[llvm-branch-commits] [llvm] [X86][APX] Fix commuting SHLD/SHRD NF, NF_ND and EVEX variants #233873 (PR #241081)

Feng Zou via llvm-branch-commits llvm-branch-commits at lists.llvm.org
Tue Nov 10 17:47:04 PST 2026


https://github.com/fzou1 created https://github.com/llvm/llvm-project/pull/241081

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>

Cherry-pick from https://github.com/llvm/llvm-project/pull/233873

>From 03fd6f83623721ce85a96e7741d631900503567b Mon Sep 17 00:00:00 2001
From: Feng Zou <feng.zou at intel.com>
Date: Sun, 25 Oct 2026 06:08:02 +0100
Subject: [PATCH 1/2] [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>
---
 llvm/lib/Target/X86/X86InstrInfo.cpp          |  50 ++-
 .../apx/compress-evex-shld-shrd-commute.mir   | 189 ++++++++
 .../X86/apx/two-address-shld-shrd-commute.mir | 422 ++++++++++++++++++
 3 files changed, 641 insertions(+), 20 deletions(-)
 create mode 100644 llvm/test/CodeGen/X86/apx/compress-evex-shld-shrd-commute.mir
 create mode 100644 llvm/test/CodeGen/X86/apx/two-address-shld-shrd-commute.mir

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
+...

>From 1e87f1c476c634964c8431d95cc8954daa69664a Mon Sep 17 00:00:00 2001
From: Feng Zou <feng.zou at intel.com>
Date: Tue, 10 Nov 2026 23:10:01 +0800
Subject: [PATCH 2/2] Fix pass name in LIT test

---
 llvm/test/CodeGen/X86/apx/two-address-shld-shrd-commute.mir | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

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
index 963a5cf27d908f..330a300023a89b 100644
--- a/llvm/test/CodeGen/X86/apx/two-address-shld-shrd-commute.mir
+++ b/llvm/test/CodeGen/X86/apx/two-address-shld-shrd-commute.mir
@@ -1,5 +1,5 @@
 # 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
+# RUN: llc -mtriple=x86_64-- -mattr=+nf -run-pass=twoaddressinstruction %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



More information about the llvm-branch-commits mailing list