[AMDGPU][True16] fix a bug in codeGen causing e64 with wrong vgpr type to shrink

[broxigarchen]

This bug is introduced in #102198

The previous path change to use realTrue16 flag, however, we have some t16 instructions that are implemented with fake16, and has Lo128 registers types. Thus we should still using hasTrue16Bit flag for shrinking check

[llvmbot]

@llvm/pr-subscribers-backend-amdgpu

Author: Brox Chen (broxigarchen)

Changes

This bug is introduced in #102198

---

Full diff: https://github.com/llvm/llvm-project/pull/102942.diff

1 Files Affected:

• (modified) llvm/lib/Target/AMDGPU/SIShrinkInstructions.cpp (+1-1)

diff --git a/llvm/lib/Target/AMDGPU/SIShrinkInstructions.cpp b/llvm/lib/Target/AMDGPU/SIShrinkInstructions.cpp
index 155747551471e..5d38cafd73dd9 100644
--- a/llvm/lib/Target/AMDGPU/SIShrinkInstructions.cpp
+++ b/llvm/lib/Target/AMDGPU/SIShrinkInstructions.cpp
@@ -1048,7 +1048,7 @@ bool SIShrinkInstructions::runOnMachineFunction(MachineFunction &MF) {
               MachineFunctionProperties::Property::NoVRegs))
         continue;
 
-      if (ST->useRealTrue16Insts() && AMDGPU::isTrue16Inst(MI.getOpcode()) &&
+      if (ST->hasTrue16BitInsts() && AMDGPU::isTrue16Inst(MI.getOpcode()) &&
           !shouldShrinkTrue16(MI))
         continue;
 

[Sisyph]

LGTM. We can revert this part of the patch until all instructions are properly updated.

[jayfoad]

Can you include a test case?

[hanhanW]

Thanks, I verified that it fixes the issue in our project!

[arsenm]

Needs test

[broxigarchen]

Needs test

Added a small test to verify and to prevent furture failure. Should be expanded when more true16 and fake16 instructions are supported