[llvm] [AMDGPU] Add an option to run `AMDGPUAttributorPass` in full LTO (PR #102086)

[Shilei Tian via llvm-commits]

https://github.com/shiltian updated https://github.com/llvm/llvm-project/pull/102086

>From ec12f76d5b50a038cd0add2cfb24f4c69c90d023 Mon Sep 17 00:00:00 2001
From: Shilei Tian <i at tianshilei.me>
Date: Mon, 5 Aug 2024 20:27:47 -0400
Subject: [PATCH] [AMDGPU] Add an option to run `AMDGPUAttributorPass` in full
 LTO

Currently `AMDGPUAttributorPass` is registered in default optimizer pipeline.
This will allow the pass to run in default pipeline as well as at thinLTO post
link stage. However, it will not run in full LTO post link stage. This patch
adds an option to run it in full LTO. It is set to false by default.
---
 llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp | 15 +++++++--------
 1 file changed, 7 insertions(+), 8 deletions(-)

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp b/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
index c8fb68d1c0b0c..a5a5bf184e9ef 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
@@ -301,6 +301,11 @@ static cl::opt<bool> EnableSIModeRegisterPass(
   cl::init(true),
   cl::Hidden);
 
+static cl::opt<bool> EnableAMDGPUAttributorInFullLTO(
+    "amdgpu-attributor-enable-full-lto",
+    cl::desc("Enable AMDGPUAttributorPass in full LTO post link stage"),
+    cl::init(false), cl::Hidden);
+
 // Enable GFX11.5+ s_singleuse_vdst insertion
 static cl::opt<bool>
     EnableInsertSingleUseVDST("amdgpu-enable-single-use-vdst",
@@ -734,14 +739,6 @@ void AMDGPUTargetMachine::registerPassBuilderCallbacks(PassBuilder &PB) {
         PM.addPass(createCGSCCToFunctionPassAdaptor(std::move(FPM)));
       });
 
-  // FIXME: Why is AMDGPUAttributor not in CGSCC?
-  PB.registerOptimizerLastEPCallback(
-      [this](ModulePassManager &MPM, OptimizationLevel Level) {
-        if (Level != OptimizationLevel::O0) {
-          MPM.addPass(AMDGPUAttributorPass(*this));
-        }
-      });
-
   PB.registerFullLinkTimeOptimizationLastEPCallback(
       [this](ModulePassManager &PM, OptimizationLevel Level) {
         // We want to support the -lto-partitions=N option as "best effort".
@@ -749,6 +746,8 @@ void AMDGPUTargetMachine::registerPassBuilderCallbacks(PassBuilder &PB) {
         // module is partitioned for codegen.
         if (EnableLowerModuleLDS)
           PM.addPass(AMDGPULowerModuleLDSPass(*this));
+
+        PM.addPass(AMDGPUAttributorPass(*this));
       });
 
   PB.registerRegClassFilterParsingCallback(