK-SSE1-NEXT: movq %r8, (%rax)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: popq %rbx
; CHECK-SSE1-NEXT: retq
;
; CHECK-SSE2-LABEL: out_v4i64:
; CHECK-SSE2: # %bb.0:
; CHECK-SSE2-NEXT: movaps (%rdx), %xmm0
; CHECK-SSE2-NEXT: movaps 16(%rdx), %xmm1
; CHECK-SSE2-NEXT: movaps 16(%rdi), %xmm2
; CHECK-SSE2-NEXT: andps %xmm1, %xmm2
; CHECK-SSE2-NEXT: movaps (%rdi), %xmm3
; CHECK-SSE2-NEXT: andps %xmm0, %xmm3
; CHECK-SSE2-NEXT: andnps 16(%rsi), %xmm1
; CHECK-SSE2-NEXT: orps %xmm2, %xmm1
; CHECK-SSE2-NEXT: andnps (%rsi), %xmm0
; CHECK-SSE2-NEXT: orps %xmm3, %xmm0
; CHECK-SSE2-NEXT: retq
;
; CHECK-XOP-LABEL: out_v4i64:
; CHECK-XOP: # %bb.0:
; CHECK-XOP-NEXT: vmovdqa (%rdi), %ymm0
; CHECK-XOP-NEXT: vmovdqa (%rdx), %ymm1
; CHECK-XOP-NEXT: vpcmov %ymm1, (%rsi), %ymm0, %ymm0
; CHECK-XOP-NEXT: retq
%x = load <4 x i64>, <4 x i64> *%px, align 32
%y = load <4 x i64>, <4 x i64> *%py, align 32
%mask = load <4 x i64>, <4 x i64> *%pmask, align 32
%mx = and <4 x i64> %x, %mask
%notmask = xor <4 x i64> %mask, <i64 -1, i64 -1, i64 -1, i64 -1>
%my = and <4 x i64> %y, %notmask
%r = or <4 x i64> %mx, %my
ret <4 x i64> %r
}
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
; Should be the same as the previous one.
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
; ============================================================================ ;
; 8-bit vector width
; ============================================================================ ;
define <1 x i8> @in_v1i8(<1 x i8> %x, <1 x i8> %y, <1 x i8> %mask) nounwind {
; CHECK-LABEL: in_v1i8:
; CHECK: # %bb.0:
; CHECK-NEXT: movl %edi, %eax
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-NEXT: xorl %esi, %eax
; CHECK-NEXT: andl %edx, %eax
; CHECK-NEXT: xorl %esi, %eax
; CHECK-NEXT: # kill: def $al killed $al killed $eax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-NEXT: retq
%n0 = xor <1 x i8> %x, %y
%n1 = and <1 x i8> %n0, %mask
%r = xor <1 x i8> %n1, %y
ret <1 x i8> %r
}
; ============================================================================ ;
; 16-bit vector width
; ============================================================================ ;
define <2 x i8> @in_v2i8(<2 x i8> %x, <2 x i8> %y, <2 x i8> %mask) nounwind {
; CHECK-BASELINE-LABEL: in_v2i8:
; CHECK-BASELINE: # %bb.0:
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movl %edi, %eax
; CHECK-BASELINE-NEXT: xorl %edx, %eax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: xorl %ecx, %esi
[DAGCombiner] use root SDLoc for all nodes created by logic fold If this is not a valid way to assign an SDLoc, then we get this wrong all over SDAG. I don't know enough about the SDAG to explain this. IIUC, theoretically, debug info is not supposed to affect codegen. But here it has clearly affected 3 different targets, and the x86 change is an actual improvement. llvm-svn: 348552
2018-12-07 08:01:57 +08:00
; CHECK-BASELINE-NEXT: andl %r9d, %esi
; CHECK-BASELINE-NEXT: andl %r8d, %eax
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: xorl %edx, %eax
[DAGCombiner] use root SDLoc for all nodes created by logic fold If this is not a valid way to assign an SDLoc, then we get this wrong all over SDAG. I don't know enough about the SDAG to explain this. IIUC, theoretically, debug info is not supposed to affect codegen. But here it has clearly affected 3 different targets, and the x86 change is an actual improvement. llvm-svn: 348552
2018-12-07 08:01:57 +08:00
; CHECK-BASELINE-NEXT: xorl %ecx, %esi
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: # kill: def $al killed $al killed $eax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl %esi, %edx
; CHECK-BASELINE-NEXT: retq
;
; CHECK-SSE1-LABEL: in_v2i8:
; CHECK-SSE1: # %bb.0:
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movl %edi, %eax
; CHECK-SSE1-NEXT: xorl %edx, %eax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: xorl %ecx, %esi
[DAGCombiner] use root SDLoc for all nodes created by logic fold If this is not a valid way to assign an SDLoc, then we get this wrong all over SDAG. I don't know enough about the SDAG to explain this. IIUC, theoretically, debug info is not supposed to affect codegen. But here it has clearly affected 3 different targets, and the x86 change is an actual improvement. llvm-svn: 348552
2018-12-07 08:01:57 +08:00
; CHECK-SSE1-NEXT: andl %r9d, %esi
; CHECK-SSE1-NEXT: andl %r8d, %eax
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: xorl %edx, %eax
[DAGCombiner] use root SDLoc for all nodes created by logic fold If this is not a valid way to assign an SDLoc, then we get this wrong all over SDAG. I don't know enough about the SDAG to explain this. IIUC, theoretically, debug info is not supposed to affect codegen. But here it has clearly affected 3 different targets, and the x86 change is an actual improvement. llvm-svn: 348552
2018-12-07 08:01:57 +08:00
; CHECK-SSE1-NEXT: xorl %ecx, %esi
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: # kill: def $al killed $al killed $eax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movl %esi, %edx
; CHECK-SSE1-NEXT: retq
;
; CHECK-SSE2-LABEL: in_v2i8:
; CHECK-SSE2: # %bb.0:
; CHECK-SSE2-NEXT: andps %xmm2, %xmm0
[DAGCombine][X86][AArch64] Masked merge unfolding: vector edition. Summary: This **appears** to be the last missing piece for the masked merge pattern handling in the backend. This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`), and we need to make sure that they are generated. Differential Revision: https://reviews.llvm.org/D46528 llvm-svn: 332904
2018-05-22 05:41:02 +08:00
; CHECK-SSE2-NEXT: andnps %xmm1, %xmm2
; CHECK-SSE2-NEXT: orps %xmm2, %xmm0
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE2-NEXT: retq
;
; CHECK-XOP-LABEL: in_v2i8:
; CHECK-XOP: # %bb.0:
[DAGCombine][X86][AArch64] Masked merge unfolding: vector edition. Summary: This **appears** to be the last missing piece for the masked merge pattern handling in the backend. This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`), and we need to make sure that they are generated. Differential Revision: https://reviews.llvm.org/D46528 llvm-svn: 332904
2018-05-22 05:41:02 +08:00
; CHECK-XOP-NEXT: vpcmov %xmm2, %xmm1, %xmm0, %xmm0
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-XOP-NEXT: retq
%n0 = xor <2 x i8> %x, %y
%n1 = and <2 x i8> %n0, %mask
%r = xor <2 x i8> %n1, %y
ret <2 x i8> %r
}
define <1 x i16> @in_v1i16(<1 x i16> %x, <1 x i16> %y, <1 x i16> %mask) nounwind {
; CHECK-LABEL: in_v1i16:
; CHECK: # %bb.0:
; CHECK-NEXT: movl %edi, %eax
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-NEXT: xorl %esi, %eax
; CHECK-NEXT: andl %edx, %eax
; CHECK-NEXT: xorl %esi, %eax
; CHECK-NEXT: # kill: def $ax killed $ax killed $eax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-NEXT: retq
%n0 = xor <1 x i16> %x, %y
%n1 = and <1 x i16> %n0, %mask
%r = xor <1 x i16> %n1, %y
ret <1 x i16> %r
}
; ============================================================================ ;
; 32-bit vector width
; ============================================================================ ;
define <4 x i8> @in_v4i8(<4 x i8> %x, <4 x i8> %y, <4 x i8> %mask) nounwind {
; CHECK-BASELINE-LABEL: in_v4i8:
; CHECK-BASELINE: # %bb.0:
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movq %rdi, %rax
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %dil
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %r10b
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %r11b
[DAGCombiner] use root SDLoc for all nodes created by logic fold If this is not a valid way to assign an SDLoc, then we get this wrong all over SDAG. I don't know enough about the SDAG to explain this. IIUC, theoretically, debug info is not supposed to affect codegen. But here it has clearly affected 3 different targets, and the x86 change is an actual improvement. llvm-svn: 348552
2018-12-07 08:01:57 +08:00
; CHECK-BASELINE-NEXT: xorl %r9d, %esi
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: xorb %r11b, %dl
; CHECK-BASELINE-NEXT: xorb %r10b, %cl
; CHECK-BASELINE-NEXT: xorb %dil, %r8b
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %r8b
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %cl
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %dl
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %sil
; CHECK-BASELINE-NEXT: xorb %r9b, %sil
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: xorb %r11b, %dl
; CHECK-BASELINE-NEXT: xorb %r10b, %cl
; CHECK-BASELINE-NEXT: xorb %dil, %r8b
; CHECK-BASELINE-NEXT: movb %r8b, 3(%rax)
; CHECK-BASELINE-NEXT: movb %cl, 2(%rax)
; CHECK-BASELINE-NEXT: movb %dl, 1(%rax)
; CHECK-BASELINE-NEXT: movb %sil, (%rax)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: retq
;
; CHECK-SSE1-LABEL: in_v4i8:
; CHECK-SSE1: # %bb.0:
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movq %rdi, %rax
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %dil
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %r10b
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %r11b
[DAGCombiner] use root SDLoc for all nodes created by logic fold If this is not a valid way to assign an SDLoc, then we get this wrong all over SDAG. I don't know enough about the SDAG to explain this. IIUC, theoretically, debug info is not supposed to affect codegen. But here it has clearly affected 3 different targets, and the x86 change is an actual improvement. llvm-svn: 348552
2018-12-07 08:01:57 +08:00
; CHECK-SSE1-NEXT: xorl %r9d, %esi
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: xorb %r11b, %dl
; CHECK-SSE1-NEXT: xorb %r10b, %cl
; CHECK-SSE1-NEXT: xorb %dil, %r8b
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %r8b
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %cl
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %dl
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %sil
; CHECK-SSE1-NEXT: xorb %r9b, %sil
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: xorb %r11b, %dl
; CHECK-SSE1-NEXT: xorb %r10b, %cl
; CHECK-SSE1-NEXT: xorb %dil, %r8b
; CHECK-SSE1-NEXT: movb %r8b, 3(%rax)
; CHECK-SSE1-NEXT: movb %cl, 2(%rax)
; CHECK-SSE1-NEXT: movb %dl, 1(%rax)
; CHECK-SSE1-NEXT: movb %sil, (%rax)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: retq
;
; CHECK-SSE2-LABEL: in_v4i8:
; CHECK-SSE2: # %bb.0:
; CHECK-SSE2-NEXT: andps %xmm2, %xmm0
[DAGCombine][X86][AArch64] Masked merge unfolding: vector edition. Summary: This **appears** to be the last missing piece for the masked merge pattern handling in the backend. This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`), and we need to make sure that they are generated. Differential Revision: https://reviews.llvm.org/D46528 llvm-svn: 332904
2018-05-22 05:41:02 +08:00
; CHECK-SSE2-NEXT: andnps %xmm1, %xmm2
; CHECK-SSE2-NEXT: orps %xmm2, %xmm0
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE2-NEXT: retq
;
; CHECK-XOP-LABEL: in_v4i8:
; CHECK-XOP: # %bb.0:
[DAGCombine][X86][AArch64] Masked merge unfolding: vector edition. Summary: This **appears** to be the last missing piece for the masked merge pattern handling in the backend. This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`), and we need to make sure that they are generated. Differential Revision: https://reviews.llvm.org/D46528 llvm-svn: 332904
2018-05-22 05:41:02 +08:00
; CHECK-XOP-NEXT: vpcmov %xmm2, %xmm1, %xmm0, %xmm0
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-XOP-NEXT: retq
%n0 = xor <4 x i8> %x, %y
%n1 = and <4 x i8> %n0, %mask
%r = xor <4 x i8> %n1, %y
ret <4 x i8> %r
}
define <2 x i16> @in_v2i16(<2 x i16> %x, <2 x i16> %y, <2 x i16> %mask) nounwind {
; CHECK-BASELINE-LABEL: in_v2i16:
; CHECK-BASELINE: # %bb.0:
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movl %edi, %eax
; CHECK-BASELINE-NEXT: xorl %edx, %eax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: xorl %ecx, %esi
[DAGCombiner] use root SDLoc for all nodes created by logic fold If this is not a valid way to assign an SDLoc, then we get this wrong all over SDAG. I don't know enough about the SDAG to explain this. IIUC, theoretically, debug info is not supposed to affect codegen. But here it has clearly affected 3 different targets, and the x86 change is an actual improvement. llvm-svn: 348552
2018-12-07 08:01:57 +08:00
; CHECK-BASELINE-NEXT: andl %r9d, %esi
; CHECK-BASELINE-NEXT: andl %r8d, %eax
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: xorl %edx, %eax
[DAGCombiner] use root SDLoc for all nodes created by logic fold If this is not a valid way to assign an SDLoc, then we get this wrong all over SDAG. I don't know enough about the SDAG to explain this. IIUC, theoretically, debug info is not supposed to affect codegen. But here it has clearly affected 3 different targets, and the x86 change is an actual improvement. llvm-svn: 348552
2018-12-07 08:01:57 +08:00
; CHECK-BASELINE-NEXT: xorl %ecx, %esi
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: # kill: def $ax killed $ax killed $eax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl %esi, %edx
; CHECK-BASELINE-NEXT: retq
;
; CHECK-SSE1-LABEL: in_v2i16:
; CHECK-SSE1: # %bb.0:
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movl %edi, %eax
; CHECK-SSE1-NEXT: xorl %edx, %eax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: xorl %ecx, %esi
[DAGCombiner] use root SDLoc for all nodes created by logic fold If this is not a valid way to assign an SDLoc, then we get this wrong all over SDAG. I don't know enough about the SDAG to explain this. IIUC, theoretically, debug info is not supposed to affect codegen. But here it has clearly affected 3 different targets, and the x86 change is an actual improvement. llvm-svn: 348552
2018-12-07 08:01:57 +08:00
; CHECK-SSE1-NEXT: andl %r9d, %esi
; CHECK-SSE1-NEXT: andl %r8d, %eax
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: xorl %edx, %eax
[DAGCombiner] use root SDLoc for all nodes created by logic fold If this is not a valid way to assign an SDLoc, then we get this wrong all over SDAG. I don't know enough about the SDAG to explain this. IIUC, theoretically, debug info is not supposed to affect codegen. But here it has clearly affected 3 different targets, and the x86 change is an actual improvement. llvm-svn: 348552
2018-12-07 08:01:57 +08:00
; CHECK-SSE1-NEXT: xorl %ecx, %esi
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: # kill: def $ax killed $ax killed $eax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movl %esi, %edx
; CHECK-SSE1-NEXT: retq
;
; CHECK-SSE2-LABEL: in_v2i16:
; CHECK-SSE2: # %bb.0:
; CHECK-SSE2-NEXT: andps %xmm2, %xmm0
[DAGCombine][X86][AArch64] Masked merge unfolding: vector edition. Summary: This **appears** to be the last missing piece for the masked merge pattern handling in the backend. This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`), and we need to make sure that they are generated. Differential Revision: https://reviews.llvm.org/D46528 llvm-svn: 332904
2018-05-22 05:41:02 +08:00
; CHECK-SSE2-NEXT: andnps %xmm1, %xmm2
; CHECK-SSE2-NEXT: orps %xmm2, %xmm0
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE2-NEXT: retq
;
; CHECK-XOP-LABEL: in_v2i16:
; CHECK-XOP: # %bb.0:
[DAGCombine][X86][AArch64] Masked merge unfolding: vector edition. Summary: This **appears** to be the last missing piece for the masked merge pattern handling in the backend. This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`), and we need to make sure that they are generated. Differential Revision: https://reviews.llvm.org/D46528 llvm-svn: 332904
