[llvm-branch-commits] [llvm] f4c8597 - [X86][APX] Fix commuting SHLD/SHRD NF, NF_ND and EVEX variants (#233873)

Tobias Hieta via llvm-branch-commits llvm-branch-commits at lists.llvm.org
Mon Nov 16 04:46:48 PST 2026


Author: Feng Zou
Date: 2026-11-16T13:46:04+01:00
New Revision: f4c8597e999118143db55f26ee708ec38a80f00f

URL: https://github.com/llvm/llvm-project/commit/f4c8597e999118143db55f26ee708ec38a80f00f
DIFF: https://github.com/llvm/llvm-project/commit/f4c8597e999118143db55f26ee708ec38a80f00f.diff

LOG: [X86][APX] Fix commuting SHLD/SHRD NF, NF_ND and EVEX variants (#233873)

All SHLD/SHRD rri8 variants are marked commutable, but
commuteInstructionImpl only rewrote the legacy and _ND opcodes as SHLD
B, C, I <-> SHRD C, B, (BitWidth - I). The _NF, _NF_ND and _EVEX
variants fell through to the generic commute, which only swaps the
sources and changes the result.

This is reachable when optimizeCompareInstr converts an EFLAGS clobber
to its NF form (e.g. SHLD16rri8_ND -> SHLD16rri8_NF_ND) and RA assigns
the destination to the second source: X86CompressEVEX commutes it to
drop the redundant NDD destination and emits the wrong shift. The
TwoAddressInstruction pass can hit the same bug when commuting the
non-NDD _NF and _EVEX forms.

Assisted-By: Claude Opus 5.5 <noreply at anthropic.com>

Added: 
    llvm/test/CodeGen/X86/apx/compress-evex-shld-shrd-commute.mir
    llvm/test/CodeGen/X86/apx/two-address-shld-shrd-commute.mir

Modified: 
    llvm/lib/Target/X86/X86InstrInfo.cpp

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/X86/X86InstrInfo.cpp b/llvm/lib/Target/X86/X86InstrInfo.cpp
index 7b459f76b012fd..64c60c8b36f7b8 100644
--- a/llvm/lib/Target/X86/X86InstrInfo.cpp
+++ b/llvm/lib/Target/X86/X86InstrInfo.cpp
@@ -2303,46 +2303,56 @@ MachineInstr *X86InstrInfo::commuteInstructionImpl(MachineInstr &MI, bool NewMI,
   case X86::OP:                                                                \
   case X86::OP##_ND:
 
+#define CASE_SHXD(OP)                                                          \
+  CASE_ND(OP)                                                                  \
+  case X86::OP##_NF:                                                           \
+  case X86::OP##_NF_ND:                                                        \
+  case X86::OP##_EVEX:
+
+  // clang-format off
   switch (Opc) {
-  // SHLD B, C, I <-> SHRD C, B, (BitWidth - I)
-  CASE_ND(SHRD16rri8)
-  CASE_ND(SHLD16rri8)
-  CASE_ND(SHRD32rri8)
-  CASE_ND(SHLD32rri8)
-  CASE_ND(SHRD64rri8)
-  CASE_ND(SHLD64rri8) {
+  // SHLD B, C, I <-> SHRD C, B, (BitWidth - I). All rri8 variants (legacy,
+  // _ND, _NF, _NF_ND, _EVEX) are commutable; swapping only the sources would
+  // change the result, so the opcode and immediate must be updated too.
+  CASE_SHXD(SHRD16rri8)
+  CASE_SHXD(SHLD16rri8)
+  CASE_SHXD(SHRD32rri8)
+  CASE_SHXD(SHLD32rri8)
+  CASE_SHXD(SHRD64rri8)
+  CASE_SHXD(SHLD64rri8) {
     unsigned Size;
     switch (Opc) {
     default:
       llvm_unreachable("Unreachable!");
-#define FROM_TO_SIZE(A, B, S)                                                  \
-  case X86::A:                                                                 \
-    Opc = X86::B;                                                              \
-    Size = S;                                                                  \
-    break;                                                                     \
-  case X86::A##_ND:                                                            \
-    Opc = X86::B##_ND;                                                         \
+#define FROM_TO_SUFFIX(A, B, S, SFX)                                           \
+  case X86::A##SFX:                                                            \
+    Opc = X86::B##SFX;                                                         \
     Size = S;                                                                  \
     break;                                                                     \
-  case X86::B:                                                                 \
-    Opc = X86::A;                                                              \
-    Size = S;                                                                  \
-    break;                                                                     \
-  case X86::B##_ND:                                                            \
-    Opc = X86::A##_ND;                                                         \
+  case X86::B##SFX:                                                            \
+    Opc = X86::A##SFX;                                                         \
     Size = S;                                                                  \
     break;
+#define FROM_TO_SIZE(A, B, S)                                                  \
+  FROM_TO_SUFFIX(A, B, S, )                                                    \
+  FROM_TO_SUFFIX(A, B, S, _ND)                                                 \
+  FROM_TO_SUFFIX(A, B, S, _NF)                                                 \
+  FROM_TO_SUFFIX(A, B, S, _NF_ND)                                              \
+  FROM_TO_SUFFIX(A, B, S, _EVEX)
 
