AMDGPU: VIMAGE/VSAMPLE are bundle candidates

[ruiling]

No description provided.

[llvmbot]

@llvm/pr-subscribers-backend-amdgpu

Author: Ruiling, Song (ruiling)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/140854.diff

1 Files Affected:

• (modified) llvm/lib/Target/AMDGPU/SIPostRABundler.cpp (+2-1)

diff --git a/llvm/lib/Target/AMDGPU/SIPostRABundler.cpp b/llvm/lib/Target/AMDGPU/SIPostRABundler.cpp
index 87ff4b4d36940..efdc55b8e68be 100644
--- a/llvm/lib/Target/AMDGPU/SIPostRABundler.cpp
+++ b/llvm/lib/Target/AMDGPU/SIPostRABundler.cpp
@@ -64,7 +64,8 @@ class SIPostRABundler {
 
 constexpr uint64_t MemFlags = SIInstrFlags::MTBUF | SIInstrFlags::MUBUF |
                               SIInstrFlags::SMRD | SIInstrFlags::DS |
-                              SIInstrFlags::FLAT | SIInstrFlags::MIMG;
+                              SIInstrFlags::FLAT | SIInstrFlags::MIMG |
+                              SIInstrFlags::VIMAGE | SIInstrFlags::VSAMPLE;
 
 } // End anonymous namespace.
 

[arsenm]

needs test

[jayfoad]

No objection to the patch, but I don't understand why this pass is useful. The original description (from https://reviews.llvm.org/D72737) said that we bundle the clauses formed by the pre-RA scheduler to prevent the post-RA scheduler from breaking them, instead of adding clustering mutations to the post-RA scheduler. But something has changed since then, and we do add clustering mutations to the post-RA scheduler.