2018-05-22 05:41:02 +08:00
; CHECK-XOP-NEXT: vpcmov %xmm2, %xmm1, %xmm0, %xmm0
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-XOP-NEXT: retq
%n0 = xor <2 x i16> %x, %y
%n1 = and <2 x i16> %n0, %mask
%r = xor <2 x i16> %n1, %y
ret <2 x i16> %r
}
define <1 x i32> @in_v1i32(<1 x i32> %x, <1 x i32> %y, <1 x i32> %mask) nounwind {
; CHECK-LABEL: in_v1i32:
; CHECK: # %bb.0:
; CHECK-NEXT: movl %edi, %eax
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-NEXT: xorl %esi, %eax
; CHECK-NEXT: andl %edx, %eax
; CHECK-NEXT: xorl %esi, %eax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-NEXT: retq
%n0 = xor <1 x i32> %x, %y
%n1 = and <1 x i32> %n0, %mask
%r = xor <1 x i32> %n1, %y
ret <1 x i32> %r
}
; ============================================================================ ;
; 64-bit vector width
; ============================================================================ ;
define <8 x i8> @in_v8i8(<8 x i8> %x, <8 x i8> %y, <8 x i8> %mask) nounwind {
; CHECK-BASELINE-LABEL: in_v8i8:
; CHECK-BASELINE: # %bb.0:
; CHECK-BASELINE-NEXT: pushq %rbp
; CHECK-BASELINE-NEXT: pushq %r15
; CHECK-BASELINE-NEXT: pushq %r14
; CHECK-BASELINE-NEXT: pushq %r13
; CHECK-BASELINE-NEXT: pushq %r12
; CHECK-BASELINE-NEXT: pushq %rbx
; CHECK-BASELINE-NEXT: movl %ecx, %r10d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %r11b
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %bpl
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %r14b
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %r15b
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %r12b
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %r13b
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: xorb %r13b, %sil
; CHECK-BASELINE-NEXT: xorb %r12b, %dl
; CHECK-BASELINE-NEXT: xorb %r15b, %r10b
; CHECK-BASELINE-NEXT: xorb %r14b, %r8b
; CHECK-BASELINE-NEXT: xorb %bpl, %r9b
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %bl
; CHECK-BASELINE-NEXT: xorb {{[0-9]+}}(%rsp), %bl
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %cl
; CHECK-BASELINE-NEXT: xorb {{[0-9]+}}(%rsp), %cl
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %al
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: xorb %r11b, %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %r9b
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %r8b
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %r10b
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %dl
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %sil
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %al
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %cl
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %bl
; CHECK-BASELINE-NEXT: xorb %r13b, %sil
; CHECK-BASELINE-NEXT: xorb %r12b, %dl
; CHECK-BASELINE-NEXT: xorb %r15b, %r10b
; CHECK-BASELINE-NEXT: xorb %r14b, %r8b
; CHECK-BASELINE-NEXT: xorb %bpl, %r9b
; CHECK-BASELINE-NEXT: xorb {{[0-9]+}}(%rsp), %bl
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: xorb {{[0-9]+}}(%rsp), %cl
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: xorb %r11b, %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb %al, 7(%rdi)
; CHECK-BASELINE-NEXT: movb %cl, 6(%rdi)
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movb %bl, 5(%rdi)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb %r9b, 4(%rdi)
; CHECK-BASELINE-NEXT: movb %r8b, 3(%rdi)
; CHECK-BASELINE-NEXT: movb %r10b, 2(%rdi)
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movb %dl, 1(%rdi)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb %sil, (%rdi)
; CHECK-BASELINE-NEXT: movq %rdi, %rax
; CHECK-BASELINE-NEXT: popq %rbx
; CHECK-BASELINE-NEXT: popq %r12
; CHECK-BASELINE-NEXT: popq %r13
; CHECK-BASELINE-NEXT: popq %r14
; CHECK-BASELINE-NEXT: popq %r15
; CHECK-BASELINE-NEXT: popq %rbp
; CHECK-BASELINE-NEXT: retq
;
; CHECK-SSE1-LABEL: in_v8i8:
; CHECK-SSE1: # %bb.0:
; CHECK-SSE1-NEXT: pushq %rbp
; CHECK-SSE1-NEXT: pushq %r15
; CHECK-SSE1-NEXT: pushq %r14
; CHECK-SSE1-NEXT: pushq %r13
; CHECK-SSE1-NEXT: pushq %r12
; CHECK-SSE1-NEXT: pushq %rbx
; CHECK-SSE1-NEXT: movl %ecx, %r10d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %r11b
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %bpl
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %r14b
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %r15b
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %r12b
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %r13b
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: xorb %r13b, %sil
; CHECK-SSE1-NEXT: xorb %r12b, %dl
; CHECK-SSE1-NEXT: xorb %r15b, %r10b
; CHECK-SSE1-NEXT: xorb %r14b, %r8b
; CHECK-SSE1-NEXT: xorb %bpl, %r9b
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %bl
; CHECK-SSE1-NEXT: xorb {{[0-9]+}}(%rsp), %bl
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %cl
; CHECK-SSE1-NEXT: xorb {{[0-9]+}}(%rsp), %cl
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %al
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: xorb %r11b, %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %r9b
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %r8b
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %r10b
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %dl
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %sil
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %al
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %cl
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %bl
; CHECK-SSE1-NEXT: xorb %r13b, %sil
; CHECK-SSE1-NEXT: xorb %r12b, %dl
; CHECK-SSE1-NEXT: xorb %r15b, %r10b
; CHECK-SSE1-NEXT: xorb %r14b, %r8b
; CHECK-SSE1-NEXT: xorb %bpl, %r9b
; CHECK-SSE1-NEXT: xorb {{[0-9]+}}(%rsp), %bl
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: xorb {{[0-9]+}}(%rsp), %cl
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: xorb %r11b, %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb %al, 7(%rdi)
; CHECK-SSE1-NEXT: movb %cl, 6(%rdi)
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movb %bl, 5(%rdi)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb %r9b, 4(%rdi)
; CHECK-SSE1-NEXT: movb %r8b, 3(%rdi)
; CHECK-SSE1-NEXT: movb %r10b, 2(%rdi)
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movb %dl, 1(%rdi)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb %sil, (%rdi)
; CHECK-SSE1-NEXT: movq %rdi, %rax
; CHECK-SSE1-NEXT: popq %rbx
; CHECK-SSE1-NEXT: popq %r12
; CHECK-SSE1-NEXT: popq %r13
; CHECK-SSE1-NEXT: popq %r14
; CHECK-SSE1-NEXT: popq %r15
; CHECK-SSE1-NEXT: popq %rbp
; CHECK-SSE1-NEXT: retq
;
; CHECK-SSE2-LABEL: in_v8i8:
; CHECK-SSE2: # %bb.0:
; CHECK-SSE2-NEXT: andps %xmm2, %xmm0
[DAGCombine][X86][AArch64] Masked merge unfolding: vector edition. Summary: This **appears** to be the last missing piece for the masked merge pattern handling in the backend. This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`), and we need to make sure that they are generated. Differential Revision: https://reviews.llvm.org/D46528 llvm-svn: 332904
2018-05-22 05:41:02 +08:00
; CHECK-SSE2-NEXT: andnps %xmm1, %xmm2
; CHECK-SSE2-NEXT: orps %xmm2, %xmm0
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE2-NEXT: retq
;
; CHECK-XOP-LABEL: in_v8i8:
; CHECK-XOP: # %bb.0:
[DAGCombine][X86][AArch64] Masked merge unfolding: vector edition. Summary: This **appears** to be the last missing piece for the masked merge pattern handling in the backend. This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`), and we need to make sure that they are generated. Differential Revision: https://reviews.llvm.org/D46528 llvm-svn: 332904
2018-05-22 05:41:02 +08:00
; CHECK-XOP-NEXT: vpcmov %xmm2, %xmm1, %xmm0, %xmm0
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-XOP-NEXT: retq
%n0 = xor <8 x i8> %x, %y
%n1 = and <8 x i8> %n0, %mask
%r = xor <8 x i8> %n1, %y
ret <8 x i8> %r
}
define <4 x i16> @in_v4i16(<4 x i16> %x, <4 x i16> %y, <4 x i16> %mask) nounwind {
; CHECK-BASELINE-LABEL: in_v4i16:
; CHECK-BASELINE: # %bb.0:
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movq %rdi, %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl {{[0-9]+}}(%rsp), %r10d
; CHECK-BASELINE-NEXT: movl {{[0-9]+}}(%rsp), %r11d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movl {{[0-9]+}}(%rsp), %edi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: xorl %r9d, %esi
[DAGCombiner] use root SDLoc for all nodes created by logic fold If this is not a valid way to assign an SDLoc, then we get this wrong all over SDAG. I don't know enough about the SDAG to explain this. IIUC, theoretically, debug info is not supposed to affect codegen. But here it has clearly affected 3 different targets, and the x86 change is an actual improvement. llvm-svn: 348552
2018-12-07 08:01:57 +08:00
; CHECK-BASELINE-NEXT: xorl %edi, %edx
; CHECK-BASELINE-NEXT: xorl %r11d, %ecx
; CHECK-BASELINE-NEXT: xorl %r10d, %r8d
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andw {{[0-9]+}}(%rsp), %r8w
; CHECK-BASELINE-NEXT: andw {{[0-9]+}}(%rsp), %cx
; CHECK-BASELINE-NEXT: andw {{[0-9]+}}(%rsp), %dx
; CHECK-BASELINE-NEXT: andw {{[0-9]+}}(%rsp), %si
; CHECK-BASELINE-NEXT: xorl %r9d, %esi
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: xorl %edi, %edx
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: xorl %r11d, %ecx
; CHECK-BASELINE-NEXT: xorl %r10d, %r8d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movw %r8w, 6(%rax)
; CHECK-BASELINE-NEXT: movw %cx, 4(%rax)
; CHECK-BASELINE-NEXT: movw %dx, 2(%rax)
; CHECK-BASELINE-NEXT: movw %si, (%rax)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: retq
;
; CHECK-SSE1-LABEL: in_v4i16:
; CHECK-SSE1: # %bb.0:
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movq %rdi, %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movl {{[0-9]+}}(%rsp), %r10d
; CHECK-SSE1-NEXT: movl {{[0-9]+}}(%rsp), %r11d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movl {{[0-9]+}}(%rsp), %edi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: xorl %r9d, %esi
[DAGCombiner] use root SDLoc for all nodes created by logic fold If this is not a valid way to assign an SDLoc, then we get this wrong all over SDAG. I don't know enough about the SDAG to explain this. IIUC, theoretically, debug info is not supposed to affect codegen. But here it has clearly affected 3 different targets, and the x86 change is an actual improvement. llvm-svn: 348552
2018-12-07 08:01:57 +08:00
; CHECK-SSE1-NEXT: xorl %edi, %edx
; CHECK-SSE1-NEXT: xorl %r11d, %ecx
; CHECK-SSE1-NEXT: xorl %r10d, %r8d
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: andw {{[0-9]+}}(%rsp), %r8w
; CHECK-SSE1-NEXT: andw {{[0-9]+}}(%rsp), %cx
; CHECK-SSE1-NEXT: andw {{[0-9]+}}(%rsp), %dx
; CHECK-SSE1-NEXT: andw {{[0-9]+}}(%rsp), %si
; CHECK-SSE1-NEXT: xorl %r9d, %esi
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: xorl %edi, %edx
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: xorl %r11d, %ecx
; CHECK-SSE1-NEXT: xorl %r10d, %r8d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movw %r8w, 6(%rax)
; CHECK-SSE1-NEXT: movw %cx, 4(%rax)
; CHECK-SSE1-NEXT: movw %dx, 2(%rax)
; CHECK-SSE1-NEXT: movw %si, (%rax)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: retq
;
; CHECK-SSE2-LABEL: in_v4i16:
; CHECK-SSE2: # %bb.0:
; CHECK-SSE2-NEXT: andps %xmm2, %xmm0
[DAGCombine][X86][AArch64] Masked merge unfolding: vector edition. Summary: This **appears** to be the last missing piece for the masked merge pattern handling in the backend. This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`), and we need to make sure that they are generated. Differential Revision: https://reviews.llvm.org/D46528 llvm-svn: 332904
2018-05-22 05:41:02 +08:00
; CHECK-SSE2-NEXT: andnps %xmm1, %xmm2
; CHECK-SSE2-NEXT: orps %xmm2, %xmm0
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE2-NEXT: retq
;
; CHECK-XOP-LABEL: in_v4i16:
; CHECK-XOP: # %bb.0:
[DAGCombine][X86][AArch64] Masked merge unfolding: vector edition. Summary: This **appears** to be the last missing piece for the masked merge pattern handling in the backend. This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`), and we need to make sure that they are generated. Differential Revision: https://reviews.llvm.org/D46528 llvm-svn: 332904
2018-05-22 05:41:02 +08:00
; CHECK-XOP-NEXT: vpcmov %xmm2, %xmm1, %xmm0, %xmm0
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-XOP-NEXT: retq
%n0 = xor <4 x i16> %x, %y
%n1 = and <4 x i16> %n0, %mask
%r = xor <4 x i16> %n1, %y
ret <4 x i16> %r
}
define <2 x i32> @in_v2i32(<2 x i32> %x, <2 x i32> %y, <2 x i32> %mask) nounwind {
; CHECK-BASELINE-LABEL: in_v2i32:
; CHECK-BASELINE: # %bb.0:
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movl %edi, %eax
; CHECK-BASELINE-NEXT: xorl %edx, %eax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: xorl %ecx, %esi
; CHECK-BASELINE-NEXT: andl %r9d, %esi
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: andl %r8d, %eax
; CHECK-BASELINE-NEXT: xorl %edx, %eax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: xorl %ecx, %esi
; CHECK-BASELINE-NEXT: movl %esi, %edx
; CHECK-BASELINE-NEXT: retq
;
; CHECK-SSE1-LABEL: in_v2i32:
; CHECK-SSE1: # %bb.0:
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movl %edi, %eax
; CHECK-SSE1-NEXT: xorl %edx, %eax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: xorl %ecx, %esi
; CHECK-SSE1-NEXT: andl %r9d, %esi
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: andl %r8d, %eax
; CHECK-SSE1-NEXT: xorl %edx, %eax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: xorl %ecx, %esi
; CHECK-SSE1-NEXT: movl %esi, %edx
; CHECK-SSE1-NEXT: retq
;
; CHECK-SSE2-LABEL: in_v2i32:
; CHECK-SSE2: # %bb.0:
; CHECK-SSE2-NEXT: andps %xmm2, %xmm0
[DAGCombine][X86][AArch64] Masked merge unfolding: vector edition. Summary: This **appears** to be the last missing piece for the masked merge pattern handling in the backend. This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`), and we need to make sure that they are generated. Differential Revision: https://reviews.llvm.org/D46528 llvm-svn: 332904
2018-05-22 05:41:02 +08:00
; CHECK-SSE2-NEXT: andnps %xmm1, %xmm2
; CHECK-SSE2-NEXT: orps %xmm2, %xmm0
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE2-NEXT: retq
;
; CHECK-XOP-LABEL: in_v2i32:
; CHECK-XOP: # %bb.0:
[DAGCombine][X86][AArch64] Masked merge unfolding: vector edition. Summary: This **appears** to be the last missing piece for the masked merge pattern handling in the backend. This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`), and we need to make sure that they are generated. Differential Revision: https://reviews.llvm.org/D46528 llvm-svn: 332904
2018-05-22 05:41:02 +08:00
; CHECK-XOP-NEXT: vpcmov %xmm2, %xmm1, %xmm0, %xmm0
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-XOP-NEXT: retq
%n0 = xor <2 x i32> %x, %y
%n1 = and <2 x i32> %n0, %mask
%r = xor <2 x i32> %n1, %y
ret <2 x i32> %r
}
define <1 x i64> @in_v1i64(<1 x i64> %x, <1 x i64> %y, <1 x i64> %mask) nounwind {
; CHECK-LABEL: in_v1i64:
; CHECK: # %bb.0:
; CHECK-NEXT: movq %rdi, %rax
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-NEXT: xorq %rsi, %rax
; CHECK-NEXT: andq %rdx, %rax
; CHECK-NEXT: xorq %rsi, %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-NEXT: retq
%n0 = xor <1 x i64> %x, %y
%n1 = and <1 x i64> %n0, %mask
%r = xor <1 x i64> %n1, %y
ret <1 x i64> %r
}
; ============================================================================ ;
; 128-bit vector width
