[AArch64] Fix ldp rename through a bundle

[davemgreen]

std::prev(Paired) will get the previous instruction, that might skip over the instructions in a bundle to the BUNDLE itself. Change it to Paired->getPrevNode() to make sure we update the registers in each instruction in the bundle.

[llvmbot]

@llvm/pr-subscribers-backend-aarch64

Author: David Green (davemgreen)

Changes

std::prev(Paired) will get the previous instruction, that might skip over the instructions in a bundle to the BUNDLE itself. Change it to Paired->getPrevNode() to make sure we update the registers in each instruction in the bundle.

---

Full diff: https://github.com/llvm/llvm-project/pull/146415.diff

2 Files Affected:

• (modified) llvm/lib/Target/AArch64/AArch64LoadStoreOptimizer.cpp (+1-1)
• (modified) llvm/test/CodeGen/AArch64/stp-opt-with-renaming.mir (+29)

diff --git a/llvm/lib/Target/AArch64/AArch64LoadStoreOptimizer.cpp b/llvm/lib/Target/AArch64/AArch64LoadStoreOptimizer.cpp
index f51f0d11ef9d8..5f4aa80980ca5 100644
--- a/llvm/lib/Target/AArch64/AArch64LoadStoreOptimizer.cpp
+++ b/llvm/lib/Target/AArch64/AArch64LoadStoreOptimizer.cpp
@@ -1081,7 +1081,7 @@ AArch64LoadStoreOpt::mergePairedInsns(MachineBasicBlock::iterator I,
           LLVM_DEBUG(dbgs() << "Renamed " << MI);
           return true;
         };
-    forAllMIsUntilDef(MergeForward ? *I : *std::prev(Paired), RegToRename, TRI,
+    forAllMIsUntilDef(MergeForward ? *I : *Paired->getPrevNode(), RegToRename, TRI,
                       UINT32_MAX, UpdateMIs);
 
 #if !defined(NDEBUG)
diff --git a/llvm/test/CodeGen/AArch64/stp-opt-with-renaming.mir b/llvm/test/CodeGen/AArch64/stp-opt-with-renaming.mir
index cbb7f09513aa1..02fef348592cb 100644
--- a/llvm/test/CodeGen/AArch64/stp-opt-with-renaming.mir
+++ b/llvm/test/CodeGen/AArch64/stp-opt-with-renaming.mir
@@ -735,6 +735,35 @@ body:             |
     RET undef $lr
 ...
 ---
+name:            bundled_load_rename
+alignment:       4
+tracksRegLiveness: true
+frameInfo:
+  maxAlignment:     1
+  maxCallFrameSize: 0
+machineFunctionInfo:
+  hasRedZone:      false
+body:             |
+  bb.0.entry:
+    liveins: $z3, $z19, $p0, $z16
+    ; CHECK-LABEL: name: bundled_load_rename
+    ; CHECK: liveins: $z3, $z19, $p0, $z16
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: renamable $q16, $q20 = LDPQi $sp, 0 :: (load (s128)), (load (s128), align 32)
+    ; CHECK-NEXT: BUNDLE implicit-def $z3, implicit-def $q3, implicit-def $d3, implicit-def $s3, implicit-def $h3, implicit-def $b3, implicit $z19, implicit $p0, implicit $z20 {
+    ; CHECK-NEXT:   $z3 = MOVPRFX_ZZ $z19
+    ; CHECK-NEXT:   $z3 = FMUL_ZPmZ_S renamable $p0, killed $z3, $z20
+    ; CHECK-NEXT: }
+    ; CHECK-NEXT: RET undef $lr
+    renamable $q16 = LDRQui $sp, 1 :: (load 16)
+    BUNDLE implicit-def $z3, implicit-def $q3, implicit-def $d3, implicit-def $s3, implicit-def $h3, implicit-def $b3, implicit $z19, implicit $p0, implicit $z16 {
+      $z3 = MOVPRFX_ZZ $z19
+      $z3 = FMUL_ZPmZ_S renamable $p0, killed $z3, renamable $z16
+    }
+    renamable $q16 = LDRQui $sp, 0 :: (load 16, align 32)
+    RET undef $lr
+...
+---
 #
 # CHECK-LABEL: name: bundled_use
 # CHECK: renamable $q0 = LDRQui $sp, 1

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.

[pawosm-arm]

It fixes the miscomputation issue I've faced on Neoverse-V2.