     FROM_TO_SIZE(SHRD16rri8, SHLD16rri8, 16)
     FROM_TO_SIZE(SHRD32rri8, SHLD32rri8, 32)
     FROM_TO_SIZE(SHRD64rri8, SHLD64rri8, 64)
 #undef FROM_TO_SIZE
+#undef FROM_TO_SUFFIX
+#undef CASE_SHXD
     }
     WorkingMI = CloneIfNew(MI);
     WorkingMI->setDesc(get(Opc));
     WorkingMI->getOperand(3).setImm(Size - MI.getOperand(3).getImm());
     break;
   }
+  // clang-format on
   case X86::PFSUBrr:
   case X86::PFSUBRrr:
     // PFSUB  x, y: x = x - y

diff  --git a/llvm/test/CodeGen/X86/apx/compress-evex-shld-shrd-commute.mir b/llvm/test/CodeGen/X86/apx/compress-evex-shld-shrd-commute.mir
new file mode 100644
index 00000000000000..eed1ce2f64b355
--- /dev/null
+++ b/llvm/test/CodeGen/X86/apx/compress-evex-shld-shrd-commute.mir
@@ -0,0 +1,189 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 6
+# RUN: llc -mtriple=x86_64-- -mattr=+ndd,+nf -run-pass=x86-compress-evex %s -o - | FileCheck %s
+
+# When the destination of an NDD SHLD/SHRD matches its second source,
+# x86-compress-evex commutes it to drop the NDD form. The commute must also
+# switch SHLD <-> SHRD and use BitWidth - imm, not just swap the sources:
+#   SHLD B, C, I <-> SHRD C, B, (BitWidth - I)
+---
+name: shld16_nf_nd
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $cx, $r8w
+    ; CHECK-LABEL: name: shld16_nf_nd
+    ; CHECK: liveins: $cx, $r8w
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: $r8w = SHRD16rri8_NF $r8w, $cx, 8
+    ; CHECK-NEXT: RET64 $r8w
+    $r8w = SHLD16rri8_NF_ND $cx, $r8w, 8
+    RET64 $r8w
+...
+---
+name: shld32_nf_nd
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $ecx, $r8d
+    ; CHECK-LABEL: name: shld32_nf_nd
+    ; CHECK: liveins: $ecx, $r8d
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: $r8d = SHRD32rri8_NF $r8d, $ecx, 27
+    ; CHECK-NEXT: RET64 $r8d
+    $r8d = SHLD32rri8_NF_ND $ecx, $r8d, 5
+    RET64 $r8d
+...
+---
+name: shld64_nf_nd
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $rcx, $r8
+    ; CHECK-LABEL: name: shld64_nf_nd
+    ; CHECK: liveins: $rcx, $r8
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: $r8 = SHRD64rri8_NF $r8, $rcx, 51
+    ; CHECK-NEXT: RET64 $r8
+    $r8 = SHLD64rri8_NF_ND $rcx, $r8, 13
+    RET64 $r8
+...
+---
+name: shrd16_nf_nd
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $cx, $r8w
+    ; CHECK-LABEL: name: shrd16_nf_nd
+    ; CHECK: liveins: $cx, $r8w
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: $r8w = SHLD16rri8_NF $r8w, $cx, 13
+    ; CHECK-NEXT: RET64 $r8w
+    $r8w = SHRD16rri8_NF_ND $cx, $r8w, 3
+    RET64 $r8w
+...
+---
+name: shrd32_nf_nd
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $ecx, $r8d
+    ; CHECK-LABEL: name: shrd32_nf_nd
+    ; CHECK: liveins: $ecx, $r8d
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: $r8d = SHLD32rri8_NF $r8d, $ecx, 25
+    ; CHECK-NEXT: RET64 $r8d
+    $r8d = SHRD32rri8_NF_ND $ecx, $r8d, 7
+    RET64 $r8d
+...
+---
+name: shrd64_nf_nd
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $rcx, $r8