; ============================================================================ ;
define <16 x i8> @in_v16i8(<16 x i8> %x, <16 x i8> %y, <16 x i8> %mask) nounwind {
; CHECK-BASELINE-LABEL: in_v16i8:
; CHECK-BASELINE: # %bb.0:
; CHECK-BASELINE-NEXT: pushq %rbp
; CHECK-BASELINE-NEXT: pushq %r15
; CHECK-BASELINE-NEXT: pushq %r14
; CHECK-BASELINE-NEXT: pushq %r13
; CHECK-BASELINE-NEXT: pushq %r12
; CHECK-BASELINE-NEXT: pushq %rbx
; CHECK-BASELINE-NEXT: movl %ecx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-BASELINE-NEXT: movl %edx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-BASELINE-NEXT: movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movq %rdi, %rdx
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %sil
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %cl
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %al
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %bpl
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %r14b
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %r15b
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %r12b
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %r13b
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %bl
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %r11b
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %r10b
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %dil
; CHECK-BASELINE-NEXT: xorb %dil, %r9b
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %r9b
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: xorb %dil, %r9b
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %dil
; CHECK-BASELINE-NEXT: xorb %r10b, %dil
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %dil
; CHECK-BASELINE-NEXT: xorb %r10b, %dil
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %r10b
; CHECK-BASELINE-NEXT: xorb %r11b, %r10b
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %r10b
; CHECK-BASELINE-NEXT: xorb %r11b, %r10b
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %r11b
; CHECK-BASELINE-NEXT: xorb %bl, %r11b
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %r11b
; CHECK-BASELINE-NEXT: xorb %bl, %r11b
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %bl
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: xorb %r13b, %bl
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %bl
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: xorb %r13b, %bl
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %r13b
; CHECK-BASELINE-NEXT: xorb %r12b, %r13b
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %r13b
; CHECK-BASELINE-NEXT: xorb %r12b, %r13b
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %r12b
; CHECK-BASELINE-NEXT: xorb %r15b, %r12b
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %r12b
; CHECK-BASELINE-NEXT: xorb %r15b, %r12b
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %r15b
; CHECK-BASELINE-NEXT: xorb %r14b, %r15b
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %r15b
; CHECK-BASELINE-NEXT: xorb %r14b, %r15b
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %r14b
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: xorb %bpl, %r14b
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %r14b
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: xorb %bpl, %r14b
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %bpl
; CHECK-BASELINE-NEXT: xorb %al, %bpl
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %bpl
; CHECK-BASELINE-NEXT: xorb %al, %bpl
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %al
; CHECK-BASELINE-NEXT: xorb %cl, %al
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %al
; CHECK-BASELINE-NEXT: xorb %cl, %al
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %cl
; CHECK-BASELINE-NEXT: xorb %sil, %cl
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %cl
; CHECK-BASELINE-NEXT: xorb %sil, %cl
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movb %cl, 15(%rdx)
; CHECK-BASELINE-NEXT: movb %al, 14(%rdx)
; CHECK-BASELINE-NEXT: movb %bpl, 13(%rdx)
; CHECK-BASELINE-NEXT: movb %r14b, 12(%rdx)
; CHECK-BASELINE-NEXT: movb %r15b, 11(%rdx)
; CHECK-BASELINE-NEXT: movb %r12b, 10(%rdx)
; CHECK-BASELINE-NEXT: movb %r13b, 9(%rdx)
; CHECK-BASELINE-NEXT: movb %bl, 8(%rdx)
; CHECK-BASELINE-NEXT: movb %r11b, 7(%rdx)
; CHECK-BASELINE-NEXT: movb %r10b, 6(%rdx)
; CHECK-BASELINE-NEXT: movb %dil, 5(%rdx)
; CHECK-BASELINE-NEXT: movb %r9b, 4(%rdx)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %al
; CHECK-BASELINE-NEXT: xorb %al, %r8b
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %r8b
; CHECK-BASELINE-NEXT: xorb %al, %r8b
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movb %r8b, 3(%rdx)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %al
; CHECK-BASELINE-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %ecx # 4-byte Reload
; CHECK-BASELINE-NEXT: xorb %al, %cl
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %cl
; CHECK-BASELINE-NEXT: xorb %al, %cl
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movb %cl, 2(%rdx)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %al
; CHECK-BASELINE-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %ecx # 4-byte Reload
; CHECK-BASELINE-NEXT: xorb %al, %cl
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %cl
; CHECK-BASELINE-NEXT: xorb %al, %cl
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movb %cl, 1(%rdx)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb {{[0-9]+}}(%rsp), %al
; CHECK-BASELINE-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %ecx # 4-byte Reload
; CHECK-BASELINE-NEXT: xorb %al, %cl
; CHECK-BASELINE-NEXT: andb {{[0-9]+}}(%rsp), %cl
; CHECK-BASELINE-NEXT: xorb %al, %cl
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movb %cl, (%rdx)
; CHECK-BASELINE-NEXT: movq %rdx, %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: popq %rbx
; CHECK-BASELINE-NEXT: popq %r12
; CHECK-BASELINE-NEXT: popq %r13
; CHECK-BASELINE-NEXT: popq %r14
; CHECK-BASELINE-NEXT: popq %r15
; CHECK-BASELINE-NEXT: popq %rbp
; CHECK-BASELINE-NEXT: retq
;
; CHECK-SSE1-LABEL: in_v16i8:
; CHECK-SSE1: # %bb.0:
; CHECK-SSE1-NEXT: pushq %rbp
; CHECK-SSE1-NEXT: pushq %r15
; CHECK-SSE1-NEXT: pushq %r14
; CHECK-SSE1-NEXT: pushq %r13
; CHECK-SSE1-NEXT: pushq %r12
; CHECK-SSE1-NEXT: pushq %rbx
; CHECK-SSE1-NEXT: movl %ecx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-SSE1-NEXT: movl %edx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-SSE1-NEXT: movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movq %rdi, %rdx
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %sil
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %cl
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %al
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %bpl
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %r14b
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %r15b
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %r12b
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %r13b
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %bl
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %r11b
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %r10b
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %dil
; CHECK-SSE1-NEXT: xorb %dil, %r9b
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %r9b
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: xorb %dil, %r9b
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %dil
; CHECK-SSE1-NEXT: xorb %r10b, %dil
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %dil
; CHECK-SSE1-NEXT: xorb %r10b, %dil
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %r10b
; CHECK-SSE1-NEXT: xorb %r11b, %r10b
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %r10b
; CHECK-SSE1-NEXT: xorb %r11b, %r10b
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %r11b
; CHECK-SSE1-NEXT: xorb %bl, %r11b
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %r11b
; CHECK-SSE1-NEXT: xorb %bl, %r11b
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %bl
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: xorb %r13b, %bl
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %bl
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: xorb %r13b, %bl
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %r13b
; CHECK-SSE1-NEXT: xorb %r12b, %r13b
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %r13b
; CHECK-SSE1-NEXT: xorb %r12b, %r13b
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %r12b
; CHECK-SSE1-NEXT: xorb %r15b, %r12b
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %r12b
; CHECK-SSE1-NEXT: xorb %r15b, %r12b
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %r15b
; CHECK-SSE1-NEXT: xorb %r14b, %r15b
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %r15b
; CHECK-SSE1-NEXT: xorb %r14b, %r15b
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %r14b
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: xorb %bpl, %r14b
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %r14b
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: xorb %bpl, %r14b
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %bpl
; CHECK-SSE1-NEXT: xorb %al, %bpl
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %bpl
; CHECK-SSE1-NEXT: xorb %al, %bpl
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %al
; CHECK-SSE1-NEXT: xorb %cl, %al
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %al
; CHECK-SSE1-NEXT: xorb %cl, %al
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %cl
; CHECK-SSE1-NEXT: xorb %sil, %cl
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %cl
; CHECK-SSE1-NEXT: xorb %sil, %cl
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movb %cl, 15(%rdx)
; CHECK-SSE1-NEXT: movb %al, 14(%rdx)
; CHECK-SSE1-NEXT: movb %bpl, 13(%rdx)
; CHECK-SSE1-NEXT: movb %r14b, 12(%rdx)
; CHECK-SSE1-NEXT: movb %r15b, 11(%rdx)
; CHECK-SSE1-NEXT: movb %r12b, 10(%rdx)
; CHECK-SSE1-NEXT: movb %r13b, 9(%rdx)
; CHECK-SSE1-NEXT: movb %bl, 8(%rdx)
; CHECK-SSE1-NEXT: movb %r11b, 7(%rdx)
; CHECK-SSE1-NEXT: movb %r10b, 6(%rdx)
; CHECK-SSE1-NEXT: movb %dil, 5(%rdx)
; CHECK-SSE1-NEXT: movb %r9b, 4(%rdx)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %al
; CHECK-SSE1-NEXT: xorb %al, %r8b
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %r8b
; CHECK-SSE1-NEXT: xorb %al, %r8b
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movb %r8b, 3(%rdx)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %al
; CHECK-SSE1-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %ecx # 4-byte Reload
; CHECK-SSE1-NEXT: xorb %al, %cl
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %cl
; CHECK-SSE1-NEXT: xorb %al, %cl
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movb %cl, 2(%rdx)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %al
; CHECK-SSE1-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %ecx # 4-byte Reload
; CHECK-SSE1-NEXT: xorb %al, %cl
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %cl
; CHECK-SSE1-NEXT: xorb %al, %cl
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movb %cl, 1(%rdx)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb {{[0-9]+}}(%rsp), %al
; CHECK-SSE1-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %ecx # 4-byte Reload
; CHECK-SSE1-NEXT: xorb %al, %cl
; CHECK-SSE1-NEXT: andb {{[0-9]+}}(%rsp), %cl
; CHECK-SSE1-NEXT: xorb %al, %cl
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movb %cl, (%rdx)
; CHECK-SSE1-NEXT: movq %rdx, %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: popq %rbx
; CHECK-SSE1-NEXT: popq %r12
; CHECK-SSE1-NEXT: popq %r13
; CHECK-SSE1-NEXT: popq %r14
; CHECK-SSE1-NEXT: popq %r15
; CHECK-SSE1-NEXT: popq %rbp
; CHECK-SSE1-NEXT: retq
;
; CHECK-SSE2-LABEL: in_v16i8:
; CHECK-SSE2: # %bb.0:
; CHECK-SSE2-NEXT: andps %xmm2, %xmm0
[DAGCombine][X86][AArch64] Masked merge unfolding: vector edition. Summary: This **appears** to be the last missing piece for the masked merge pattern handling in the backend. This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`), and we need to make sure that they are generated. Differential Revision: https://reviews.llvm.org/D46528 llvm-svn: 332904
2018-05-22 05:41:02 +08:00
; CHECK-SSE2-NEXT: andnps %xmm1, %xmm2
; CHECK-SSE2-NEXT: orps %xmm2, %xmm0
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE2-NEXT: retq
;
; CHECK-XOP-LABEL: in_v16i8:
; CHECK-XOP: # %bb.0:
[DAGCombine][X86][AArch64] Masked merge unfolding: vector edition. Summary: This **appears** to be the last missing piece for the masked merge pattern handling in the backend. This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`), and we need to make sure that they are generated. Differential Revision: https://reviews.llvm.org/D46528 llvm-svn: 332904
2018-05-22 05:41:02 +08:00
; CHECK-XOP-NEXT: vpcmov %xmm2, %xmm1, %xmm0, %xmm0
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-XOP-NEXT: retq
%n0 = xor <16 x i8> %x, %y
%n1 = and <16 x i8> %n0, %mask
%r = xor <16 x i8> %n1, %y
ret <16 x i8> %r
}
define <8 x i16> @in_v8i16(<8 x i16> %x, <8 x i16> %y, <8 x i16> %mask) nounwind {
; CHECK-BASELINE-LABEL: in_v8i16:
; CHECK-BASELINE: # %bb.0:
; CHECK-BASELINE-NEXT: pushq %rbp
; CHECK-BASELINE-NEXT: pushq %rbx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movq %rdi, %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl {{[0-9]+}}(%rsp), %r10d
; CHECK-BASELINE-NEXT: movl {{[0-9]+}}(%rsp), %r11d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movl {{[0-9]+}}(%rsp), %edi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl {{[0-9]+}}(%rsp), %ebx
; CHECK-BASELINE-NEXT: xorl %ebx, %esi
; CHECK-BASELINE-NEXT: andw {{[0-9]+}}(%rsp), %si
; CHECK-BASELINE-NEXT: xorl %ebx, %esi
; CHECK-BASELINE-NEXT: movl {{[0-9]+}}(%rsp), %ebx
; CHECK-BASELINE-NEXT: xorl %ebx, %edx
; CHECK-BASELINE-NEXT: andw {{[0-9]+}}(%rsp), %dx
; CHECK-BASELINE-NEXT: xorl %ebx, %edx
[DAGCombiner] use root SDLoc for all nodes created by logic fold If this is not a valid way to assign an SDLoc, then we get this wrong all over SDAG. I don't know enough about the SDAG to explain this. IIUC, theoretically, debug info is not supposed to affect codegen. But here it has clearly affected 3 different targets, and the x86 change is an actual improvement. llvm-svn: 348552
2018-12-07 08:01:57 +08:00
; CHECK-BASELINE-NEXT: movl {{[0-9]+}}(%rsp), %ebx
; CHECK-BASELINE-NEXT: xorl %ebx, %ecx
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andw {{[0-9]+}}(%rsp), %cx
[DAGCombiner] use root SDLoc for all nodes created by logic fold If this is not a valid way to assign an SDLoc, then we get this wrong all over SDAG. I don't know enough about the SDAG to explain this. IIUC, theoretically, debug info is not supposed to affect codegen. But here it has clearly affected 3 different targets, and the x86 change is an actual improvement. llvm-svn: 348552
2018-12-07 08:01:57 +08:00
; CHECK-BASELINE-NEXT: xorl %ebx, %ecx
; CHECK-BASELINE-NEXT: movl {{[0-9]+}}(%rsp), %ebx
; CHECK-BASELINE-NEXT: xorl %ebx, %r8d
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andw {{[0-9]+}}(%rsp), %r8w
[DAGCombiner] use root SDLoc for all nodes created by logic fold If this is not a valid way to assign an SDLoc, then we get this wrong all over SDAG. I don't know enough about the SDAG to explain this. IIUC, theoretically, debug info is not supposed to affect codegen. But here it has clearly affected 3 different targets, and the x86 change is an actual improvement. llvm-svn: 348552
2018-12-07 08:01:57 +08:00
; CHECK-BASELINE-NEXT: xorl %ebx, %r8d
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl {{[0-9]+}}(%rsp), %ebx
