[CodeGen][NPM] Stitch up loop passes in codegen pipeline

[optimisan]

Adding in some missing passes in NPM path.

[optimisan]

• [CodeGen][NPM] Stitch up loop passes in codegen pipeline #133050 👈 (View in Graphite)
• users/optimisan/04-02-_amdgpu_npm_add_isrequired_to_passes_missing_it

This stack of pull requests is managed by Graphite. Learn more about stacking.

[llvmbot]

@llvm/pr-subscribers-backend-amdgpu

Author: Akshat Oke (optimisan)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/133050.diff

3 Files Affected:

• (modified) llvm/include/llvm/Passes/CodeGenPassBuilder.h (+14-2)
• (modified) llvm/include/llvm/Target/CGPassBuilderOption.h (+2)
• (modified) llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp (+7-1)

diff --git a/llvm/include/llvm/Passes/CodeGenPassBuilder.h b/llvm/include/llvm/Passes/CodeGenPassBuilder.h
index 61e377de0c424..a1be22cda96c8 100644
--- a/llvm/include/llvm/Passes/CodeGenPassBuilder.h
+++ b/llvm/include/llvm/Passes/CodeGenPassBuilder.h
@@ -107,13 +107,16 @@
 #include "llvm/Target/CGPassBuilderOption.h"
 #include "llvm/Target/TargetMachine.h"
 #include "llvm/Transforms/CFGuard.h"
+#include "llvm/Transforms/ObjCARC.h"
 #include "llvm/Transforms/Scalar/ConstantHoisting.h"
 #include "llvm/Transforms/Scalar/LoopPassManager.h"
 #include "llvm/Transforms/Scalar/LoopStrengthReduce.h"
+#include "llvm/Transforms/Scalar/LoopTermFold.h"
 #include "llvm/Transforms/Scalar/LowerConstantIntrinsics.h"
 #include "llvm/Transforms/Scalar/MergeICmps.h"
 #include "llvm/Transforms/Scalar/PartiallyInlineLibCalls.h"
 #include "llvm/Transforms/Scalar/ScalarizeMaskedMemIntrin.h"
+#include "llvm/Transforms/Utils/CanonicalizeFreezeInLoops.h"
 #include "llvm/Transforms/Utils/EntryExitInstrumenter.h"
 #include "llvm/Transforms/Utils/LowerInvoke.h"
 #include <cassert>
@@ -690,7 +693,12 @@ void CodeGenPassBuilder<Derived, TargetMachineT>::addIRPasses(
 
   // Run loop strength reduction before anything else.
   if (getOptLevel() != CodeGenOptLevel::None && !Opt.DisableLSR) {
-    addPass(createFunctionToLoopPassAdaptor(LoopStrengthReducePass(),
+    LoopPassManager LPM;
+    LPM.addPass(CanonicalizeFreezeInLoopsPass());
+    LPM.addPass(LoopStrengthReducePass());
+    if (Opt.EnableLoopTermFold)
+      LPM.addPass(LoopTermFoldPass());
+    addPass(createFunctionToLoopPassAdaptor(std::move(LPM),
                                             /*UseMemorySSA=*/true));
   }
 
@@ -735,7 +743,8 @@ void CodeGenPassBuilder<Derived, TargetMachineT>::addIRPasses(
   addPass(ScalarizeMaskedMemIntrinPass());
 
   // Expand reduction intrinsics into shuffle sequences if the target wants to.
-  addPass(ExpandReductionsPass());
+  if (!Opt.DisableExpandReductions)
+    addPass(ExpandReductionsPass());
 
   // Convert conditional moves to conditional jumps when profitable.
   if (getOptLevel() != CodeGenOptLevel::None && !Opt.DisableSelectOptimize)
@@ -810,6 +819,9 @@ void CodeGenPassBuilder<Derived, TargetMachineT>::addISelPrepare(
     AddIRPass &addPass) const {
   derived().addPreISel(addPass);
 
+  if (getOptLevel() != CodeGenOptLevel::None)
+    addPass(ObjCARCContractPass());
+
   addPass(CallBrPreparePass());
   // Add both the safe stack and the stack protection passes: each of them will
   // only protect functions that have corresponding attributes.
diff --git a/llvm/include/llvm/Target/CGPassBuilderOption.h b/llvm/include/llvm/Target/CGPassBuilderOption.h
index 51f25c1360b87..00d8a99c3984c 100644
--- a/llvm/include/llvm/Target/CGPassBuilderOption.h
+++ b/llvm/include/llvm/Target/CGPassBuilderOption.h
@@ -51,6 +51,8 @@ struct CGPassBuilderOption {
   bool EnableMachineFunctionSplitter = false;
   bool EnableSinkAndFold = false;
   bool EnableTailMerge = true;
+  /// Enable LoopTermFold immediately after LSR
+  bool EnableLoopTermFold = false;
   bool MISchedPostRA = false;
   bool EarlyLiveIntervals = false;
   bool GCEmptyBlocks = false;
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp b/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
index b6cc5137d711a..84efff6fe79be 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
@@ -101,7 +101,9 @@
 #include "llvm/Transforms/Scalar/FlattenCFG.h"
 #include "llvm/Transforms/Scalar/GVN.h"
 #include "llvm/Transforms/Scalar/InferAddressSpaces.h"
+#include "llvm/Transforms/Scalar/LICM.h"
 #include "llvm/Transforms/Scalar/LoopDataPrefetch.h"
+#include "llvm/Transforms/Scalar/LoopPassManager.h"
 #include "llvm/Transforms/Scalar/NaryReassociate.h"
 #include "llvm/Transforms/Scalar/SeparateConstOffsetFromGEP.h"
 #include "llvm/Transforms/Scalar/Sink.h"
@@ -2021,7 +2023,11 @@ void AMDGPUCodeGenPassBuilder::addIRPasses(AddIRPass &addPass) const {
     // TODO: May want to move later or split into an early and late one.
     addPass(AMDGPUCodeGenPreparePass(TM));
 
-    // TODO: LICM
+    // Try to hoist loop invariant parts of divisions AMDGPUCodeGenPrepare may
+    // have expanded.
+    if (TM.getOptLevel() > CodeGenOptLevel::Less)
+      addPass(createFunctionToLoopPassAdaptor(LICMPass(LICMOptions()),
+                                              /*UseMemorySSA=*/true));
   }
 
   Base::addIRPasses(addPass);

[optimisan]

Stack updated manually:

• [CodeGen][NPM] Stitch up loop passes in codegen pipeline #133050 👈
• [AMDGPU][NPM] Add isRequired to passes missing it #134033
• [CodeGen][NPM] Differentiate pipeline-required and opt-required passes #135752
• main

[cdevadas]

LGTM except for the comment.

[cdevadas]

Terminate the comment with a period.