+    ; CHECK-LABEL: name: shrd64_nf_nd
+    ; CHECK: liveins: $rcx, $r8
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: $r8 = SHLD64rri8_NF $r8, $rcx, 24
+    ; CHECK-NEXT: RET64 $r8
+    $r8 = SHRD64rri8_NF_ND $rcx, $r8, 40
+    RET64 $r8
+...
+---
+name: shld16_nd
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $cx, $r8w
+    ; CHECK-LABEL: name: shld16_nd
+    ; CHECK: liveins: $cx, $r8w
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: $r8w = SHRD16rri8 $r8w, $cx, 8, implicit-def dead $eflags
+    ; CHECK-NEXT: RET64 $r8w
+    $r8w = SHLD16rri8_ND $cx, $r8w, 8, implicit-def dead $eflags
+    RET64 $r8w
+...
+---
+name: shld32_nd
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $ecx, $r8d
+    ; CHECK-LABEL: name: shld32_nd
+    ; CHECK: liveins: $ecx, $r8d
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: $r8d = SHRD32rri8 $r8d, $ecx, 27, implicit-def dead $eflags
+    ; CHECK-NEXT: RET64 $r8d
+    $r8d = SHLD32rri8_ND $ecx, $r8d, 5, implicit-def dead $eflags
+    RET64 $r8d
+...
+---
+name: shld64_nd
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $rcx, $r8
+    ; CHECK-LABEL: name: shld64_nd
+    ; CHECK: liveins: $rcx, $r8
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: $r8 = SHRD64rri8 $r8, $rcx, 51, implicit-def dead $eflags
+    ; CHECK-NEXT: RET64 $r8
+    $r8 = SHLD64rri8_ND $rcx, $r8, 13, implicit-def dead $eflags
+    RET64 $r8
+...
+---
+name: shrd16_nd
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $cx, $r8w
+    ; CHECK-LABEL: name: shrd16_nd
+    ; CHECK: liveins: $cx, $r8w
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: $r8w = SHLD16rri8 $r8w, $cx, 13, implicit-def dead $eflags
+    ; CHECK-NEXT: RET64 $r8w
+    $r8w = SHRD16rri8_ND $cx, $r8w, 3, implicit-def dead $eflags
+    RET64 $r8w
+...
+---
+name: shrd32_nd
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $ecx, $r8d
+    ; CHECK-LABEL: name: shrd32_nd
+    ; CHECK: liveins: $ecx, $r8d
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: $r8d = SHLD32rri8 $r8d, $ecx, 25, implicit-def dead $eflags
+    ; CHECK-NEXT: RET64 $r8d
+    $r8d = SHRD32rri8_ND $ecx, $r8d, 7, implicit-def dead $eflags
+    RET64 $r8d
+...
+---
+name: shrd64_nd
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $rcx, $r8
+    ; CHECK-LABEL: name: shrd64_nd
+    ; CHECK: liveins: $rcx, $r8
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: $r8 = SHLD64rri8 $r8, $rcx, 24, implicit-def dead $eflags
+    ; CHECK-NEXT: RET64 $r8
+    $r8 = SHRD64rri8_ND $rcx, $r8, 40, implicit-def dead $eflags
+    RET64 $r8
+...
+---
+name: shld32_nf_nd_dst_is_src1
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $r8d, $ecx
+    ; CHECK-LABEL: name: shld32_nf_nd_dst_is_src1
+    ; CHECK: liveins: $r8d, $ecx
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: $r8d = SHLD32rri8_NF $r8d, $ecx, 5
+    ; CHECK-NEXT: RET64 $r8d
+    $r8d = SHLD32rri8_NF_ND $r8d, $ecx, 5
+    RET64 $r8d
+...