[DAGCombiner] use root SDLoc for all nodes created by logic fold If this is not a valid way to assign an SDLoc, then we get this wrong all over SDAG. I don't know enough about the SDAG to explain this. IIUC, theoretically, debug info is not supposed to affect codegen. But here it has clearly affected 3 different targets, and the x86 change is an actual improvement. llvm-svn: 348552
2018-12-07 08:01:57 +08:00
; CHECK-BASELINE-NEXT: xorl %ebx, %r9d
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andw {{[0-9]+}}(%rsp), %r9w
[DAGCombiner] use root SDLoc for all nodes created by logic fold If this is not a valid way to assign an SDLoc, then we get this wrong all over SDAG. I don't know enough about the SDAG to explain this. IIUC, theoretically, debug info is not supposed to affect codegen. But here it has clearly affected 3 different targets, and the x86 change is an actual improvement. llvm-svn: 348552
2018-12-07 08:01:57 +08:00
; CHECK-BASELINE-NEXT: xorl %ebx, %r9d
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movzwl {{[0-9]+}}(%rsp), %ebp
[DAGCombiner] use root SDLoc for all nodes created by logic fold If this is not a valid way to assign an SDLoc, then we get this wrong all over SDAG. I don't know enough about the SDAG to explain this. IIUC, theoretically, debug info is not supposed to affect codegen. But here it has clearly affected 3 different targets, and the x86 change is an actual improvement. llvm-svn: 348552
2018-12-07 08:01:57 +08:00
; CHECK-BASELINE-NEXT: xorw %di, %bp
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andw {{[0-9]+}}(%rsp), %bp
[DAGCombiner] use root SDLoc for all nodes created by logic fold If this is not a valid way to assign an SDLoc, then we get this wrong all over SDAG. I don't know enough about the SDAG to explain this. IIUC, theoretically, debug info is not supposed to affect codegen. But here it has clearly affected 3 different targets, and the x86 change is an actual improvement. llvm-svn: 348552
2018-12-07 08:01:57 +08:00
; CHECK-BASELINE-NEXT: xorl %edi, %ebp
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movzwl {{[0-9]+}}(%rsp), %edi
[DAGCombiner] use root SDLoc for all nodes created by logic fold If this is not a valid way to assign an SDLoc, then we get this wrong all over SDAG. I don't know enough about the SDAG to explain this. IIUC, theoretically, debug info is not supposed to affect codegen. But here it has clearly affected 3 different targets, and the x86 change is an actual improvement. llvm-svn: 348552
2018-12-07 08:01:57 +08:00
; CHECK-BASELINE-NEXT: xorw %r11w, %di
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: andw {{[0-9]+}}(%rsp), %di
[DAGCombiner] use root SDLoc for all nodes created by logic fold If this is not a valid way to assign an SDLoc, then we get this wrong all over SDAG. I don't know enough about the SDAG to explain this. IIUC, theoretically, debug info is not supposed to affect codegen. But here it has clearly affected 3 different targets, and the x86 change is an actual improvement. llvm-svn: 348552
2018-12-07 08:01:57 +08:00
; CHECK-BASELINE-NEXT: xorl %r11d, %edi
; CHECK-BASELINE-NEXT: movzwl {{[0-9]+}}(%rsp), %ebx
; CHECK-BASELINE-NEXT: xorw %r10w, %bx
; CHECK-BASELINE-NEXT: andw {{[0-9]+}}(%rsp), %bx
; CHECK-BASELINE-NEXT: xorl %r10d, %ebx
; CHECK-BASELINE-NEXT: movw %bx, 14(%rax)
; CHECK-BASELINE-NEXT: movw %di, 12(%rax)
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movw %bp, 10(%rax)
; CHECK-BASELINE-NEXT: movw %r9w, 8(%rax)
; CHECK-BASELINE-NEXT: movw %r8w, 6(%rax)
; CHECK-BASELINE-NEXT: movw %cx, 4(%rax)
; CHECK-BASELINE-NEXT: movw %dx, 2(%rax)
; CHECK-BASELINE-NEXT: movw %si, (%rax)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: popq %rbx
; CHECK-BASELINE-NEXT: popq %rbp
; CHECK-BASELINE-NEXT: retq
;
; CHECK-SSE1-LABEL: in_v8i16:
; CHECK-SSE1: # %bb.0:
; CHECK-SSE1-NEXT: pushq %rbp
; CHECK-SSE1-NEXT: pushq %rbx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movq %rdi, %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movl {{[0-9]+}}(%rsp), %r10d
; CHECK-SSE1-NEXT: movl {{[0-9]+}}(%rsp), %r11d
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movl {{[0-9]+}}(%rsp), %edi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movl {{[0-9]+}}(%rsp), %ebx
; CHECK-SSE1-NEXT: xorl %ebx, %esi
; CHECK-SSE1-NEXT: andw {{[0-9]+}}(%rsp), %si
; CHECK-SSE1-NEXT: xorl %ebx, %esi
; CHECK-SSE1-NEXT: movl {{[0-9]+}}(%rsp), %ebx
; CHECK-SSE1-NEXT: xorl %ebx, %edx
; CHECK-SSE1-NEXT: andw {{[0-9]+}}(%rsp), %dx
; CHECK-SSE1-NEXT: xorl %ebx, %edx
[DAGCombiner] use root SDLoc for all nodes created by logic fold If this is not a valid way to assign an SDLoc, then we get this wrong all over SDAG. I don't know enough about the SDAG to explain this. IIUC, theoretically, debug info is not supposed to affect codegen. But here it has clearly affected 3 different targets, and the x86 change is an actual improvement. llvm-svn: 348552
2018-12-07 08:01:57 +08:00
; CHECK-SSE1-NEXT: movl {{[0-9]+}}(%rsp), %ebx
; CHECK-SSE1-NEXT: xorl %ebx, %ecx
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: andw {{[0-9]+}}(%rsp), %cx
[DAGCombiner] use root SDLoc for all nodes created by logic fold If this is not a valid way to assign an SDLoc, then we get this wrong all over SDAG. I don't know enough about the SDAG to explain this. IIUC, theoretically, debug info is not supposed to affect codegen. But here it has clearly affected 3 different targets, and the x86 change is an actual improvement. llvm-svn: 348552
2018-12-07 08:01:57 +08:00
; CHECK-SSE1-NEXT: xorl %ebx, %ecx
; CHECK-SSE1-NEXT: movl {{[0-9]+}}(%rsp), %ebx
; CHECK-SSE1-NEXT: xorl %ebx, %r8d
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: andw {{[0-9]+}}(%rsp), %r8w
[DAGCombiner] use root SDLoc for all nodes created by logic fold If this is not a valid way to assign an SDLoc, then we get this wrong all over SDAG. I don't know enough about the SDAG to explain this. IIUC, theoretically, debug info is not supposed to affect codegen. But here it has clearly affected 3 different targets, and the x86 change is an actual improvement. llvm-svn: 348552
2018-12-07 08:01:57 +08:00
; CHECK-SSE1-NEXT: xorl %ebx, %r8d
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movl {{[0-9]+}}(%rsp), %ebx
[DAGCombiner] use root SDLoc for all nodes created by logic fold If this is not a valid way to assign an SDLoc, then we get this wrong all over SDAG. I don't know enough about the SDAG to explain this. IIUC, theoretically, debug info is not supposed to affect codegen. But here it has clearly affected 3 different targets, and the x86 change is an actual improvement. llvm-svn: 348552
2018-12-07 08:01:57 +08:00
; CHECK-SSE1-NEXT: xorl %ebx, %r9d
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: andw {{[0-9]+}}(%rsp), %r9w
[DAGCombiner] use root SDLoc for all nodes created by logic fold If this is not a valid way to assign an SDLoc, then we get this wrong all over SDAG. I don't know enough about the SDAG to explain this. IIUC, theoretically, debug info is not supposed to affect codegen. But here it has clearly affected 3 different targets, and the x86 change is an actual improvement. llvm-svn: 348552
2018-12-07 08:01:57 +08:00
; CHECK-SSE1-NEXT: xorl %ebx, %r9d
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movzwl {{[0-9]+}}(%rsp), %ebp
[DAGCombiner] use root SDLoc for all nodes created by logic fold If this is not a valid way to assign an SDLoc, then we get this wrong all over SDAG. I don't know enough about the SDAG to explain this. IIUC, theoretically, debug info is not supposed to affect codegen. But here it has clearly affected 3 different targets, and the x86 change is an actual improvement. llvm-svn: 348552
2018-12-07 08:01:57 +08:00
; CHECK-SSE1-NEXT: xorw %di, %bp
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: andw {{[0-9]+}}(%rsp), %bp
[DAGCombiner] use root SDLoc for all nodes created by logic fold If this is not a valid way to assign an SDLoc, then we get this wrong all over SDAG. I don't know enough about the SDAG to explain this. IIUC, theoretically, debug info is not supposed to affect codegen. But here it has clearly affected 3 different targets, and the x86 change is an actual improvement. llvm-svn: 348552
2018-12-07 08:01:57 +08:00
; CHECK-SSE1-NEXT: xorl %edi, %ebp
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movzwl {{[0-9]+}}(%rsp), %edi
[DAGCombiner] use root SDLoc for all nodes created by logic fold If this is not a valid way to assign an SDLoc, then we get this wrong all over SDAG. I don't know enough about the SDAG to explain this. IIUC, theoretically, debug info is not supposed to affect codegen. But here it has clearly affected 3 different targets, and the x86 change is an actual improvement. llvm-svn: 348552
2018-12-07 08:01:57 +08:00
; CHECK-SSE1-NEXT: xorw %r11w, %di
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: andw {{[0-9]+}}(%rsp), %di
[DAGCombiner] use root SDLoc for all nodes created by logic fold If this is not a valid way to assign an SDLoc, then we get this wrong all over SDAG. I don't know enough about the SDAG to explain this. IIUC, theoretically, debug info is not supposed to affect codegen. But here it has clearly affected 3 different targets, and the x86 change is an actual improvement. llvm-svn: 348552
2018-12-07 08:01:57 +08:00
; CHECK-SSE1-NEXT: xorl %r11d, %edi
; CHECK-SSE1-NEXT: movzwl {{[0-9]+}}(%rsp), %ebx
; CHECK-SSE1-NEXT: xorw %r10w, %bx
; CHECK-SSE1-NEXT: andw {{[0-9]+}}(%rsp), %bx
; CHECK-SSE1-NEXT: xorl %r10d, %ebx
; CHECK-SSE1-NEXT: movw %bx, 14(%rax)
; CHECK-SSE1-NEXT: movw %di, 12(%rax)
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movw %bp, 10(%rax)
; CHECK-SSE1-NEXT: movw %r9w, 8(%rax)
; CHECK-SSE1-NEXT: movw %r8w, 6(%rax)
; CHECK-SSE1-NEXT: movw %cx, 4(%rax)
; CHECK-SSE1-NEXT: movw %dx, 2(%rax)
; CHECK-SSE1-NEXT: movw %si, (%rax)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: popq %rbx
; CHECK-SSE1-NEXT: popq %rbp
; CHECK-SSE1-NEXT: retq
;
; CHECK-SSE2-LABEL: in_v8i16:
; CHECK-SSE2: # %bb.0:
; CHECK-SSE2-NEXT: andps %xmm2, %xmm0
[DAGCombine][X86][AArch64] Masked merge unfolding: vector edition. Summary: This **appears** to be the last missing piece for the masked merge pattern handling in the backend. This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`), and we need to make sure that they are generated. Differential Revision: https://reviews.llvm.org/D46528 llvm-svn: 332904
2018-05-22 05:41:02 +08:00
; CHECK-SSE2-NEXT: andnps %xmm1, %xmm2
; CHECK-SSE2-NEXT: orps %xmm2, %xmm0
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE2-NEXT: retq
;
; CHECK-XOP-LABEL: in_v8i16:
; CHECK-XOP: # %bb.0:
[DAGCombine][X86][AArch64] Masked merge unfolding: vector edition. Summary: This **appears** to be the last missing piece for the masked merge pattern handling in the backend. This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`), and we need to make sure that they are generated. Differential Revision: https://reviews.llvm.org/D46528 llvm-svn: 332904
2018-05-22 05:41:02 +08:00
; CHECK-XOP-NEXT: vpcmov %xmm2, %xmm1, %xmm0, %xmm0
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-XOP-NEXT: retq
%n0 = xor <8 x i16> %x, %y
%n1 = and <8 x i16> %n0, %mask
%r = xor <8 x i16> %n1, %y
ret <8 x i16> %r
}
define <4 x i32> @in_v4i32(<4 x i32> *%px, <4 x i32> *%py, <4 x i32> *%pmask) nounwind {
; CHECK-BASELINE-LABEL: in_v4i32:
; CHECK-BASELINE: # %bb.0:
; CHECK-BASELINE-NEXT: pushq %rbx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movq %rdi, %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl 12(%rdx), %r8d
; CHECK-BASELINE-NEXT: movl 8(%rdx), %r9d
; CHECK-BASELINE-NEXT: movl (%rdx), %r11d
; CHECK-BASELINE-NEXT: movl 4(%rdx), %r10d
; CHECK-BASELINE-NEXT: movl (%rsi), %edx
; CHECK-BASELINE-NEXT: xorl %r11d, %edx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movl 4(%rsi), %edi
; CHECK-BASELINE-NEXT: xorl %r10d, %edi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl 8(%rsi), %ebx
; CHECK-BASELINE-NEXT: xorl %r9d, %ebx
; CHECK-BASELINE-NEXT: movl 12(%rsi), %esi
; CHECK-BASELINE-NEXT: xorl %r8d, %esi
; CHECK-BASELINE-NEXT: andl 12(%rcx), %esi
; CHECK-BASELINE-NEXT: andl 8(%rcx), %ebx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: andl 4(%rcx), %edi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andl (%rcx), %edx
; CHECK-BASELINE-NEXT: xorl %r11d, %edx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: xorl %r10d, %edi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: xorl %r9d, %ebx
; CHECK-BASELINE-NEXT: xorl %r8d, %esi
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movl %esi, 12(%rax)
; CHECK-BASELINE-NEXT: movl %ebx, 8(%rax)
; CHECK-BASELINE-NEXT: movl %edi, 4(%rax)
; CHECK-BASELINE-NEXT: movl %edx, (%rax)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: popq %rbx
; CHECK-BASELINE-NEXT: retq
;
; CHECK-SSE1-LABEL: in_v4i32:
; CHECK-SSE1: # %bb.0:
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movq %rdi, %rax
[DAGCombine][X86][AArch64] Masked merge unfolding: vector edition. Summary: This **appears** to be the last missing piece for the masked merge pattern handling in the backend. This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`), and we need to make sure that they are generated. Differential Revision: https://reviews.llvm.org/D46528 llvm-svn: 332904
2018-05-22 05:41:02 +08:00
; CHECK-SSE1-NEXT: movaps (%rcx), %xmm0
; CHECK-SSE1-NEXT: movaps %xmm0, %xmm1
; CHECK-SSE1-NEXT: andnps (%rdx), %xmm1
; CHECK-SSE1-NEXT: andps (%rsi), %xmm0
; CHECK-SSE1-NEXT: orps %xmm1, %xmm0
; CHECK-SSE1-NEXT: movaps %xmm0, (%rdi)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: retq
;
; CHECK-SSE2-LABEL: in_v4i32:
; CHECK-SSE2: # %bb.0:
[DAGCombine][X86][AArch64] Masked merge unfolding: vector edition. Summary: This **appears** to be the last missing piece for the masked merge pattern handling in the backend. This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`), and we need to make sure that they are generated. Differential Revision: https://reviews.llvm.org/D46528 llvm-svn: 332904
2018-05-22 05:41:02 +08:00
; CHECK-SSE2-NEXT: movaps (%rdx), %xmm0
; CHECK-SSE2-NEXT: movaps %xmm0, %xmm1
; CHECK-SSE2-NEXT: andnps (%rsi), %xmm1
; CHECK-SSE2-NEXT: andps (%rdi), %xmm0
; CHECK-SSE2-NEXT: orps %xmm1, %xmm0
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE2-NEXT: retq
;
; CHECK-XOP-LABEL: in_v4i32:
; CHECK-XOP: # %bb.0:
[DAGCombine][X86][AArch64] Masked merge unfolding: vector edition. Summary: This **appears** to be the last missing piece for the masked merge pattern handling in the backend. This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`), and we need to make sure that they are generated. Differential Revision: https://reviews.llvm.org/D46528 llvm-svn: 332904
2018-05-22 05:41:02 +08:00
; CHECK-XOP-NEXT: vmovdqa (%rdi), %xmm0
; CHECK-XOP-NEXT: vmovdqa (%rdx), %xmm1
; CHECK-XOP-NEXT: vpcmov %xmm1, (%rsi), %xmm0, %xmm0
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-XOP-NEXT: retq
%x = load <4 x i32>, <4 x i32> *%px, align 16
%y = load <4 x i32>, <4 x i32> *%py, align 16
%mask = load <4 x i32>, <4 x i32> *%pmask, align 16
%n0 = xor <4 x i32> %x, %y
%n1 = and <4 x i32> %n0, %mask
%r = xor <4 x i32> %n1, %y
ret <4 x i32> %r
}
define <2 x i64> @in_v2i64(<2 x i64> %x, <2 x i64> %y, <2 x i64> %mask) nounwind {
; CHECK-BASELINE-LABEL: in_v2i64:
; CHECK-BASELINE: # %bb.0:
