[llvm] [MachineSink] Update register dependency correctly (PR #109763)

[via llvm-commits]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-backend-amdgpu

Author: Ruiling, Song (ruiling)

<details>
<summary>Changes</summary>

The accumulateUsedDefed() was missing if block prologue interference check does not pass. This would cause incorrect register dependency, which cause incorrect sinking.

---
Full diff: https://github.com/llvm/llvm-project/pull/109763.diff


2 Files Affected:

- (modified) llvm/lib/CodeGen/MachineSink.cpp (+3-2) 
- (added) llvm/test/CodeGen/AMDGPU/postra-sink-update-dependency.mir (+66) 


``````````diff
diff --git a/llvm/lib/CodeGen/MachineSink.cpp b/llvm/lib/CodeGen/MachineSink.cpp
index 609f9af9767f5d..658ebd47488c71 100644
--- a/llvm/lib/CodeGen/MachineSink.cpp
+++ b/llvm/lib/CodeGen/MachineSink.cpp
@@ -2152,8 +2152,9 @@ bool PostRAMachineSinking::tryToSinkCopy(MachineBasicBlock &CurBB,
     MachineBasicBlock::iterator InsertPos =
         SuccBB->SkipPHIsAndLabels(SuccBB->begin());
     if (blockPrologueInterferes(SuccBB, InsertPos, MI, TRI, TII, nullptr)) {
-      LLVM_DEBUG(
-          dbgs() << " *** Not sinking: prologue interference\n");
+      LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, UsedRegUnits,
+                                        TRI);
+      LLVM_DEBUG(dbgs() << " *** Not sinking: prologue interference\n");
       continue;
     }
 
diff --git a/llvm/test/CodeGen/AMDGPU/postra-sink-update-dependency.mir b/llvm/test/CodeGen/AMDGPU/postra-sink-update-dependency.mir
new file mode 100644
index 00000000000000..14617e066f954c
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/postra-sink-update-dependency.mir
@@ -0,0 +1,66 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 5
+# RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -run-pass=postra-machine-sink -verify-machineinstrs -o - %s | FileCheck %s
+#
+# In the example, the ` $sgpr4 = COPY $sgpr2` was incorrectly sunk into bb.3. This happened because we did not update
+# register uses when we found that `$sgpr2 = COPY $sgpr3` should not be sunk because of conflict with the successor's
+# prologue instructions.
+---
+name:            update_dependency_correctly
+tracksRegLiveness: true
+body:             |
+  ; CHECK-LABEL: name: update_dependency_correctly
+  ; CHECK: bb.0:
+  ; CHECK-NEXT:   successors: %bb.2(0x40000000), %bb.1(0x40000000)
+  ; CHECK-NEXT:   liveins: $sgpr0, $sgpr3, $sgpr2
+  ; CHECK-NEXT: {{  $}}
+  ; CHECK-NEXT:   $vgpr1 = IMPLICIT_DEF
+  ; CHECK-NEXT:   renamable $sgpr4 = COPY $sgpr2
+  ; CHECK-NEXT:   renamable $sgpr2 = COPY $sgpr3
+  ; CHECK-NEXT:   $vgpr1 = SI_SPILL_S32_TO_VGPR $sgpr0, 0, $vgpr1
+  ; CHECK-NEXT:   $sgpr1 = S_AND_SAVEEXEC_B32 $sgpr0, implicit-def $exec, implicit-def $scc, implicit $exec
+  ; CHECK-NEXT:   S_CBRANCH_EXECZ %bb.1, implicit $exec
+  ; CHECK-NEXT:   S_BRANCH %bb.2
+  ; CHECK-NEXT: {{  $}}
+  ; CHECK-NEXT: bb.1:
+  ; CHECK-NEXT:   S_ENDPGM 0
+  ; CHECK-NEXT: {{  $}}
+  ; CHECK-NEXT: bb.2:
+  ; CHECK-NEXT:   successors: %bb.1(0x80000000)
+  ; CHECK-NEXT:   liveins: $sgpr0, $sgpr2, $sgpr4, $vgpr1
+  ; CHECK-NEXT: {{  $}}
+  ; CHECK-NEXT:   $sgpr3 = SI_RESTORE_S32_FROM_VGPR $vgpr1, 0
+  ; CHECK-NEXT:   renamable $sgpr0_sgpr1 = S_GETPC_B64_pseudo
+  ; CHECK-NEXT:   renamable $sgpr5 = COPY $sgpr1
+  ; CHECK-NEXT:   renamable $sgpr0_sgpr1 = S_LOAD_DWORDX2_IMM renamable $sgpr4_sgpr5, 32, 0
+  ; CHECK-NEXT:   S_BRANCH %bb.1
+  bb.0:
+    successors: %bb.3(0x40000000), %bb.2(0x40000000)
+    liveins: $sgpr0, $sgpr3, $sgpr2
+
+    $vgpr1 = IMPLICIT_DEF
+
+    renamable $sgpr4 = COPY $sgpr2
+    renamable $sgpr2 = COPY $sgpr3
+
+    $vgpr1 = SI_SPILL_S32_TO_VGPR $sgpr0, 0, $vgpr1
+
+    $sgpr1 = S_AND_SAVEEXEC_B32 $sgpr0, implicit-def $exec, implicit-def $scc, implicit $exec
+    S_CBRANCH_EXECZ %bb.2, implicit $exec
+    S_BRANCH %bb.3
+
+  bb.2:
+    S_ENDPGM 0
+
+  bb.3:
+    successors: %bb.2(0x40000000)
+    liveins: $sgpr0, $sgpr2, $sgpr4, $vgpr1
+
+    $sgpr3 = SI_RESTORE_S32_FROM_VGPR $vgpr1, 0
+
+    renamable $sgpr0_sgpr1 = S_GETPC_B64_pseudo
+    renamable $sgpr5 = COPY $sgpr1
+    renamable $sgpr0_sgpr1 = S_LOAD_DWORDX2_IMM renamable $sgpr4_sgpr5, 32, 0
+
+    S_BRANCH %bb.2
+
+...

``````````

</details>


https://github.com/llvm/llvm-project/pull/109763