diff  --git a/llvm/test/CodeGen/X86/apx/two-address-shld-shrd-commute.mir b/llvm/test/CodeGen/X86/apx/two-address-shld-shrd-commute.mir
new file mode 100644
index 00000000000000..963a5cf27d908f
--- /dev/null
+++ b/llvm/test/CodeGen/X86/apx/two-address-shld-shrd-commute.mir
@@ -0,0 +1,422 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 6
+# RUN: llc -mtriple=x86_64-- -mattr=+nf -run-pass=two-address-instruction %s -o - | FileCheck %s
+
+# When the first source stays live and the second is killed, two-address
+# commutes SHLD/SHRD to tie the second source to the destination. The commute
+# must also switch SHLD <-> SHRD and use BitWidth - imm, not just swap the
+# sources:
+#   SHLD B, C, I <-> SHRD C, B, (BitWidth - I)
+---
+name: shld16_legacy
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $di, $si
+    ; CHECK-LABEL: name: shld16_legacy
+    ; CHECK: liveins: $di, $si
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:gr16 = COPY $di
+    ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gr16 = COPY $si
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr16 = COPY [[COPY1]]
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr16 = SHRD16rri8 [[COPY2]], [[COPY]], 13, implicit-def dead $eflags
+    ; CHECK-NEXT: $ax = COPY [[COPY2]]
+    ; CHECK-NEXT: $cx = COPY [[COPY]]
+    ; CHECK-NEXT: RET64 $ax, $cx
+    %0:gr16 = COPY $di
+    %1:gr16 = COPY $si
+    %2:gr16 = SHLD16rri8 %0, killed %1, 3, implicit-def dead $eflags
+    $ax = COPY %2
+    $cx = COPY %0
+    RET64 $ax, $cx
+...
+---
+name: shld32_legacy
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $edi, $esi
+    ; CHECK-LABEL: name: shld32_legacy
+    ; CHECK: liveins: $edi, $esi
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:gr32 = COPY $edi
+    ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gr32 = COPY $esi
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr32 = COPY [[COPY1]]
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr32 = SHRD32rri8 [[COPY2]], [[COPY]], 29, implicit-def dead $eflags
+    ; CHECK-NEXT: $eax = COPY [[COPY2]]
+    ; CHECK-NEXT: $ecx = COPY [[COPY]]
+    ; CHECK-NEXT: RET64 $eax, $ecx
+    %0:gr32 = COPY $edi
+    %1:gr32 = COPY $esi
+    %2:gr32 = SHLD32rri8 %0, killed %1, 3, implicit-def dead $eflags
+    $eax = COPY %2
+    $ecx = COPY %0
+    RET64 $eax, $ecx
+...
+---
+name: shld64_legacy
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $rdi, $rsi
+    ; CHECK-LABEL: name: shld64_legacy
+    ; CHECK: liveins: $rdi, $rsi
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:gr64 = COPY $rdi
+    ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gr64 = COPY $rsi
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr64 = COPY [[COPY1]]
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr64 = SHRD64rri8 [[COPY2]], [[COPY]], 61, implicit-def dead $eflags
+    ; CHECK-NEXT: $rax = COPY [[COPY2]]
+    ; CHECK-NEXT: $rcx = COPY [[COPY]]
+    ; CHECK-NEXT: RET64 $rax, $rcx
+    %0:gr64 = COPY $rdi
+    %1:gr64 = COPY $rsi
+    %2:gr64 = SHLD64rri8 %0, killed %1, 3, implicit-def dead $eflags
+    $rax = COPY %2
+    $rcx = COPY %0
+    RET64 $rax, $rcx
+...
+---
+name: shld16_nf
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $di, $si
+    ; CHECK-LABEL: name: shld16_nf
+    ; CHECK: liveins: $di, $si
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:gr16 = COPY $di
+    ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gr16 = COPY $si
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr16 = COPY [[COPY1]]
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr16 = SHRD16rri8_NF [[COPY2]], [[COPY]], 13
+    ; CHECK-NEXT: $ax = COPY [[COPY2]]
+    ; CHECK-NEXT: $cx = COPY [[COPY]]
+    ; CHECK-NEXT: RET64 $ax, $cx
+    %0:gr16 = COPY $di