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movq %rdi, %rax
; CHECK-BASELINE-NEXT: xorq %rdx, %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: xorq %rcx, %rsi
; CHECK-BASELINE-NEXT: andq %r9, %rsi
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: andq %r8, %rax
; CHECK-BASELINE-NEXT: xorq %rdx, %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: xorq %rcx, %rsi
; CHECK-BASELINE-NEXT: movq %rsi, %rdx
; CHECK-BASELINE-NEXT: retq
;
; CHECK-SSE1-LABEL: in_v2i64:
; CHECK-SSE1: # %bb.0:
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movq %rdi, %rax
; CHECK-SSE1-NEXT: xorq %rdx, %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: xorq %rcx, %rsi
; CHECK-SSE1-NEXT: andq %r9, %rsi
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: andq %r8, %rax
; CHECK-SSE1-NEXT: xorq %rdx, %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: xorq %rcx, %rsi
; CHECK-SSE1-NEXT: movq %rsi, %rdx
; CHECK-SSE1-NEXT: retq
;
; CHECK-SSE2-LABEL: in_v2i64:
; CHECK-SSE2: # %bb.0:
; CHECK-SSE2-NEXT: andps %xmm2, %xmm0
[DAGCombine][X86][AArch64] Masked merge unfolding: vector edition. Summary: This **appears** to be the last missing piece for the masked merge pattern handling in the backend. This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`), and we need to make sure that they are generated. Differential Revision: https://reviews.llvm.org/D46528 llvm-svn: 332904
2018-05-22 05:41:02 +08:00
; CHECK-SSE2-NEXT: andnps %xmm1, %xmm2
; CHECK-SSE2-NEXT: orps %xmm2, %xmm0
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE2-NEXT: retq
;
; CHECK-XOP-LABEL: in_v2i64:
; CHECK-XOP: # %bb.0:
[DAGCombine][X86][AArch64] Masked merge unfolding: vector edition. Summary: This **appears** to be the last missing piece for the masked merge pattern handling in the backend. This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`), and we need to make sure that they are generated. Differential Revision: https://reviews.llvm.org/D46528 llvm-svn: 332904
2018-05-22 05:41:02 +08:00
; CHECK-XOP-NEXT: vpcmov %xmm2, %xmm1, %xmm0, %xmm0
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-XOP-NEXT: retq
%n0 = xor <2 x i64> %x, %y
%n1 = and <2 x i64> %n0, %mask
%r = xor <2 x i64> %n1, %y
ret <2 x i64> %r
}
; ============================================================================ ;
; 256-bit vector width
; ============================================================================ ;
define <32 x i8> @in_v32i8(<32 x i8> *%px, <32 x i8> *%py, <32 x i8> *%pmask) nounwind {
; CHECK-BASELINE-LABEL: in_v32i8:
; CHECK-BASELINE: # %bb.0:
; CHECK-BASELINE-NEXT: pushq %rbp
; CHECK-BASELINE-NEXT: pushq %r15
; CHECK-BASELINE-NEXT: pushq %r14
; CHECK-BASELINE-NEXT: pushq %r13
; CHECK-BASELINE-NEXT: pushq %r12
; CHECK-BASELINE-NEXT: pushq %rbx
; CHECK-BASELINE-NEXT: movq %rdx, %r13
; CHECK-BASELINE-NEXT: movq %rsi, %rbx
; CHECK-BASELINE-NEXT: movq %rdi, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-BASELINE-NEXT: movb 15(%rdx), %r12b
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb 14(%rdx), %al
; CHECK-BASELINE-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 13(%rdx), %al
; CHECK-BASELINE-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 12(%rdx), %al
; CHECK-BASELINE-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 11(%rdx), %al
; CHECK-BASELINE-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 10(%rdx), %al
; CHECK-BASELINE-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-BASELINE-NEXT: movb 9(%rdx), %r9b
; CHECK-BASELINE-NEXT: movb 8(%rdx), %r10b
; CHECK-BASELINE-NEXT: movb 7(%rdx), %r11b
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb 6(%rdx), %r8b
; CHECK-BASELINE-NEXT: movb 5(%rdx), %bpl
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-BASELINE-NEXT: movb 4(%rdx), %sil
; CHECK-BASELINE-NEXT: movb 3(%rdx), %dil
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb 2(%rdx), %r14b
; CHECK-BASELINE-NEXT: movb (%rdx), %al
; CHECK-BASELINE-NEXT: movb 1(%rdx), %r15b
; CHECK-BASELINE-NEXT: movb (%rbx), %dl
; CHECK-BASELINE-NEXT: xorb %al, %dl
; CHECK-BASELINE-NEXT: andb (%rcx), %dl
; CHECK-BASELINE-NEXT: xorb %al, %dl
; CHECK-BASELINE-NEXT: movb %dl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 1(%rbx), %al
; CHECK-BASELINE-NEXT: xorb %r15b, %al
; CHECK-BASELINE-NEXT: andb 1(%rcx), %al
; CHECK-BASELINE-NEXT: xorb %r15b, %al
; CHECK-BASELINE-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 2(%rbx), %al
; CHECK-BASELINE-NEXT: xorb %r14b, %al
; CHECK-BASELINE-NEXT: andb 2(%rcx), %al
; CHECK-BASELINE-NEXT: xorb %r14b, %al
; CHECK-BASELINE-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 3(%rbx), %al
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-BASELINE-NEXT: xorb %dil, %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andb 3(%rcx), %al
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-BASELINE-NEXT: xorb %dil, %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 4(%rbx), %al
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-BASELINE-NEXT: xorb %sil, %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andb 4(%rcx), %al
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-BASELINE-NEXT: xorb %sil, %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 5(%rbx), %al
; CHECK-BASELINE-NEXT: xorb %bpl, %al
; CHECK-BASELINE-NEXT: andb 5(%rcx), %al
; CHECK-BASELINE-NEXT: xorb %bpl, %al
; CHECK-BASELINE-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 6(%rbx), %al
; CHECK-BASELINE-NEXT: xorb %r8b, %al
; CHECK-BASELINE-NEXT: andb 6(%rcx), %al
; CHECK-BASELINE-NEXT: xorb %r8b, %al
; CHECK-BASELINE-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 7(%rbx), %al
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-BASELINE-NEXT: xorb %r11b, %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andb 7(%rcx), %al
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-BASELINE-NEXT: xorb %r11b, %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 8(%rbx), %al
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-BASELINE-NEXT: xorb %r10b, %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andb 8(%rcx), %al
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-BASELINE-NEXT: xorb %r10b, %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 9(%rbx), %al
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-BASELINE-NEXT: xorb %r9b, %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andb 9(%rcx), %al
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-BASELINE-NEXT: xorb %r9b, %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 10(%rbx), %dl
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-BASELINE-NEXT: xorb %al, %dl
; CHECK-BASELINE-NEXT: andb 10(%rcx), %dl
; CHECK-BASELINE-NEXT: xorb %al, %dl
; CHECK-BASELINE-NEXT: movb %dl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 11(%rbx), %dl
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-BASELINE-NEXT: xorb %al, %dl
; CHECK-BASELINE-NEXT: andb 11(%rcx), %dl
; CHECK-BASELINE-NEXT: xorb %al, %dl
; CHECK-BASELINE-NEXT: movb %dl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 12(%rbx), %dl
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-BASELINE-NEXT: xorb %al, %dl
; CHECK-BASELINE-NEXT: andb 12(%rcx), %dl
; CHECK-BASELINE-NEXT: xorb %al, %dl
; CHECK-BASELINE-NEXT: movb %dl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 13(%rbx), %dl
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-BASELINE-NEXT: xorb %al, %dl
; CHECK-BASELINE-NEXT: andb 13(%rcx), %dl
; CHECK-BASELINE-NEXT: xorb %al, %dl
; CHECK-BASELINE-NEXT: movb %dl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 14(%rbx), %dl
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-BASELINE-NEXT: xorb %al, %dl
; CHECK-BASELINE-NEXT: andb 14(%rcx), %dl
; CHECK-BASELINE-NEXT: xorb %al, %dl
; CHECK-BASELINE-NEXT: movb %dl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-BASELINE-NEXT: movb 15(%rbx), %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: xorb %r12b, %al
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-BASELINE-NEXT: andb 15(%rcx), %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: xorb %r12b, %al
; CHECK-BASELINE-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-BASELINE-NEXT: movb 16(%r13), %al
; CHECK-BASELINE-NEXT: movb 16(%rbx), %dl
; CHECK-BASELINE-NEXT: xorb %al, %dl
; CHECK-BASELINE-NEXT: andb 16(%rcx), %dl
; CHECK-BASELINE-NEXT: xorb %al, %dl
; CHECK-BASELINE-NEXT: movb %dl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb 17(%r13), %al
; CHECK-BASELINE-NEXT: movb 17(%rbx), %dl
; CHECK-BASELINE-NEXT: xorb %al, %dl
; CHECK-BASELINE-NEXT: andb 17(%rcx), %dl
; CHECK-BASELINE-NEXT: xorb %al, %dl
; CHECK-BASELINE-NEXT: movb %dl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 18(%r13), %al
; CHECK-BASELINE-NEXT: movb 18(%rbx), %dl
; CHECK-BASELINE-NEXT: xorb %al, %dl
; CHECK-BASELINE-NEXT: andb 18(%rcx), %dl
; CHECK-BASELINE-NEXT: xorb %al, %dl
; CHECK-BASELINE-NEXT: movb %dl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 19(%r13), %al
; CHECK-BASELINE-NEXT: movb 19(%rbx), %r12b
; CHECK-BASELINE-NEXT: xorb %al, %r12b
; CHECK-BASELINE-NEXT: andb 19(%rcx), %r12b
; CHECK-BASELINE-NEXT: xorb %al, %r12b
; CHECK-BASELINE-NEXT: movb 20(%r13), %al
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-BASELINE-NEXT: movb 20(%rbx), %r15b
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: xorb %al, %r15b
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-BASELINE-NEXT: andb 20(%rcx), %r15b
; CHECK-BASELINE-NEXT: movq %rcx, %rsi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: xorb %al, %r15b
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-BASELINE-NEXT: movb 21(%r13), %al
; CHECK-BASELINE-NEXT: movb 21(%rbx), %r14b
; CHECK-BASELINE-NEXT: xorb %al, %r14b
; CHECK-BASELINE-NEXT: andb 21(%rcx), %r14b
; CHECK-BASELINE-NEXT: xorb %al, %r14b
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb 22(%r13), %al
; CHECK-BASELINE-NEXT: movb 22(%rbx), %bpl
; CHECK-BASELINE-NEXT: xorb %al, %bpl
; CHECK-BASELINE-NEXT: andb 22(%rcx), %bpl
; CHECK-BASELINE-NEXT: xorb %al, %bpl
; CHECK-BASELINE-NEXT: movb 23(%r13), %al
; CHECK-BASELINE-NEXT: movb 23(%rbx), %r11b
; CHECK-BASELINE-NEXT: xorb %al, %r11b
; CHECK-BASELINE-NEXT: andb 23(%rcx), %r11b
; CHECK-BASELINE-NEXT: xorb %al, %r11b
; CHECK-BASELINE-NEXT: movb 24(%r13), %al
; CHECK-BASELINE-NEXT: movb 24(%rbx), %r10b
; CHECK-BASELINE-NEXT: xorb %al, %r10b
; CHECK-BASELINE-NEXT: andb 24(%rcx), %r10b
; CHECK-BASELINE-NEXT: xorb %al, %r10b
; CHECK-BASELINE-NEXT: movb 25(%r13), %al
; CHECK-BASELINE-NEXT: movb 25(%rbx), %r9b
; CHECK-BASELINE-NEXT: xorb %al, %r9b
; CHECK-BASELINE-NEXT: andb 25(%rcx), %r9b
; CHECK-BASELINE-NEXT: xorb %al, %r9b
; CHECK-BASELINE-NEXT: movb 26(%r13), %al
; CHECK-BASELINE-NEXT: movb 26(%rbx), %r8b
; CHECK-BASELINE-NEXT: xorb %al, %r8b
; CHECK-BASELINE-NEXT: andb 26(%rcx), %r8b
; CHECK-BASELINE-NEXT: xorb %al, %r8b
; CHECK-BASELINE-NEXT: movb 27(%r13), %al
; CHECK-BASELINE-NEXT: movb 27(%rbx), %dil
; CHECK-BASELINE-NEXT: xorb %al, %dil
; CHECK-BASELINE-NEXT: andb 27(%rcx), %dil
; CHECK-BASELINE-NEXT: xorb %al, %dil
; CHECK-BASELINE-NEXT: movb 28(%r13), %al
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-BASELINE-NEXT: movb 28(%rbx), %dl
; CHECK-BASELINE-NEXT: xorb %al, %dl
; CHECK-BASELINE-NEXT: andb 28(%rcx), %dl
; CHECK-BASELINE-NEXT: xorb %al, %dl
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb 29(%r13), %al
; CHECK-BASELINE-NEXT: movb 29(%rbx), %cl
; CHECK-BASELINE-NEXT: xorb %al, %cl
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-BASELINE-NEXT: andb 29(%rsi), %cl
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: xorb %al, %cl
; CHECK-BASELINE-NEXT: movb 30(%r13), %al
; CHECK-BASELINE-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-BASELINE-NEXT: movb 30(%rbx), %al
; CHECK-BASELINE-NEXT: xorb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Folded Reload
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-BASELINE-NEXT: andb 30(%rsi), %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: xorb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Folded Reload
; CHECK-BASELINE-NEXT: movb 31(%r13), %r13b
; CHECK-BASELINE-NEXT: movb 31(%rbx), %bl
; CHECK-BASELINE-NEXT: xorb %r13b, %bl
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-BASELINE-NEXT: andb 31(%rsi), %bl
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: xorb %r13b, %bl
; CHECK-BASELINE-NEXT: movq {{[-0-9]+}}(%r{{[sb]}}p), %r13 # 8-byte Reload
; CHECK-BASELINE-NEXT: movb %bl, 31(%r13)
; CHECK-BASELINE-NEXT: movb %al, 30(%r13)
; CHECK-BASELINE-NEXT: movb %cl, 29(%r13)
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-BASELINE-NEXT: movb %dl, 28(%r13)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb %dil, 27(%r13)
; CHECK-BASELINE-NEXT: movb %r8b, 26(%r13)
; CHECK-BASELINE-NEXT: movb %r9b, 25(%r13)
; CHECK-BASELINE-NEXT: movb %r10b, 24(%r13)
; CHECK-BASELINE-NEXT: movb %r11b, 23(%r13)
; CHECK-BASELINE-NEXT: movb %bpl, 22(%r13)
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-BASELINE-NEXT: movb %r14b, 21(%r13)
; CHECK-BASELINE-NEXT: movb %r15b, 20(%r13)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movb %r12b, 19(%r13)
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %al, 18(%r13)
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %al, 17(%r13)
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %al, 16(%r13)
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %al, 15(%r13)
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %al, 14(%r13)
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %al, 13(%r13)
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %al, 12(%r13)
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %al, 11(%r13)
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %al, 10(%r13)
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %al, 9(%r13)
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %al, 8(%r13)
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %al, 7(%r13)
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %al, 6(%r13)
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %al, 5(%r13)
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %al, 4(%r13)
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %al, 3(%r13)
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %al, 2(%r13)
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %al, 1(%r13)
; CHECK-BASELINE-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-BASELINE-NEXT: movb %al, (%r13)
; CHECK-BASELINE-NEXT: movq %r13, %rax
; CHECK-BASELINE-NEXT: popq %rbx
; CHECK-BASELINE-NEXT: popq %r12
; CHECK-BASELINE-NEXT: popq %r13
; CHECK-BASELINE-NEXT: popq %r14
; CHECK-BASELINE-NEXT: popq %r15
; CHECK-BASELINE-NEXT: popq %rbp
; CHECK-BASELINE-NEXT: retq
;
; CHECK-SSE1-LABEL: in_v32i8:
; CHECK-SSE1: # %bb.0:
; CHECK-SSE1-NEXT: pushq %rbp
; CHECK-SSE1-NEXT: pushq %r15
; CHECK-SSE1-NEXT: pushq %r14
; CHECK-SSE1-NEXT: pushq %r13
; CHECK-SSE1-NEXT: pushq %r12
; CHECK-SSE1-NEXT: pushq %rbx
; CHECK-SSE1-NEXT: movq %rdx, %r13
; CHECK-SSE1-NEXT: movq %rsi, %rbx
; CHECK-SSE1-NEXT: movq %rdi, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-SSE1-NEXT: movb 15(%rdx), %r12b
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb 14(%rdx), %al
; CHECK-SSE1-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 13(%rdx), %al
; CHECK-SSE1-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 12(%rdx), %al
; CHECK-SSE1-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 11(%rdx), %al
