[llvm] [AMDGPU] AMDGPULateCodeGenPrepare Legacy PM: replace `setPreservesAll()` with `setPreservesCFG()` (PR #148167)

[Jim M. R. Teichgräber via llvm-commits]

https://github.com/J-MR-T created https://github.com/llvm/llvm-project/pull/148167

<details><summary>This PR depends on #148165; the first commit (90f1d0a881a21a8b4f192622d798c290770fda63) belongs to that PR.</summary>
<p>
The changes are distinct, so separate PRs seemed the best option. I don't have commit access, so I couldn't use user-branches to mark the dependency.
</p>
</details> 

As AMDGPULateCodeGenPrepare actually performs changes that invalidate Uniformity Analysis; use `setPreservesCFG()` to mark this.

Note that before #148165, this would still have preserved Uniformity Analysis, hence the dependency. In addition, `amdgpu/llc-pipeline.cc` needs to be changed when both changes are in effect, but those changes would make the test fail if the PRs weren't based on one another.

<details><summary>Note on why this hasn't caused issues so far</summary>
<p>

It just so happens that AMDGPULateCodeGenPrepare is always immediately followed by AMDGPUUnifyDivergentExitNodes, which *does* invalidate most analyses, including Uniformity. And because UnifyDivergentExitNodes only looks at terminators, and LateCGP seemingly does not replace uniform values with divergent values, or divergent values with uniform values, and it only *inserts new values that are not looked at by UnifyDivergentExitNodes*, this bug remained hidden.

</p>
</details> 

---

I ran `git-clang-format` on my changes. I tested them using the `check-llvm` target; no unexpected failures occurred after I made the change to `amdgpu/llc-pipeline.ll`.


>From 90f1d0a881a21a8b4f192622d798c290770fda63 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Jim=20Teichgr=C3=A4ber?= <Jim.Teichgraber at amd.com>
Date: Fri, 11 Jul 2025 10:16:25 +0200
Subject: [PATCH 1/2] set legacy PM UniformityInfoWrapperPass isCFGOnly to
 false

---
 llvm/lib/Analysis/UniformityAnalysis.cpp | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/llvm/lib/Analysis/UniformityAnalysis.cpp b/llvm/lib/Analysis/UniformityAnalysis.cpp
index 2101fdfacfc8f..0daf4041a72f3 100644
--- a/llvm/lib/Analysis/UniformityAnalysis.cpp
+++ b/llvm/lib/Analysis/UniformityAnalysis.cpp
@@ -150,8 +150,11 @@ INITIALIZE_PASS_BEGIN(UniformityInfoWrapperPass, "uniformity",
 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
 INITIALIZE_PASS_DEPENDENCY(CycleInfoWrapperPass)
 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
+// Though Uniformity Analysis depends on the CFG,
+// it also needs to be invalidated if values are changed, so isCFGOnly: false.
+// See NOTE on updatability at the start of GenericUniformityImpl.h
 INITIALIZE_PASS_END(UniformityInfoWrapperPass, "uniformity",
-                    "Uniformity Analysis", true, true)
+                    "Uniformity Analysis", false, true)
 
 void UniformityInfoWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
   AU.setPreservesAll();

>From 9a8d96a14bf9fd0e7cad0aab7d6e200d3bc652f6 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Jim=20Teichgr=C3=A4ber?= <Jim.Teichgraber at amd.com>
Date: Fri, 11 Jul 2025 10:33:41 +0200
Subject: [PATCH 2/2] [AMDGPU] AMDGPULateCodeGenPrepare: legacy-pm replace
 setPreservesAll() with setPreservesCFG()

---
 llvm/lib/Target/AMDGPU/AMDGPULateCodeGenPrepare.cpp | 4 +++-
 llvm/test/CodeGen/AMDGPU/llc-pipeline.ll            | 4 ++++
 2 files changed, 7 insertions(+), 1 deletion(-)

diff --git a/llvm/lib/Target/AMDGPU/AMDGPULateCodeGenPrepare.cpp b/llvm/lib/Target/AMDGPU/AMDGPULateCodeGenPrepare.cpp
index f0d63f523088b..b109fb73deaa5 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPULateCodeGenPrepare.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPULateCodeGenPrepare.cpp
@@ -546,7 +546,9 @@ class AMDGPULateCodeGenPrepareLegacy : public FunctionPass {
     AU.addRequired<TargetPassConfig>();
     AU.addRequired<AssumptionCacheTracker>();
     AU.addRequired<UniformityInfoWrapperPass>();
-    AU.setPreservesAll();
+    // makes changes that can invalidate Uniformity Analysis,
+    // so don't preserveAll here (see new PM version above)
+    AU.setPreservesCFG();
   }
 
   bool runOnFunction(Function &F) override;
diff --git a/llvm/test/CodeGen/AMDGPU/llc-pipeline.ll b/llvm/test/CodeGen/AMDGPU/llc-pipeline.ll
index dd2ff2e013cc8..b23662c63338c 100644
--- a/llvm/test/CodeGen/AMDGPU/llc-pipeline.ll
+++ b/llvm/test/CodeGen/AMDGPU/llc-pipeline.ll
@@ -255,6 +255,7 @@
 ; GCN-O1-NEXT:        Uniformity Analysis
 ; GCN-O1-NEXT:        AMDGPU IR late optimizations
 ; GCN-O1-NEXT:        Post-Dominator Tree Construction
+; GCN-O1-NEXT:        Uniformity Analysis
 ; GCN-O1-NEXT:        Unify divergent function exit nodes
 ; GCN-O1-NEXT:        Dominator Tree Construction
 ; GCN-O1-NEXT:        Cycle Info Analysis
@@ -557,6 +558,7 @@
 ; GCN-O1-OPTS-NEXT:        Uniformity Analysis
 ; GCN-O1-OPTS-NEXT:        AMDGPU IR late optimizations
 ; GCN-O1-OPTS-NEXT:        Post-Dominator Tree Construction
+; GCN-O1-OPTS-NEXT:        Uniformity Analysis
 ; GCN-O1-OPTS-NEXT:        Unify divergent function exit nodes
 ; GCN-O1-OPTS-NEXT:        Dominator Tree Construction
 ; GCN-O1-OPTS-NEXT:        Cycle Info Analysis
@@ -871,6 +873,7 @@
 ; GCN-O2-NEXT:        Uniformity Analysis
 ; GCN-O2-NEXT:        AMDGPU IR late optimizations
 ; GCN-O2-NEXT:        Post-Dominator Tree Construction
+; GCN-O2-NEXT:        Uniformity Analysis
 ; GCN-O2-NEXT:        Unify divergent function exit nodes
 ; GCN-O2-NEXT:        Dominator Tree Construction
 ; GCN-O2-NEXT:        Cycle Info Analysis
@@ -1200,6 +1203,7 @@
 ; GCN-O3-NEXT:        Uniformity Analysis
 ; GCN-O3-NEXT:        AMDGPU IR late optimizations
 ; GCN-O3-NEXT:        Post-Dominator Tree Construction
+; GCN-O3-NEXT:        Uniformity Analysis
 ; GCN-O3-NEXT:        Unify divergent function exit nodes
 ; GCN-O3-NEXT:        Dominator Tree Construction
 ; GCN-O3-NEXT:        Cycle Info Analysis