+    %1:gr16 = COPY $si
+    %2:gr16 = SHLD16rri8_NF %0, killed %1, 3
+    $ax = COPY %2
+    $cx = COPY %0
+    RET64 $ax, $cx
+...
+---
+name: shld32_nf
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $edi, $esi
+    ; CHECK-LABEL: name: shld32_nf
+    ; CHECK: liveins: $edi, $esi
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:gr32 = COPY $edi
+    ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gr32 = COPY $esi
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr32 = COPY [[COPY1]]
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr32 = SHRD32rri8_NF [[COPY2]], [[COPY]], 29
+    ; CHECK-NEXT: $eax = COPY [[COPY2]]
+    ; CHECK-NEXT: $ecx = COPY [[COPY]]
+    ; CHECK-NEXT: RET64 $eax, $ecx
+    %0:gr32 = COPY $edi
+    %1:gr32 = COPY $esi
+    %2:gr32 = SHLD32rri8_NF %0, killed %1, 3
+    $eax = COPY %2
+    $ecx = COPY %0
+    RET64 $eax, $ecx
+...
+---
+name: shld64_nf
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $rdi, $rsi
+    ; CHECK-LABEL: name: shld64_nf
+    ; CHECK: liveins: $rdi, $rsi
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:gr64 = COPY $rdi
+    ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gr64 = COPY $rsi
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr64 = COPY [[COPY1]]
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr64 = SHRD64rri8_NF [[COPY2]], [[COPY]], 61
+    ; CHECK-NEXT: $rax = COPY [[COPY2]]
+    ; CHECK-NEXT: $rcx = COPY [[COPY]]
+    ; CHECK-NEXT: RET64 $rax, $rcx
+    %0:gr64 = COPY $rdi
+    %1:gr64 = COPY $rsi
+    %2:gr64 = SHLD64rri8_NF %0, killed %1, 3
+    $rax = COPY %2
+    $rcx = COPY %0
+    RET64 $rax, $rcx
+...
+---
+name: shld16_evex
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $di, $si
+    ; CHECK-LABEL: name: shld16_evex
+    ; CHECK: liveins: $di, $si
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:gr16 = COPY $di
+    ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gr16 = COPY $si
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr16 = COPY [[COPY1]]
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr16 = SHRD16rri8_EVEX [[COPY2]], [[COPY]], 13, implicit-def dead $eflags
+    ; CHECK-NEXT: $ax = COPY [[COPY2]]
+    ; CHECK-NEXT: $cx = COPY [[COPY]]
+    ; CHECK-NEXT: RET64 $ax, $cx
+    %0:gr16 = COPY $di
+    %1:gr16 = COPY $si
+    %2:gr16 = SHLD16rri8_EVEX %0, killed %1, 3, implicit-def dead $eflags
+    $ax = COPY %2
+    $cx = COPY %0
+    RET64 $ax, $cx
+...
+---
+name: shld32_evex
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $edi, $esi
+    ; CHECK-LABEL: name: shld32_evex
+    ; CHECK: liveins: $edi, $esi
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:gr32 = COPY $edi
+    ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gr32 = COPY $esi
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr32 = COPY [[COPY1]]
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr32 = SHRD32rri8_EVEX [[COPY2]], [[COPY]], 29, implicit-def dead $eflags
+    ; CHECK-NEXT: $eax = COPY [[COPY2]]
+    ; CHECK-NEXT: $ecx = COPY [[COPY]]
+    ; CHECK-NEXT: RET64 $eax, $ecx
+    %0:gr32 = COPY $edi
+    %1:gr32 = COPY $esi
+    %2:gr32 = SHLD32rri8_EVEX %0, killed %1, 3, implicit-def dead $eflags
+    $eax = COPY %2
+    $ecx = COPY %0
+    RET64 $eax, $ecx
+...
+---
+name: shld64_evex
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $rdi, $rsi
+    ; CHECK-LABEL: name: shld64_evex
+    ; CHECK: liveins: $rdi, $rsi
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:gr64 = COPY $rdi
+    ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gr64 = COPY $rsi
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr64 = COPY [[COPY1]]
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr64 = SHRD64rri8_EVEX [[COPY2]], [[COPY]], 61, implicit-def dead $eflags