; CHECK-SSE1-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 10(%rdx), %al
; CHECK-SSE1-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-SSE1-NEXT: movb 9(%rdx), %r9b
; CHECK-SSE1-NEXT: movb 8(%rdx), %r10b
; CHECK-SSE1-NEXT: movb 7(%rdx), %r11b
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb 6(%rdx), %r8b
; CHECK-SSE1-NEXT: movb 5(%rdx), %bpl
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-SSE1-NEXT: movb 4(%rdx), %sil
; CHECK-SSE1-NEXT: movb 3(%rdx), %dil
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb 2(%rdx), %r14b
; CHECK-SSE1-NEXT: movb (%rdx), %al
; CHECK-SSE1-NEXT: movb 1(%rdx), %r15b
; CHECK-SSE1-NEXT: movb (%rbx), %dl
; CHECK-SSE1-NEXT: xorb %al, %dl
; CHECK-SSE1-NEXT: andb (%rcx), %dl
; CHECK-SSE1-NEXT: xorb %al, %dl
; CHECK-SSE1-NEXT: movb %dl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 1(%rbx), %al
; CHECK-SSE1-NEXT: xorb %r15b, %al
; CHECK-SSE1-NEXT: andb 1(%rcx), %al
; CHECK-SSE1-NEXT: xorb %r15b, %al
; CHECK-SSE1-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 2(%rbx), %al
; CHECK-SSE1-NEXT: xorb %r14b, %al
; CHECK-SSE1-NEXT: andb 2(%rcx), %al
; CHECK-SSE1-NEXT: xorb %r14b, %al
; CHECK-SSE1-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 3(%rbx), %al
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-SSE1-NEXT: xorb %dil, %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: andb 3(%rcx), %al
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-SSE1-NEXT: xorb %dil, %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 4(%rbx), %al
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-SSE1-NEXT: xorb %sil, %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: andb 4(%rcx), %al
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-SSE1-NEXT: xorb %sil, %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 5(%rbx), %al
; CHECK-SSE1-NEXT: xorb %bpl, %al
; CHECK-SSE1-NEXT: andb 5(%rcx), %al
; CHECK-SSE1-NEXT: xorb %bpl, %al
; CHECK-SSE1-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 6(%rbx), %al
; CHECK-SSE1-NEXT: xorb %r8b, %al
; CHECK-SSE1-NEXT: andb 6(%rcx), %al
; CHECK-SSE1-NEXT: xorb %r8b, %al
; CHECK-SSE1-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 7(%rbx), %al
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-SSE1-NEXT: xorb %r11b, %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: andb 7(%rcx), %al
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-SSE1-NEXT: xorb %r11b, %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 8(%rbx), %al
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-SSE1-NEXT: xorb %r10b, %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: andb 8(%rcx), %al
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-SSE1-NEXT: xorb %r10b, %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 9(%rbx), %al
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-SSE1-NEXT: xorb %r9b, %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: andb 9(%rcx), %al
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-SSE1-NEXT: xorb %r9b, %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 10(%rbx), %dl
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-SSE1-NEXT: xorb %al, %dl
; CHECK-SSE1-NEXT: andb 10(%rcx), %dl
; CHECK-SSE1-NEXT: xorb %al, %dl
; CHECK-SSE1-NEXT: movb %dl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 11(%rbx), %dl
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-SSE1-NEXT: xorb %al, %dl
; CHECK-SSE1-NEXT: andb 11(%rcx), %dl
; CHECK-SSE1-NEXT: xorb %al, %dl
; CHECK-SSE1-NEXT: movb %dl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 12(%rbx), %dl
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-SSE1-NEXT: xorb %al, %dl
; CHECK-SSE1-NEXT: andb 12(%rcx), %dl
; CHECK-SSE1-NEXT: xorb %al, %dl
; CHECK-SSE1-NEXT: movb %dl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 13(%rbx), %dl
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-SSE1-NEXT: xorb %al, %dl
; CHECK-SSE1-NEXT: andb 13(%rcx), %dl
; CHECK-SSE1-NEXT: xorb %al, %dl
; CHECK-SSE1-NEXT: movb %dl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 14(%rbx), %dl
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-SSE1-NEXT: xorb %al, %dl
; CHECK-SSE1-NEXT: andb 14(%rcx), %dl
; CHECK-SSE1-NEXT: xorb %al, %dl
; CHECK-SSE1-NEXT: movb %dl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-SSE1-NEXT: movb 15(%rbx), %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: xorb %r12b, %al
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-SSE1-NEXT: andb 15(%rcx), %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: xorb %r12b, %al
; CHECK-SSE1-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-SSE1-NEXT: movb 16(%r13), %al
; CHECK-SSE1-NEXT: movb 16(%rbx), %dl
; CHECK-SSE1-NEXT: xorb %al, %dl
; CHECK-SSE1-NEXT: andb 16(%rcx), %dl
; CHECK-SSE1-NEXT: xorb %al, %dl
; CHECK-SSE1-NEXT: movb %dl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb 17(%r13), %al
; CHECK-SSE1-NEXT: movb 17(%rbx), %dl
; CHECK-SSE1-NEXT: xorb %al, %dl
; CHECK-SSE1-NEXT: andb 17(%rcx), %dl
; CHECK-SSE1-NEXT: xorb %al, %dl
; CHECK-SSE1-NEXT: movb %dl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 18(%r13), %al
; CHECK-SSE1-NEXT: movb 18(%rbx), %dl
; CHECK-SSE1-NEXT: xorb %al, %dl
; CHECK-SSE1-NEXT: andb 18(%rcx), %dl
; CHECK-SSE1-NEXT: xorb %al, %dl
; CHECK-SSE1-NEXT: movb %dl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 19(%r13), %al
; CHECK-SSE1-NEXT: movb 19(%rbx), %r12b
; CHECK-SSE1-NEXT: xorb %al, %r12b
; CHECK-SSE1-NEXT: andb 19(%rcx), %r12b
; CHECK-SSE1-NEXT: xorb %al, %r12b
; CHECK-SSE1-NEXT: movb 20(%r13), %al
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-SSE1-NEXT: movb 20(%rbx), %r15b
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: xorb %al, %r15b
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-SSE1-NEXT: andb 20(%rcx), %r15b
; CHECK-SSE1-NEXT: movq %rcx, %rsi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: xorb %al, %r15b
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-SSE1-NEXT: movb 21(%r13), %al
; CHECK-SSE1-NEXT: movb 21(%rbx), %r14b
; CHECK-SSE1-NEXT: xorb %al, %r14b
; CHECK-SSE1-NEXT: andb 21(%rcx), %r14b
; CHECK-SSE1-NEXT: xorb %al, %r14b
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb 22(%r13), %al
; CHECK-SSE1-NEXT: movb 22(%rbx), %bpl
; CHECK-SSE1-NEXT: xorb %al, %bpl
; CHECK-SSE1-NEXT: andb 22(%rcx), %bpl
; CHECK-SSE1-NEXT: xorb %al, %bpl
; CHECK-SSE1-NEXT: movb 23(%r13), %al
; CHECK-SSE1-NEXT: movb 23(%rbx), %r11b
; CHECK-SSE1-NEXT: xorb %al, %r11b
; CHECK-SSE1-NEXT: andb 23(%rcx), %r11b
; CHECK-SSE1-NEXT: xorb %al, %r11b
; CHECK-SSE1-NEXT: movb 24(%r13), %al
; CHECK-SSE1-NEXT: movb 24(%rbx), %r10b
; CHECK-SSE1-NEXT: xorb %al, %r10b
; CHECK-SSE1-NEXT: andb 24(%rcx), %r10b
; CHECK-SSE1-NEXT: xorb %al, %r10b
; CHECK-SSE1-NEXT: movb 25(%r13), %al
; CHECK-SSE1-NEXT: movb 25(%rbx), %r9b
; CHECK-SSE1-NEXT: xorb %al, %r9b
; CHECK-SSE1-NEXT: andb 25(%rcx), %r9b
; CHECK-SSE1-NEXT: xorb %al, %r9b
; CHECK-SSE1-NEXT: movb 26(%r13), %al
; CHECK-SSE1-NEXT: movb 26(%rbx), %r8b
; CHECK-SSE1-NEXT: xorb %al, %r8b
; CHECK-SSE1-NEXT: andb 26(%rcx), %r8b
; CHECK-SSE1-NEXT: xorb %al, %r8b
; CHECK-SSE1-NEXT: movb 27(%r13), %al
; CHECK-SSE1-NEXT: movb 27(%rbx), %dil
; CHECK-SSE1-NEXT: xorb %al, %dil
; CHECK-SSE1-NEXT: andb 27(%rcx), %dil
; CHECK-SSE1-NEXT: xorb %al, %dil
; CHECK-SSE1-NEXT: movb 28(%r13), %al
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-SSE1-NEXT: movb 28(%rbx), %dl
; CHECK-SSE1-NEXT: xorb %al, %dl
; CHECK-SSE1-NEXT: andb 28(%rcx), %dl
; CHECK-SSE1-NEXT: xorb %al, %dl
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb 29(%r13), %al
; CHECK-SSE1-NEXT: movb 29(%rbx), %cl
; CHECK-SSE1-NEXT: xorb %al, %cl
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-SSE1-NEXT: andb 29(%rsi), %cl
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: xorb %al, %cl
; CHECK-SSE1-NEXT: movb 30(%r13), %al
; CHECK-SSE1-NEXT: movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
; CHECK-SSE1-NEXT: movb 30(%rbx), %al
; CHECK-SSE1-NEXT: xorb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Folded Reload
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-SSE1-NEXT: andb 30(%rsi), %al
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: xorb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Folded Reload
; CHECK-SSE1-NEXT: movb 31(%r13), %r13b
; CHECK-SSE1-NEXT: movb 31(%rbx), %bl
; CHECK-SSE1-NEXT: xorb %r13b, %bl
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-SSE1-NEXT: andb 31(%rsi), %bl
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: xorb %r13b, %bl
; CHECK-SSE1-NEXT: movq {{[-0-9]+}}(%r{{[sb]}}p), %r13 # 8-byte Reload
; CHECK-SSE1-NEXT: movb %bl, 31(%r13)
; CHECK-SSE1-NEXT: movb %al, 30(%r13)
; CHECK-SSE1-NEXT: movb %cl, 29(%r13)
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-SSE1-NEXT: movb %dl, 28(%r13)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb %dil, 27(%r13)
; CHECK-SSE1-NEXT: movb %r8b, 26(%r13)
; CHECK-SSE1-NEXT: movb %r9b, 25(%r13)
; CHECK-SSE1-NEXT: movb %r10b, 24(%r13)
; CHECK-SSE1-NEXT: movb %r11b, 23(%r13)
; CHECK-SSE1-NEXT: movb %bpl, 22(%r13)
[X86] Fix register resizings for inline assembly register operands. When replacing a named register input to the appropriately sized sub/super-register. In the case of a 64-bit value being assigned to a register in 32-bit mode, match GCC's assignment. Reviewers: eli.friedman, craig.topper Subscribers: nickdesaulniers, llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D51502 llvm-svn: 342175
2018-09-14 04:33:56 +08:00
; CHECK-SSE1-NEXT: movb %r14b, 21(%r13)
; CHECK-SSE1-NEXT: movb %r15b, 20(%r13)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movb %r12b, 19(%r13)
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-SSE1-NEXT: movb %al, 18(%r13)
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-SSE1-NEXT: movb %al, 17(%r13)
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-SSE1-NEXT: movb %al, 16(%r13)
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-SSE1-NEXT: movb %al, 15(%r13)
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-SSE1-NEXT: movb %al, 14(%r13)
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-SSE1-NEXT: movb %al, 13(%r13)
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-SSE1-NEXT: movb %al, 12(%r13)
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-SSE1-NEXT: movb %al, 11(%r13)
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-SSE1-NEXT: movb %al, 10(%r13)
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-SSE1-NEXT: movb %al, 9(%r13)
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-SSE1-NEXT: movb %al, 8(%r13)
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-SSE1-NEXT: movb %al, 7(%r13)
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-SSE1-NEXT: movb %al, 6(%r13)
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-SSE1-NEXT: movb %al, 5(%r13)
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-SSE1-NEXT: movb %al, 4(%r13)
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-SSE1-NEXT: movb %al, 3(%r13)
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-SSE1-NEXT: movb %al, 2(%r13)
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-SSE1-NEXT: movb %al, 1(%r13)
; CHECK-SSE1-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
; CHECK-SSE1-NEXT: movb %al, (%r13)
; CHECK-SSE1-NEXT: movq %r13, %rax
; CHECK-SSE1-NEXT: popq %rbx
; CHECK-SSE1-NEXT: popq %r12
; CHECK-SSE1-NEXT: popq %r13
; CHECK-SSE1-NEXT: popq %r14
; CHECK-SSE1-NEXT: popq %r15
; CHECK-SSE1-NEXT: popq %rbp
; CHECK-SSE1-NEXT: retq
;
; CHECK-SSE2-LABEL: in_v32i8:
; CHECK-SSE2: # %bb.0:
[DAGCombine][X86][AArch64] Masked merge unfolding: vector edition. Summary: This **appears** to be the last missing piece for the masked merge pattern handling in the backend. This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`), and we need to make sure that they are generated. Differential Revision: https://reviews.llvm.org/D46528 llvm-svn: 332904
2018-05-22 05:41:02 +08:00
; CHECK-SSE2-NEXT: movaps (%rdx), %xmm0
; CHECK-SSE2-NEXT: movaps 16(%rdx), %xmm1
; CHECK-SSE2-NEXT: movaps %xmm0, %xmm2
; CHECK-SSE2-NEXT: andnps (%rsi), %xmm2
; CHECK-SSE2-NEXT: andps (%rdi), %xmm0
; CHECK-SSE2-NEXT: orps %xmm2, %xmm0
; CHECK-SSE2-NEXT: movaps %xmm1, %xmm2
; CHECK-SSE2-NEXT: andnps 16(%rsi), %xmm2
; CHECK-SSE2-NEXT: andps 16(%rdi), %xmm1
; CHECK-SSE2-NEXT: orps %xmm2, %xmm1
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE2-NEXT: retq
;
; CHECK-XOP-LABEL: in_v32i8:
; CHECK-XOP: # %bb.0:
[DAGCombine][X86][AArch64] Masked merge unfolding: vector edition. Summary: This **appears** to be the last missing piece for the masked merge pattern handling in the backend. This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`), and we need to make sure that they are generated. Differential Revision: https://reviews.llvm.org/D46528 llvm-svn: 332904
2018-05-22 05:41:02 +08:00
; CHECK-XOP-NEXT: vmovdqa (%rdi), %ymm0
; CHECK-XOP-NEXT: vmovdqa (%rdx), %ymm1
; CHECK-XOP-NEXT: vpcmov %ymm1, (%rsi), %ymm0, %ymm0
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-XOP-NEXT: retq
%x = load <32 x i8>, <32 x i8> *%px, align 32
%y = load <32 x i8>, <32 x i8> *%py, align 32
%mask = load <32 x i8>, <32 x i8> *%pmask, align 32
%n0 = xor <32 x i8> %x, %y
%n1 = and <32 x i8> %n0, %mask
%r = xor <32 x i8> %n1, %y
ret <32 x i8> %r
}
define <16 x i16> @in_v16i16(<16 x i16> *%px, <16 x i16> *%py, <16 x i16> *%pmask) nounwind {
; CHECK-BASELINE-LABEL: in_v16i16:
; CHECK-BASELINE: # %bb.0:
; CHECK-BASELINE-NEXT: pushq %rbp
; CHECK-BASELINE-NEXT: pushq %r15
; CHECK-BASELINE-NEXT: pushq %r14
; CHECK-BASELINE-NEXT: pushq %r13
; CHECK-BASELINE-NEXT: pushq %r12
; CHECK-BASELINE-NEXT: pushq %rbx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movq %rcx, %r9
; CHECK-BASELINE-NEXT: movq %rdi, %r10
; CHECK-BASELINE-NEXT: movzwl 30(%rdx), %edi
; CHECK-BASELINE-NEXT: movl %edi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-BASELINE-NEXT: movl 28(%rdx), %edi
; CHECK-BASELINE-NEXT: movl %edi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-BASELINE-NEXT: movzwl 26(%rdx), %edi
; CHECK-BASELINE-NEXT: movl %edi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl 24(%rdx), %eax
; CHECK-BASELINE-NEXT: movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-BASELINE-NEXT: movzwl 22(%rdx), %eax
; CHECK-BASELINE-NEXT: movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movl 20(%rdx), %r11d
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl %r11d, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movzwl 18(%rdx), %r14d
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl %r14d, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movl 16(%rdx), %r15d
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl %r15d, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movzwl 14(%rdx), %r12d
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl %r12d, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movl 12(%rdx), %r13d
; CHECK-BASELINE-NEXT: movl %r13d, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-BASELINE-NEXT: movzwl 10(%rdx), %r8d
; CHECK-BASELINE-NEXT: movl %r8d, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-BASELINE-NEXT: movl 8(%rdx), %ebx
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl %ebx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movzwl 6(%rdx), %ebp
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl %ebp, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movl (%rdx), %ecx
