[GlobalISel] Fix fewerElementsVectorPhi to insert after G_PHIs

[dfszabo]

Currently the inserted mergelike instructions will be inserted at the location of the G_PHI. Seems like the behaviour was correct before, but the rework done in https://reviews.llvm.org/D114198 forgot to include the part which makes sure the instructions will be inserted after all the G_PHIs.

[llvmbot]

@llvm/pr-subscribers-llvm-globalisel

Author: Dávid Ferenc Szabó (dfszabo)

Changes

Currently the inserted mergelike instructions will be inserted at the location of the G_PHI. Seems like the behaviour was correct before, but the rework done in https://reviews.llvm.org/D114198 forgot to include the part which makes sure the instructions will be inserted after all the G_PHIs.

---

Full diff: https://github.com/llvm/llvm-project/pull/87927.diff

1 Files Affected:

• (modified) llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp (+4)

diff --git a/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp b/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp
index 95c6a359e52ec6..9412fa85cd3856 100644
--- a/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp
+++ b/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp
@@ -4265,6 +4265,10 @@ LegalizerHelper::fewerElementsVectorPhi(GenericMachineInstr &MI,
     }
   }
 
+  // Set the insert point after the existing PHIs
+  MachineBasicBlock &MBB = *MI.getParent();
+  MIRBuilder.setInsertPt(MBB, MBB.getFirstNonPHI());
+
   // Merge small outputs into MI's def.
   if (NumLeftovers) {
     mergeMixedSubvectors(MI.getReg(0), OutputRegs);

[dfszabo]

This issue came up with our downstream backend, but the problematic examples are passing with AArch64. Not easy to create a test for this, but on the other hand the change is fairly self evident.

[AreaZR]

This issue came up with our downstream backend, but the problematic examples are passing with AArch64. Not easy to create a test for this, but on the other hand the change is fairly self evident.

There is a failing unit test. Maybe that can give an idea of how this changes things.

[dfszabo]

This issue came up with our downstream backend, but the problematic examples are passing with AArch64. Not easy to create a test for this, but on the other hand the change is fairly self evident.

There is a failing unit test. Maybe that can give an idea of how this changes things.

Yes indeed, the test fix show now nicely the expected behavior.

[AreaZR]

Ping?

[AreaZR]

/cherry-pick 2347020

[llvmbot]

/cherry-pick 2347020

Error: Command failed due to missing milestone.