+    ; CHECK-NEXT: $rax = COPY [[COPY2]]
+    ; CHECK-NEXT: $rcx = COPY [[COPY]]
+    ; CHECK-NEXT: RET64 $rax, $rcx
+    %0:gr64 = COPY $rdi
+    %1:gr64 = COPY $rsi
+    %2:gr64 = SHLD64rri8_EVEX %0, killed %1, 3, implicit-def dead $eflags
+    $rax = COPY %2
+    $rcx = COPY %0
+    RET64 $rax, $rcx
+...
+---
+name: shrd16_legacy
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $di, $si
+    ; CHECK-LABEL: name: shrd16_legacy
+    ; CHECK: liveins: $di, $si
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:gr16 = COPY $di
+    ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gr16 = COPY $si
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr16 = COPY [[COPY1]]
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr16 = SHLD16rri8 [[COPY2]], [[COPY]], 13, implicit-def dead $eflags
+    ; CHECK-NEXT: $ax = COPY [[COPY2]]
+    ; CHECK-NEXT: $cx = COPY [[COPY]]
+    ; CHECK-NEXT: RET64 $ax, $cx
+    %0:gr16 = COPY $di
+    %1:gr16 = COPY $si
+    %2:gr16 = SHRD16rri8 %0, killed %1, 3, implicit-def dead $eflags
+    $ax = COPY %2
+    $cx = COPY %0
+    RET64 $ax, $cx
+...
+---
+name: shrd32_legacy
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $edi, $esi
+    ; CHECK-LABEL: name: shrd32_legacy
+    ; CHECK: liveins: $edi, $esi
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:gr32 = COPY $edi
+    ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gr32 = COPY $esi
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr32 = COPY [[COPY1]]
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr32 = SHLD32rri8 [[COPY2]], [[COPY]], 29, implicit-def dead $eflags
+    ; CHECK-NEXT: $eax = COPY [[COPY2]]
+    ; CHECK-NEXT: $ecx = COPY [[COPY]]
+    ; CHECK-NEXT: RET64 $eax, $ecx
+    %0:gr32 = COPY $edi
+    %1:gr32 = COPY $esi
+    %2:gr32 = SHRD32rri8 %0, killed %1, 3, implicit-def dead $eflags
+    $eax = COPY %2
+    $ecx = COPY %0
+    RET64 $eax, $ecx
+...
+---
+name: shrd64_legacy
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $rdi, $rsi
+    ; CHECK-LABEL: name: shrd64_legacy
+    ; CHECK: liveins: $rdi, $rsi
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:gr64 = COPY $rdi
+    ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gr64 = COPY $rsi
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr64 = COPY [[COPY1]]
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr64 = SHLD64rri8 [[COPY2]], [[COPY]], 61, implicit-def dead $eflags
+    ; CHECK-NEXT: $rax = COPY [[COPY2]]
+    ; CHECK-NEXT: $rcx = COPY [[COPY]]
+    ; CHECK-NEXT: RET64 $rax, $rcx
+    %0:gr64 = COPY $rdi
+    %1:gr64 = COPY $rsi
+    %2:gr64 = SHRD64rri8 %0, killed %1, 3, implicit-def dead $eflags
+    $rax = COPY %2
+    $rcx = COPY %0
+    RET64 $rax, $rcx
+...
+---
+name: shrd16_nf
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $di, $si
+    ; CHECK-LABEL: name: shrd16_nf
+    ; CHECK: liveins: $di, $si
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:gr16 = COPY $di
+    ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gr16 = COPY $si
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr16 = COPY [[COPY1]]
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr16 = SHLD16rri8_NF [[COPY2]], [[COPY]], 13
+    ; CHECK-NEXT: $ax = COPY [[COPY2]]
+    ; CHECK-NEXT: $cx = COPY [[COPY]]
+    ; CHECK-NEXT: RET64 $ax, $cx
+    %0:gr16 = COPY $di
+    %1:gr16 = COPY $si
+    %2:gr16 = SHRD16rri8_NF %0, killed %1, 3
+    $ax = COPY %2
+    $cx = COPY %0
+    RET64 $ax, $cx
+...
+---
+name: shrd32_nf
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $edi, $esi
+    ; CHECK-LABEL: name: shrd32_nf
+    ; CHECK: liveins: $edi, $esi
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:gr32 = COPY $edi
