Skip to content

[SSAUpdater] Avoid scanning basic blocks to find instruction order. #123803

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 3 commits into from
Jan 22, 2025

Conversation

jcranmer-intel
Copy link
Contributor

This fixes a compile-time regression caused by #116645, where an entry basic block with a very large number of allocas and other instructions caused SROA to take ~100× its expected runtime, as every alloca (with ~2 uses) now calls this method to find the order of those few instructions, rescanning the very large basic block every single time.

Since this code was originally written, Instructions now have ordering numbers available to determine relative order without unnecessarily scanning the basic block.

This fixes a compile-time regression caused by llvm#116645, where an entry
basic block with a very large number of allocas and other instructions
caused SROA to take ~100× its expected runtime, as every alloca (with ~2
uses) now calls this method to find the order of those few instructions,
rescanning the very large basic block every single time.

Since this code was originally written, Instructions now have ordering
numbers available to determine relative order without unnecessarily
scanning the basic block.
@llvmbot
Copy link
Member

llvmbot commented Jan 21, 2025

@llvm/pr-subscribers-llvm-transforms

Author: Joshua Cranmer (jcranmer-intel)

Changes

This fixes a compile-time regression caused by #116645, where an entry basic block with a very large number of allocas and other instructions caused SROA to take ~100× its expected runtime, as every alloca (with ~2 uses) now calls this method to find the order of those few instructions, rescanning the very large basic block every single time.

Since this code was originally written, Instructions now have ordering numbers available to determine relative order without unnecessarily scanning the basic block.


Full diff: https://github.com/llvm/llvm-project/pull/123803.diff

1 Files Affected:

  • (modified) llvm/lib/Transforms/Utils/SSAUpdater.cpp (+15-10)
diff --git a/llvm/lib/Transforms/Utils/SSAUpdater.cpp b/llvm/lib/Transforms/Utils/SSAUpdater.cpp
index 4bf4acd6330f58..79d46a88cb454b 100644
--- a/llvm/lib/Transforms/Utils/SSAUpdater.cpp
+++ b/llvm/lib/Transforms/Utils/SSAUpdater.cpp
@@ -432,9 +432,7 @@ void LoadAndStorePromoter::run(const SmallVectorImpl<Instruction *> &Insts) {
       }
     }
 
-    // If so, we can queue them all as live in loads.  We don't have an
-    // efficient way to tell which on is first in the block and don't want to
-    // scan large blocks, so just add all loads as live ins.
+    // If so, we can queue them all as live in loads.
     if (!HasStore) {
       for (Instruction *I : BlockUses)
         LiveInLoads.push_back(cast<LoadInst>(I));
@@ -442,16 +440,22 @@ void LoadAndStorePromoter::run(const SmallVectorImpl<Instruction *> &Insts) {
       continue;
     }
 
+    // Sort all of the interesting instructions in the block so that we don't
+    // have to scan a large block just to find a few instructions.
+    std::sort(BlockUses.begin(), BlockUses.end(),
+              [](Instruction *A, Instruction *B) { return A->comesBefore(B); });
+
     // Otherwise, we have mixed loads and stores (or just a bunch of stores).
     // Since SSAUpdater is purely for cross-block values, we need to determine
     // the order of these instructions in the block.  If the first use in the
     // block is a load, then it uses the live in value.  The last store defines
-    // the live out value.  We handle this by doing a linear scan of the block.
+    // the live out value.
     Value *StoredValue = nullptr;
-    for (Instruction &I : *BB) {
-      if (LoadInst *L = dyn_cast<LoadInst>(&I)) {
+    for (Instruction *I : BlockUses) {
+      if (LoadInst *L = dyn_cast<LoadInst>(I)) {
         // If this is a load from an unrelated pointer, ignore it.
-        if (!isInstInList(L, Insts)) continue;
+        if (!isInstInList(L, Insts))
+          continue;
 
         // If we haven't seen a store yet, this is a live in use, otherwise
         // use the stored value.
@@ -465,14 +469,15 @@ void LoadAndStorePromoter::run(const SmallVectorImpl<Instruction *> &Insts) {
         continue;
       }
 
-      if (StoreInst *SI = dyn_cast<StoreInst>(&I)) {
+      if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
         // If this is a store to an unrelated pointer, ignore it.
-        if (!isInstInList(SI, Insts)) continue;
+        if (!isInstInList(SI, Insts))
+          continue;
         updateDebugInfo(SI);
 
         // Remember that this is the active value in the block.
         StoredValue = SI->getOperand(0);
-      } else if (auto *AI = dyn_cast<AllocaInst>(&I)) {
+      } else if (auto *AI = dyn_cast<AllocaInst>(I)) {
         // Check if this an alloca, in which case we treat it as a store of
         // getValueToUseForAlloca.
         if (!isInstInList(AI, Insts))

@davemgreen davemgreen requested a review from nikic January 21, 2025 20:10
@davemgreen
Copy link
Collaborator

Sounds OK to me.

// If this is a load from an unrelated pointer, ignore it.
if (!isInstInList(L, Insts)) continue;
if (!isInstInList(L, Insts))
continue;
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can't we also remove the isInstInList checks now?

Comment on lines 445 to 446
std::sort(BlockUses.begin(), BlockUses.end(),
[](Instruction *A, Instruction *B) { return A->comesBefore(B); });
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
std::sort(BlockUses.begin(), BlockUses.end(),
[](Instruction *A, Instruction *B) { return A->comesBefore(B); });
llvm::sort(BlockUses,
[](Instruction *A, Instruction *B) { return A->comesBefore(B); });

Copy link
Contributor

@nikic nikic left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM

@jcranmer-intel jcranmer-intel merged commit b0d35cf into llvm:main Jan 22, 2025
8 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
Development

Successfully merging this pull request may close these issues.

4 participants