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl %ecx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movl 4(%rdx), %edi
; CHECK-BASELINE-NEXT: movl %edi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-BASELINE-NEXT: movzwl 2(%rdx), %eax
; CHECK-BASELINE-NEXT: movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movzwl (%rsi), %edx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: xorw %cx, %dx
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl %edx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movzwl 2(%rsi), %ecx
; CHECK-BASELINE-NEXT: xorw %ax, %cx
; CHECK-BASELINE-NEXT: movl %ecx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movzwl 4(%rsi), %eax
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: xorw %di, %ax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-BASELINE-NEXT: movzwl 6(%rsi), %edx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: xorw %bp, %dx
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl %edx, %eax
; CHECK-BASELINE-NEXT: movzwl 8(%rsi), %ecx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: xorw %bx, %cx
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movzwl 10(%rsi), %edx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: xorw %r8w, %dx
; CHECK-BASELINE-NEXT: movl %edx, %r8d
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movzwl 12(%rsi), %edx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: xorw %r13w, %dx
; CHECK-BASELINE-NEXT: movzwl 14(%rsi), %r13d
; CHECK-BASELINE-NEXT: xorw %r12w, %r13w
; CHECK-BASELINE-NEXT: movzwl 16(%rsi), %r12d
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: xorw %r15w, %r12w
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movzwl 18(%rsi), %r15d
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: xorw %r14w, %r15w
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movzwl 20(%rsi), %r14d
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: xorw %r11w, %r14w
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movzwl 22(%rsi), %ebp
; CHECK-BASELINE-NEXT: xorw {{[-0-9]+}}(%r{{[sb]}}p), %bp # 2-byte Folded Reload
; CHECK-BASELINE-NEXT: movzwl 24(%rsi), %ebx
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: xorw {{[-0-9]+}}(%r{{[sb]}}p), %bx # 2-byte Folded Reload
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movzwl 26(%rsi), %r11d
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: xorw {{[-0-9]+}}(%r{{[sb]}}p), %r11w # 2-byte Folded Reload
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movzwl 28(%rsi), %edi
; CHECK-BASELINE-NEXT: xorw {{[-0-9]+}}(%r{{[sb]}}p), %di # 2-byte Folded Reload
; CHECK-BASELINE-NEXT: movzwl 30(%rsi), %esi
; CHECK-BASELINE-NEXT: xorw {{[-0-9]+}}(%r{{[sb]}}p), %si # 2-byte Folded Reload
; CHECK-BASELINE-NEXT: andw 30(%r9), %si
; CHECK-BASELINE-NEXT: andw 28(%r9), %di
; CHECK-BASELINE-NEXT: andw 26(%r9), %r11w
; CHECK-BASELINE-NEXT: andw 24(%r9), %bx
; CHECK-BASELINE-NEXT: andw 22(%r9), %bp
; CHECK-BASELINE-NEXT: andw 20(%r9), %r14w
; CHECK-BASELINE-NEXT: andw 18(%r9), %r15w
; CHECK-BASELINE-NEXT: andw 16(%r9), %r12w
; CHECK-BASELINE-NEXT: andw 14(%r9), %r13w
; CHECK-BASELINE-NEXT: andw 12(%r9), %dx
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl %edx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: andw 10(%r9), %r8w
; CHECK-BASELINE-NEXT: movl %r8d, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-BASELINE-NEXT: movl %ecx, %edx
; CHECK-BASELINE-NEXT: andw 8(%r9), %dx
; CHECK-BASELINE-NEXT: andw 6(%r9), %ax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %r8d # 4-byte Reload
; CHECK-BASELINE-NEXT: andw 4(%r9), %r8w
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 4-byte Reload
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: andw 2(%r9), %ax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %ecx # 4-byte Reload
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: andw (%r9), %cx
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %ecx # 4-byte Folded Reload
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movl %ecx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-BASELINE-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 4-byte Folded Reload
; CHECK-BASELINE-NEXT: movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-BASELINE-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %r8d # 4-byte Folded Reload
; CHECK-BASELINE-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %r9d # 4-byte Reload
; CHECK-BASELINE-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %r9d # 4-byte Folded Reload
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %edx # 4-byte Folded Reload
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movl %edx, %ecx
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %edx # 4-byte Reload
; CHECK-BASELINE-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %edx # 4-byte Folded Reload
; CHECK-BASELINE-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 4-byte Reload
; CHECK-BASELINE-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 4-byte Folded Reload
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %r13d # 4-byte Folded Reload
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %r12d # 4-byte Folded Reload
; CHECK-BASELINE-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %r15d # 4-byte Folded Reload
; CHECK-BASELINE-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %r14d # 4-byte Folded Reload
; CHECK-BASELINE-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %ebp # 4-byte Folded Reload
; CHECK-BASELINE-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %ebx # 4-byte Folded Reload
; CHECK-BASELINE-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %r11d # 4-byte Folded Reload
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Folded Reload
; CHECK-BASELINE-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %esi # 4-byte Folded Reload
; CHECK-BASELINE-NEXT: movw %si, 30(%r10)
; CHECK-BASELINE-NEXT: movw %di, 28(%r10)
; CHECK-BASELINE-NEXT: movw %r11w, 26(%r10)
; CHECK-BASELINE-NEXT: movw %bx, 24(%r10)
; CHECK-BASELINE-NEXT: movw %bp, 22(%r10)
; CHECK-BASELINE-NEXT: movw %r14w, 20(%r10)
; CHECK-BASELINE-NEXT: movw %r15w, 18(%r10)
; CHECK-BASELINE-NEXT: movw %r12w, 16(%r10)
; CHECK-BASELINE-NEXT: movw %r13w, 14(%r10)
; CHECK-BASELINE-NEXT: movw %ax, 12(%r10)
; CHECK-BASELINE-NEXT: movw %dx, 10(%r10)
; CHECK-BASELINE-NEXT: movw %cx, 8(%r10)
; CHECK-BASELINE-NEXT: movw %r9w, 6(%r10)
; CHECK-BASELINE-NEXT: movw %r8w, 4(%r10)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 4-byte Reload
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movw %ax, 2(%r10)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 4-byte Reload
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movw %ax, (%r10)
; CHECK-BASELINE-NEXT: movq %r10, %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: popq %rbx
; CHECK-BASELINE-NEXT: popq %r12
; CHECK-BASELINE-NEXT: popq %r13
; CHECK-BASELINE-NEXT: popq %r14
; CHECK-BASELINE-NEXT: popq %r15
; CHECK-BASELINE-NEXT: popq %rbp
; CHECK-BASELINE-NEXT: retq
;
; CHECK-SSE1-LABEL: in_v16i16:
; CHECK-SSE1: # %bb.0:
; CHECK-SSE1-NEXT: pushq %rbp
; CHECK-SSE1-NEXT: pushq %r15
; CHECK-SSE1-NEXT: pushq %r14
; CHECK-SSE1-NEXT: pushq %r13
; CHECK-SSE1-NEXT: pushq %r12
; CHECK-SSE1-NEXT: pushq %rbx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movq %rcx, %r9
; CHECK-SSE1-NEXT: movq %rdi, %r10
; CHECK-SSE1-NEXT: movzwl 30(%rdx), %edi
; CHECK-SSE1-NEXT: movl %edi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-SSE1-NEXT: movl 28(%rdx), %edi
; CHECK-SSE1-NEXT: movl %edi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-SSE1-NEXT: movzwl 26(%rdx), %edi
; CHECK-SSE1-NEXT: movl %edi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movl 24(%rdx), %eax
; CHECK-SSE1-NEXT: movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-SSE1-NEXT: movzwl 22(%rdx), %eax
; CHECK-SSE1-NEXT: movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movl 20(%rdx), %r11d
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movl %r11d, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movzwl 18(%rdx), %r14d
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movl %r14d, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movl 16(%rdx), %r15d
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movl %r15d, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movzwl 14(%rdx), %r12d
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movl %r12d, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movl 12(%rdx), %r13d
; CHECK-SSE1-NEXT: movl %r13d, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-SSE1-NEXT: movzwl 10(%rdx), %r8d
; CHECK-SSE1-NEXT: movl %r8d, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-SSE1-NEXT: movl 8(%rdx), %ebx
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movl %ebx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movzwl 6(%rdx), %ebp
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movl %ebp, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movl (%rdx), %ecx
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movl %ecx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movl 4(%rdx), %edi
; CHECK-SSE1-NEXT: movl %edi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-SSE1-NEXT: movzwl 2(%rdx), %eax
; CHECK-SSE1-NEXT: movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movzwl (%rsi), %edx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: xorw %cx, %dx
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movl %edx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movzwl 2(%rsi), %ecx
; CHECK-SSE1-NEXT: xorw %ax, %cx
; CHECK-SSE1-NEXT: movl %ecx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movzwl 4(%rsi), %eax
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: xorw %di, %ax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-SSE1-NEXT: movzwl 6(%rsi), %edx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: xorw %bp, %dx
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movl %edx, %eax
; CHECK-SSE1-NEXT: movzwl 8(%rsi), %ecx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: xorw %bx, %cx
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movzwl 10(%rsi), %edx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: xorw %r8w, %dx
; CHECK-SSE1-NEXT: movl %edx, %r8d
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movzwl 12(%rsi), %edx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: xorw %r13w, %dx
; CHECK-SSE1-NEXT: movzwl 14(%rsi), %r13d
; CHECK-SSE1-NEXT: xorw %r12w, %r13w
; CHECK-SSE1-NEXT: movzwl 16(%rsi), %r12d
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: xorw %r15w, %r12w
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movzwl 18(%rsi), %r15d
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: xorw %r14w, %r15w
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movzwl 20(%rsi), %r14d
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: xorw %r11w, %r14w
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movzwl 22(%rsi), %ebp
; CHECK-SSE1-NEXT: xorw {{[-0-9]+}}(%r{{[sb]}}p), %bp # 2-byte Folded Reload
; CHECK-SSE1-NEXT: movzwl 24(%rsi), %ebx
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: xorw {{[-0-9]+}}(%r{{[sb]}}p), %bx # 2-byte Folded Reload
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movzwl 26(%rsi), %r11d
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: xorw {{[-0-9]+}}(%r{{[sb]}}p), %r11w # 2-byte Folded Reload
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movzwl 28(%rsi), %edi
; CHECK-SSE1-NEXT: xorw {{[-0-9]+}}(%r{{[sb]}}p), %di # 2-byte Folded Reload
; CHECK-SSE1-NEXT: movzwl 30(%rsi), %esi
; CHECK-SSE1-NEXT: xorw {{[-0-9]+}}(%r{{[sb]}}p), %si # 2-byte Folded Reload
; CHECK-SSE1-NEXT: andw 30(%r9), %si
; CHECK-SSE1-NEXT: andw 28(%r9), %di
; CHECK-SSE1-NEXT: andw 26(%r9), %r11w
; CHECK-SSE1-NEXT: andw 24(%r9), %bx
; CHECK-SSE1-NEXT: andw 22(%r9), %bp
; CHECK-SSE1-NEXT: andw 20(%r9), %r14w
; CHECK-SSE1-NEXT: andw 18(%r9), %r15w
; CHECK-SSE1-NEXT: andw 16(%r9), %r12w
; CHECK-SSE1-NEXT: andw 14(%r9), %r13w
; CHECK-SSE1-NEXT: andw 12(%r9), %dx
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movl %edx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: andw 10(%r9), %r8w
; CHECK-SSE1-NEXT: movl %r8d, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-SSE1-NEXT: movl %ecx, %edx
; CHECK-SSE1-NEXT: andw 8(%r9), %dx
; CHECK-SSE1-NEXT: andw 6(%r9), %ax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %r8d # 4-byte Reload
; CHECK-SSE1-NEXT: andw 4(%r9), %r8w
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 4-byte Reload
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: andw 2(%r9), %ax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %ecx # 4-byte Reload
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: andw (%r9), %cx
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %ecx # 4-byte Folded Reload
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movl %ecx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-SSE1-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 4-byte Folded Reload
; CHECK-SSE1-NEXT: movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-SSE1-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %r8d # 4-byte Folded Reload
; CHECK-SSE1-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %r9d # 4-byte Reload
; CHECK-SSE1-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %r9d # 4-byte Folded Reload
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %edx # 4-byte Folded Reload
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movl %edx, %ecx
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %edx # 4-byte Reload
; CHECK-SSE1-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %edx # 4-byte Folded Reload
; CHECK-SSE1-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 4-byte Reload
; CHECK-SSE1-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 4-byte Folded Reload
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %r13d # 4-byte Folded Reload
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %r12d # 4-byte Folded Reload
; CHECK-SSE1-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %r15d # 4-byte Folded Reload
; CHECK-SSE1-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %r14d # 4-byte Folded Reload
; CHECK-SSE1-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %ebp # 4-byte Folded Reload
; CHECK-SSE1-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %ebx # 4-byte Folded Reload
; CHECK-SSE1-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %r11d # 4-byte Folded Reload
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %edi # 4-byte Folded Reload
; CHECK-SSE1-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %esi # 4-byte Folded Reload
; CHECK-SSE1-NEXT: movw %si, 30(%r10)
; CHECK-SSE1-NEXT: movw %di, 28(%r10)
; CHECK-SSE1-NEXT: movw %r11w, 26(%r10)
; CHECK-SSE1-NEXT: movw %bx, 24(%r10)
; CHECK-SSE1-NEXT: movw %bp, 22(%r10)
; CHECK-SSE1-NEXT: movw %r14w, 20(%r10)
; CHECK-SSE1-NEXT: movw %r15w, 18(%r10)
; CHECK-SSE1-NEXT: movw %r12w, 16(%r10)
; CHECK-SSE1-NEXT: movw %r13w, 14(%r10)
; CHECK-SSE1-NEXT: movw %ax, 12(%r10)
; CHECK-SSE1-NEXT: movw %dx, 10(%r10)
; CHECK-SSE1-NEXT: movw %cx, 8(%r10)
; CHECK-SSE1-NEXT: movw %r9w, 6(%r10)
; CHECK-SSE1-NEXT: movw %r8w, 4(%r10)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 4-byte Reload
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movw %ax, 2(%r10)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movl {{[-0-9]+}}(%r{{[sb]}}p), %eax # 4-byte Reload