+    ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gr32 = COPY $esi
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr32 = COPY [[COPY1]]
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr32 = SHLD32rri8_NF [[COPY2]], [[COPY]], 29
+    ; CHECK-NEXT: $eax = COPY [[COPY2]]
+    ; CHECK-NEXT: $ecx = COPY [[COPY]]
+    ; CHECK-NEXT: RET64 $eax, $ecx
+    %0:gr32 = COPY $edi
+    %1:gr32 = COPY $esi
+    %2:gr32 = SHRD32rri8_NF %0, killed %1, 3
+    $eax = COPY %2
+    $ecx = COPY %0
+    RET64 $eax, $ecx
+...
+---
+name: shrd64_nf
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $rdi, $rsi
+    ; CHECK-LABEL: name: shrd64_nf
+    ; CHECK: liveins: $rdi, $rsi
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:gr64 = COPY $rdi
+    ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gr64 = COPY $rsi
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr64 = COPY [[COPY1]]
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr64 = SHLD64rri8_NF [[COPY2]], [[COPY]], 61
+    ; CHECK-NEXT: $rax = COPY [[COPY2]]
+    ; CHECK-NEXT: $rcx = COPY [[COPY]]
+    ; CHECK-NEXT: RET64 $rax, $rcx
+    %0:gr64 = COPY $rdi
+    %1:gr64 = COPY $rsi
+    %2:gr64 = SHRD64rri8_NF %0, killed %1, 3
+    $rax = COPY %2
+    $rcx = COPY %0
+    RET64 $rax, $rcx
+...
+---
+name: shrd16_evex
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $di, $si
+    ; CHECK-LABEL: name: shrd16_evex
+    ; CHECK: liveins: $di, $si
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:gr16 = COPY $di
+    ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gr16 = COPY $si
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr16 = COPY [[COPY1]]
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr16 = SHLD16rri8_EVEX [[COPY2]], [[COPY]], 13, implicit-def dead $eflags
+    ; CHECK-NEXT: $ax = COPY [[COPY2]]
+    ; CHECK-NEXT: $cx = COPY [[COPY]]
+    ; CHECK-NEXT: RET64 $ax, $cx
+    %0:gr16 = COPY $di
+    %1:gr16 = COPY $si
+    %2:gr16 = SHRD16rri8_EVEX %0, killed %1, 3, implicit-def dead $eflags
+    $ax = COPY %2
+    $cx = COPY %0
+    RET64 $ax, $cx
+...
+---
+name: shrd32_evex
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $edi, $esi
+    ; CHECK-LABEL: name: shrd32_evex
+    ; CHECK: liveins: $edi, $esi
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:gr32 = COPY $edi
+    ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gr32 = COPY $esi
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr32 = COPY [[COPY1]]
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr32 = SHLD32rri8_EVEX [[COPY2]], [[COPY]], 29, implicit-def dead $eflags
+    ; CHECK-NEXT: $eax = COPY [[COPY2]]
+    ; CHECK-NEXT: $ecx = COPY [[COPY]]
+    ; CHECK-NEXT: RET64 $eax, $ecx
+    %0:gr32 = COPY $edi
+    %1:gr32 = COPY $esi
+    %2:gr32 = SHRD32rri8_EVEX %0, killed %1, 3, implicit-def dead $eflags
+    $eax = COPY %2
+    $ecx = COPY %0
+    RET64 $eax, $ecx
+...
+---
+name: shrd64_evex
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $rdi, $rsi
+    ; CHECK-LABEL: name: shrd64_evex
+    ; CHECK: liveins: $rdi, $rsi
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:gr64 = COPY $rdi
+    ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gr64 = COPY $rsi
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr64 = COPY [[COPY1]]
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr64 = SHLD64rri8_EVEX [[COPY2]], [[COPY]], 61, implicit-def dead $eflags
+    ; CHECK-NEXT: $rax = COPY [[COPY2]]
+    ; CHECK-NEXT: $rcx = COPY [[COPY]]
+    ; CHECK-NEXT: RET64 $rax, $rcx
+    %0:gr64 = COPY $rdi
+    %1:gr64 = COPY $rsi
+    %2:gr64 = SHRD64rri8_EVEX %0, killed %1, 3, implicit-def dead $eflags
+    $rax = COPY %2
+    $rcx = COPY %0
+    RET64 $rax, $rcx
+...


        


More information about the llvm-branch-commits mailing list