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movw %ax, (%r10)
; CHECK-SSE1-NEXT: movq %r10, %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: popq %rbx
; CHECK-SSE1-NEXT: popq %r12
; CHECK-SSE1-NEXT: popq %r13
; CHECK-SSE1-NEXT: popq %r14
; CHECK-SSE1-NEXT: popq %r15
; CHECK-SSE1-NEXT: popq %rbp
; CHECK-SSE1-NEXT: retq
;
; CHECK-SSE2-LABEL: in_v16i16:
; CHECK-SSE2: # %bb.0:
[DAGCombine][X86][AArch64] Masked merge unfolding: vector edition. Summary: This **appears** to be the last missing piece for the masked merge pattern handling in the backend. This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`), and we need to make sure that they are generated. Differential Revision: https://reviews.llvm.org/D46528 llvm-svn: 332904
2018-05-22 05:41:02 +08:00
; CHECK-SSE2-NEXT: movaps (%rdx), %xmm0
; CHECK-SSE2-NEXT: movaps 16(%rdx), %xmm1
; CHECK-SSE2-NEXT: movaps %xmm0, %xmm2
; CHECK-SSE2-NEXT: andnps (%rsi), %xmm2
; CHECK-SSE2-NEXT: andps (%rdi), %xmm0
; CHECK-SSE2-NEXT: orps %xmm2, %xmm0
; CHECK-SSE2-NEXT: movaps %xmm1, %xmm2
; CHECK-SSE2-NEXT: andnps 16(%rsi), %xmm2
; CHECK-SSE2-NEXT: andps 16(%rdi), %xmm1
; CHECK-SSE2-NEXT: orps %xmm2, %xmm1
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE2-NEXT: retq
;
; CHECK-XOP-LABEL: in_v16i16:
; CHECK-XOP: # %bb.0:
[DAGCombine][X86][AArch64] Masked merge unfolding: vector edition. Summary: This **appears** to be the last missing piece for the masked merge pattern handling in the backend. This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`), and we need to make sure that they are generated. Differential Revision: https://reviews.llvm.org/D46528 llvm-svn: 332904
2018-05-22 05:41:02 +08:00
; CHECK-XOP-NEXT: vmovdqa (%rdi), %ymm0
; CHECK-XOP-NEXT: vmovdqa (%rdx), %ymm1
; CHECK-XOP-NEXT: vpcmov %ymm1, (%rsi), %ymm0, %ymm0
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-XOP-NEXT: retq
%x = load <16 x i16>, <16 x i16> *%px, align 32
%y = load <16 x i16>, <16 x i16> *%py, align 32
%mask = load <16 x i16>, <16 x i16> *%pmask, align 32
%n0 = xor <16 x i16> %x, %y
%n1 = and <16 x i16> %n0, %mask
%r = xor <16 x i16> %n1, %y
ret <16 x i16> %r
}
define <8 x i32> @in_v8i32(<8 x i32> *%px, <8 x i32> *%py, <8 x i32> *%pmask) nounwind {
; CHECK-BASELINE-LABEL: in_v8i32:
; CHECK-BASELINE: # %bb.0:
; CHECK-BASELINE-NEXT: pushq %rbp
; CHECK-BASELINE-NEXT: pushq %r15
; CHECK-BASELINE-NEXT: pushq %r14
; CHECK-BASELINE-NEXT: pushq %r13
; CHECK-BASELINE-NEXT: pushq %r12
; CHECK-BASELINE-NEXT: pushq %rbx
; CHECK-BASELINE-NEXT: movl 28(%rdx), %r15d
; CHECK-BASELINE-NEXT: movl 24(%rdx), %r14d
; CHECK-BASELINE-NEXT: movl 20(%rdx), %r10d
; CHECK-BASELINE-NEXT: movl %r10d, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-BASELINE-NEXT: movl 16(%rdx), %eax
; CHECK-BASELINE-NEXT: movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-BASELINE-NEXT: movl 12(%rdx), %ebp
; CHECK-BASELINE-NEXT: movl %ebp, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-BASELINE-NEXT: movl 8(%rdx), %ebx
; CHECK-BASELINE-NEXT: movl %ebx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-BASELINE-NEXT: movl (%rdx), %r12d
; CHECK-BASELINE-NEXT: movl 4(%rdx), %r13d
; CHECK-BASELINE-NEXT: movl (%rsi), %r11d
; CHECK-BASELINE-NEXT: xorl %r12d, %r11d
; CHECK-BASELINE-NEXT: movl 4(%rsi), %r9d
; CHECK-BASELINE-NEXT: xorl %r13d, %r9d
; CHECK-BASELINE-NEXT: movl 8(%rsi), %r8d
; CHECK-BASELINE-NEXT: xorl %ebx, %r8d
; CHECK-BASELINE-NEXT: movl 12(%rsi), %ebx
; CHECK-BASELINE-NEXT: xorl %ebp, %ebx
; CHECK-BASELINE-NEXT: movl 16(%rsi), %ebp
; CHECK-BASELINE-NEXT: xorl %eax, %ebp
; CHECK-BASELINE-NEXT: movl 20(%rsi), %edx
; CHECK-BASELINE-NEXT: xorl %r10d, %edx
; CHECK-BASELINE-NEXT: movl 24(%rsi), %eax
; CHECK-BASELINE-NEXT: xorl %r14d, %eax
; CHECK-BASELINE-NEXT: movl 28(%rsi), %esi
; CHECK-BASELINE-NEXT: xorl %r15d, %esi
; CHECK-BASELINE-NEXT: andl 28(%rcx), %esi
; CHECK-BASELINE-NEXT: andl 24(%rcx), %eax
; CHECK-BASELINE-NEXT: andl 20(%rcx), %edx
; CHECK-BASELINE-NEXT: andl 16(%rcx), %ebp
; CHECK-BASELINE-NEXT: andl 12(%rcx), %ebx
; CHECK-BASELINE-NEXT: andl 8(%rcx), %r8d
; CHECK-BASELINE-NEXT: andl 4(%rcx), %r9d
; CHECK-BASELINE-NEXT: andl (%rcx), %r11d
; CHECK-BASELINE-NEXT: xorl %r12d, %r11d
; CHECK-BASELINE-NEXT: xorl %r13d, %r9d
; CHECK-BASELINE-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %r8d # 4-byte Folded Reload
; CHECK-BASELINE-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %ebx # 4-byte Folded Reload
; CHECK-BASELINE-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %ebp # 4-byte Folded Reload
; CHECK-BASELINE-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %edx # 4-byte Folded Reload
; CHECK-BASELINE-NEXT: xorl %r14d, %eax
; CHECK-BASELINE-NEXT: xorl %r15d, %esi
; CHECK-BASELINE-NEXT: movl %esi, 28(%rdi)
; CHECK-BASELINE-NEXT: movl %eax, 24(%rdi)
; CHECK-BASELINE-NEXT: movl %edx, 20(%rdi)
; CHECK-BASELINE-NEXT: movl %ebp, 16(%rdi)
; CHECK-BASELINE-NEXT: movl %ebx, 12(%rdi)
; CHECK-BASELINE-NEXT: movl %r8d, 8(%rdi)
; CHECK-BASELINE-NEXT: movl %r9d, 4(%rdi)
; CHECK-BASELINE-NEXT: movl %r11d, (%rdi)
; CHECK-BASELINE-NEXT: movq %rdi, %rax
; CHECK-BASELINE-NEXT: popq %rbx
; CHECK-BASELINE-NEXT: popq %r12
; CHECK-BASELINE-NEXT: popq %r13
; CHECK-BASELINE-NEXT: popq %r14
; CHECK-BASELINE-NEXT: popq %r15
; CHECK-BASELINE-NEXT: popq %rbp
; CHECK-BASELINE-NEXT: retq
;
; CHECK-SSE1-LABEL: in_v8i32:
; CHECK-SSE1: # %bb.0:
; CHECK-SSE1-NEXT: pushq %rbp
; CHECK-SSE1-NEXT: pushq %r15
; CHECK-SSE1-NEXT: pushq %r14
; CHECK-SSE1-NEXT: pushq %r13
; CHECK-SSE1-NEXT: pushq %r12
; CHECK-SSE1-NEXT: pushq %rbx
; CHECK-SSE1-NEXT: movl 28(%rdx), %r15d
; CHECK-SSE1-NEXT: movl 24(%rdx), %r14d
; CHECK-SSE1-NEXT: movl 20(%rdx), %r10d
; CHECK-SSE1-NEXT: movl %r10d, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-SSE1-NEXT: movl 16(%rdx), %eax
; CHECK-SSE1-NEXT: movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-SSE1-NEXT: movl 12(%rdx), %ebp
; CHECK-SSE1-NEXT: movl %ebp, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-SSE1-NEXT: movl 8(%rdx), %ebx
; CHECK-SSE1-NEXT: movl %ebx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
; CHECK-SSE1-NEXT: movl (%rdx), %r12d
; CHECK-SSE1-NEXT: movl 4(%rdx), %r13d
; CHECK-SSE1-NEXT: movl (%rsi), %r11d
; CHECK-SSE1-NEXT: xorl %r12d, %r11d
; CHECK-SSE1-NEXT: movl 4(%rsi), %r9d
; CHECK-SSE1-NEXT: xorl %r13d, %r9d
; CHECK-SSE1-NEXT: movl 8(%rsi), %r8d
; CHECK-SSE1-NEXT: xorl %ebx, %r8d
; CHECK-SSE1-NEXT: movl 12(%rsi), %ebx
; CHECK-SSE1-NEXT: xorl %ebp, %ebx
; CHECK-SSE1-NEXT: movl 16(%rsi), %ebp
; CHECK-SSE1-NEXT: xorl %eax, %ebp
; CHECK-SSE1-NEXT: movl 20(%rsi), %edx
; CHECK-SSE1-NEXT: xorl %r10d, %edx
; CHECK-SSE1-NEXT: movl 24(%rsi), %eax
; CHECK-SSE1-NEXT: xorl %r14d, %eax
; CHECK-SSE1-NEXT: movl 28(%rsi), %esi
; CHECK-SSE1-NEXT: xorl %r15d, %esi
; CHECK-SSE1-NEXT: andl 28(%rcx), %esi
; CHECK-SSE1-NEXT: andl 24(%rcx), %eax
; CHECK-SSE1-NEXT: andl 20(%rcx), %edx
; CHECK-SSE1-NEXT: andl 16(%rcx), %ebp
; CHECK-SSE1-NEXT: andl 12(%rcx), %ebx
; CHECK-SSE1-NEXT: andl 8(%rcx), %r8d
; CHECK-SSE1-NEXT: andl 4(%rcx), %r9d
; CHECK-SSE1-NEXT: andl (%rcx), %r11d
; CHECK-SSE1-NEXT: xorl %r12d, %r11d
; CHECK-SSE1-NEXT: xorl %r13d, %r9d
; CHECK-SSE1-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %r8d # 4-byte Folded Reload
; CHECK-SSE1-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %ebx # 4-byte Folded Reload
; CHECK-SSE1-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %ebp # 4-byte Folded Reload
; CHECK-SSE1-NEXT: xorl {{[-0-9]+}}(%r{{[sb]}}p), %edx # 4-byte Folded Reload
; CHECK-SSE1-NEXT: xorl %r14d, %eax
; CHECK-SSE1-NEXT: xorl %r15d, %esi
; CHECK-SSE1-NEXT: movl %esi, 28(%rdi)
; CHECK-SSE1-NEXT: movl %eax, 24(%rdi)
; CHECK-SSE1-NEXT: movl %edx, 20(%rdi)
; CHECK-SSE1-NEXT: movl %ebp, 16(%rdi)
; CHECK-SSE1-NEXT: movl %ebx, 12(%rdi)
; CHECK-SSE1-NEXT: movl %r8d, 8(%rdi)
; CHECK-SSE1-NEXT: movl %r9d, 4(%rdi)
; CHECK-SSE1-NEXT: movl %r11d, (%rdi)
; CHECK-SSE1-NEXT: movq %rdi, %rax
; CHECK-SSE1-NEXT: popq %rbx
; CHECK-SSE1-NEXT: popq %r12
; CHECK-SSE1-NEXT: popq %r13
; CHECK-SSE1-NEXT: popq %r14
; CHECK-SSE1-NEXT: popq %r15
; CHECK-SSE1-NEXT: popq %rbp
; CHECK-SSE1-NEXT: retq
;
; CHECK-SSE2-LABEL: in_v8i32:
; CHECK-SSE2: # %bb.0:
[DAGCombine][X86][AArch64] Masked merge unfolding: vector edition. Summary: This **appears** to be the last missing piece for the masked merge pattern handling in the backend. This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`), and we need to make sure that they are generated. Differential Revision: https://reviews.llvm.org/D46528 llvm-svn: 332904
2018-05-22 05:41:02 +08:00
; CHECK-SSE2-NEXT: movaps (%rdx), %xmm0
; CHECK-SSE2-NEXT: movaps 16(%rdx), %xmm1
; CHECK-SSE2-NEXT: movaps %xmm0, %xmm2
; CHECK-SSE2-NEXT: andnps (%rsi), %xmm2
; CHECK-SSE2-NEXT: andps (%rdi), %xmm0
; CHECK-SSE2-NEXT: orps %xmm2, %xmm0
; CHECK-SSE2-NEXT: movaps %xmm1, %xmm2
; CHECK-SSE2-NEXT: andnps 16(%rsi), %xmm2
; CHECK-SSE2-NEXT: andps 16(%rdi), %xmm1
; CHECK-SSE2-NEXT: orps %xmm2, %xmm1
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE2-NEXT: retq
;
; CHECK-XOP-LABEL: in_v8i32:
; CHECK-XOP: # %bb.0:
[DAGCombine][X86][AArch64] Masked merge unfolding: vector edition. Summary: This **appears** to be the last missing piece for the masked merge pattern handling in the backend. This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`), and we need to make sure that they are generated. Differential Revision: https://reviews.llvm.org/D46528 llvm-svn: 332904
2018-05-22 05:41:02 +08:00
; CHECK-XOP-NEXT: vmovdqa (%rdi), %ymm0
; CHECK-XOP-NEXT: vmovdqa (%rdx), %ymm1
; CHECK-XOP-NEXT: vpcmov %ymm1, (%rsi), %ymm0, %ymm0
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-XOP-NEXT: retq
%x = load <8 x i32>, <8 x i32> *%px, align 32
%y = load <8 x i32>, <8 x i32> *%py, align 32
%mask = load <8 x i32>, <8 x i32> *%pmask, align 32
%n0 = xor <8 x i32> %x, %y
%n1 = and <8 x i32> %n0, %mask
%r = xor <8 x i32> %n1, %y
ret <8 x i32> %r
}
define <4 x i64> @in_v4i64(<4 x i64> *%px, <4 x i64> *%py, <4 x i64> *%pmask) nounwind {
; CHECK-BASELINE-LABEL: in_v4i64:
; CHECK-BASELINE: # %bb.0:
; CHECK-BASELINE-NEXT: pushq %rbx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movq %rdi, %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movq 24(%rdx), %r8
; CHECK-BASELINE-NEXT: movq 16(%rdx), %r9
; CHECK-BASELINE-NEXT: movq (%rdx), %r11
; CHECK-BASELINE-NEXT: movq 8(%rdx), %r10
; CHECK-BASELINE-NEXT: movq (%rsi), %rdx
; CHECK-BASELINE-NEXT: xorq %r11, %rdx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movq 8(%rsi), %rdi
; CHECK-BASELINE-NEXT: xorq %r10, %rdi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: movq 16(%rsi), %rbx
; CHECK-BASELINE-NEXT: xorq %r9, %rbx
; CHECK-BASELINE-NEXT: movq 24(%rsi), %rsi
; CHECK-BASELINE-NEXT: xorq %r8, %rsi
; CHECK-BASELINE-NEXT: andq 24(%rcx), %rsi
; CHECK-BASELINE-NEXT: andq 16(%rcx), %rbx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: andq 8(%rcx), %rdi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: andq (%rcx), %rdx
; CHECK-BASELINE-NEXT: xorq %r11, %rdx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: xorq %r10, %rdi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: xorq %r9, %rbx
; CHECK-BASELINE-NEXT: xorq %r8, %rsi
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-BASELINE-NEXT: movq %rsi, 24(%rax)
; CHECK-BASELINE-NEXT: movq %rbx, 16(%rax)
; CHECK-BASELINE-NEXT: movq %rdi, 8(%rax)
; CHECK-BASELINE-NEXT: movq %rdx, (%rax)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-BASELINE-NEXT: popq %rbx
; CHECK-BASELINE-NEXT: retq
;
; CHECK-SSE1-LABEL: in_v4i64:
; CHECK-SSE1: # %bb.0:
; CHECK-SSE1-NEXT: pushq %rbx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movq %rdi, %rax
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movq 24(%rdx), %r8
; CHECK-SSE1-NEXT: movq 16(%rdx), %r9
; CHECK-SSE1-NEXT: movq (%rdx), %r11
; CHECK-SSE1-NEXT: movq 8(%rdx), %r10
; CHECK-SSE1-NEXT: movq (%rsi), %rdx
; CHECK-SSE1-NEXT: xorq %r11, %rdx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movq 8(%rsi), %rdi
; CHECK-SSE1-NEXT: xorq %r10, %rdi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: movq 16(%rsi), %rbx
; CHECK-SSE1-NEXT: xorq %r9, %rbx
; CHECK-SSE1-NEXT: movq 24(%rsi), %rsi
; CHECK-SSE1-NEXT: xorq %r8, %rsi
; CHECK-SSE1-NEXT: andq 24(%rcx), %rsi
; CHECK-SSE1-NEXT: andq 16(%rcx), %rbx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: andq 8(%rcx), %rdi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: andq (%rcx), %rdx
; CHECK-SSE1-NEXT: xorq %r11, %rdx
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: xorq %r10, %rdi
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: xorq %r9, %rbx
; CHECK-SSE1-NEXT: xorq %r8, %rsi
[X86] Handle COPYs of physregs better (regalloc hints) Enable enableMultipleCopyHints() on X86. Original Patch by @jonpa: While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling. Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates. Differential Revision: https://reviews.llvm.org/D38128 llvm-svn: 342578
2018-09-20 02:59:08 +08:00
; CHECK-SSE1-NEXT: movq %rsi, 24(%rax)
; CHECK-SSE1-NEXT: movq %rbx, 16(%rax)
; CHECK-SSE1-NEXT: movq %rdi, 8(%rax)
; CHECK-SSE1-NEXT: movq %rdx, (%rax)
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE1-NEXT: popq %rbx
; CHECK-SSE1-NEXT: retq
;
; CHECK-SSE2-LABEL: in_v4i64:
; CHECK-SSE2: # %bb.0:
[DAGCombine][X86][AArch64] Masked merge unfolding: vector edition. Summary: This **appears** to be the last missing piece for the masked merge pattern handling in the backend. This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`), and we need to make sure that they are generated. Differential Revision: https://reviews.llvm.org/D46528 llvm-svn: 332904
2018-05-22 05:41:02 +08:00
; CHECK-SSE2-NEXT: movaps (%rdx), %xmm0
; CHECK-SSE2-NEXT: movaps 16(%rdx), %xmm1
; CHECK-SSE2-NEXT: movaps %xmm0, %xmm2
; CHECK-SSE2-NEXT: andnps (%rsi), %xmm2
; CHECK-SSE2-NEXT: andps (%rdi), %xmm0
; CHECK-SSE2-NEXT: orps %xmm2, %xmm0
; CHECK-SSE2-NEXT: movaps %xmm1, %xmm2
; CHECK-SSE2-NEXT: andnps 16(%rsi), %xmm2
; CHECK-SSE2-NEXT: andps 16(%rdi), %xmm1
; CHECK-SSE2-NEXT: orps %xmm2, %xmm1
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-SSE2-NEXT: retq
;
; CHECK-XOP-LABEL: in_v4i64:
; CHECK-XOP: # %bb.0:
[DAGCombine][X86][AArch64] Masked merge unfolding: vector edition. Summary: This **appears** to be the last missing piece for the masked merge pattern handling in the backend. This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`), and we need to make sure that they are generated. Differential Revision: https://reviews.llvm.org/D46528 llvm-svn: 332904
2018-05-22 05:41:02 +08:00
; CHECK-XOP-NEXT: vmovdqa (%rdi), %ymm0
; CHECK-XOP-NEXT: vmovdqa (%rdx), %ymm1
; CHECK-XOP-NEXT: vpcmov %ymm1, (%rsi), %ymm0, %ymm0
[X86][AArch64][NFC] Add tests for vector masked merge unfolding Summary: This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]]. [[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly. Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`) Now, they would no longer be generated (see `@in`). Differential Revision: https://reviews.llvm.org/D46008 llvm-svn: 332903
2018-05-22 05:40:51 +08:00
; CHECK-XOP-NEXT: retq
%x = load <4 x i64>, <4 x i64> *%px, align 32
%y = load <4 x i64>, <4 x i64> *%py, align 32
%mask = load <4 x i64>, <4 x i64> *%pmask, align 32
%n0 = xor <4 x i64> %x, %y
%n1 = and <4 x i64> %n0, %mask
%r = xor <4 x i64> %n1, %y
ret <4 x i64> %r
}