Fuse matching product scf.if regions by 1sgtpepper · Pull Request #484 · project-llzk/llzk-lib

1sgtpepper · 2026-05-26T21:23:59Z

Summary

Partially addresses #300.

Adds conservative scf.if fusion to -llzk-fuse-product-loops for matching compute/constrain product regions.

The fusion handles narrow cases where matching compute and constrain scf.if regions have the same SSA condition and can be fused without changing member read/write ordering. It rejects constrain-side member reads, storage writes, calls, witness generation, and write-like memory effects, so member reads remain tied to their original storage semantics until a more precise remapping analysis is added.

Product control-flow fusion remains scoped to explicitly marked product-source control ops.

Testing

git diff --check
Added focused lit/FileCheck coverage under test/Transforms/FuseProductLoops/ for:
- fusable scf.if cases, including direct compute-result uses and multi-result remapping
- condition mismatch, member-read, read-after-write, and call-after-write no-fuse cases
- moved constrain-branch rejection for array writes/inserts, witness calls, member writes, global writes, POD writes, and RAM stores
- untagged outer scf.if/scf.for product-source inference no-fuse cases

iangneal · 2026-05-29T12:38:10Z

@codex review

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: b4213c67d7

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

raghav198

The constrain use is remapped to the branch-local compute value instead of hoisting the read.

I may be misunderstanding, but I think this is not safe when the struct.readm is a constrain op:

%0 = felt.add %a, %b {source = "compute"}
struct.writem %self[@foo] = %0 {source = "compute"}
%foo = struct.readm %self[@foo] {source = "constrain"}
constrain.eq %foo, %c {source = "constrain"}

emits a constraint guaranteeing that the @foo signal equals the result of %c, whereas in the remapped

%0 = felt.add %a, %b {source = "compute"}
struct.writem %self[@foo] = %0 {source = "compute"}
constrain.eq %0, %c {source = "constrain"}

the emitted constraint doesn't refer to a struct signal at all.

1sgtpepper · 2026-05-29T20:07:08Z

Thanks for catching this, Raghav. You're right. Remapping the struct.readm result to the branch local compute value can remove the member signal from the emitted constraint. I'll make this fusion case conservative for now and add a regression test so we keep the constraint tied to the member read.

Copilot

Pull request overview

This PR extends the -llzk-fuse-product-loops pass to also fuse matching compute/constrain scf.if regions in struct product programs (Fixes #300), and adds regression tests covering fusable and non-fusable scf.if patterns.

Changes:

Add scf.if fusion logic that clones compute + constrain branches into a single fused scf.if when conditions and structural constraints match.
Refactor pass entry/recursion to fuse both scf.if and scf.for control flow via a shared fuseMatchingRegionControlFlow routine.
Add lit/FileCheck tests for scf.if fusion and “do not fuse” guardrails; add an unreleased changelog entry.

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
lib/Transforms/LLZKFuseProductLoopsPass.cpp	Implements `scf.if` fusion and refactors recursion over region control flow.
test/Transforms/FuseProductLoops/fuse_if_no_member_read.llzk	Positive test ensuring `scf.if` fusion occurs in a safe case.
test/Transforms/FuseProductLoops/fuse_loop_then_if.llzk	Ensures loop fusion still works and enables nested `scf.if` fusion after loop fusion.
test/Transforms/FuseProductLoops/no_fuse_if_call_after_write.llzk	Negative test: blocks fusion when a call occurs after a mapped write.
test/Transforms/FuseProductLoops/no_fuse_if_condition_mismatch.llzk	Negative test: blocks fusion when conditions differ.
test/Transforms/FuseProductLoops/no_fuse_if_internal_read.llzk	Negative test: blocks fusion when constrain side reads members inside the `scf.if`.
test/Transforms/FuseProductLoops/no_fuse_if_read_after_write.llzk	Negative test: blocks fusion when a member read appears between the ifs.
changelogs/unreleased/fuse-product-if-regions.yaml	Changelog entry for the new `scf.if` fusion capability.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

iangneal · 2026-06-25T19:46:15Z

@codex review

chatgpt-codex-connector · 2026-06-25T19:52:53Z

Codex Review: Didn't find any major issues. Can't wait for the next one!

Reviewed commit: 326acc24af

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

1sgtpepper · 2026-06-26T13:56:10Z

I seem to have found some loose ends, will update in a bit.

1sgtpepper · 2026-06-26T14:36:26Z

Pushed a conservative safety guard for moved constrain branches.

iangneal · 2026-07-14T18:41:42Z

@codex review

chatgpt-codex-connector · 2026-07-14T18:49:00Z

Codex Review: Didn't find any major issues. Hooray!

Reviewed commit: 8725011bd2

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

iangneal · 2026-07-14T18:59:15Z

+  return effectOp.hasEffect<MemoryEffects::Write>() ||
+         effectOp.hasEffect<MemoryEffects::Allocate>() || effectOp.hasEffect<MemoryEffects::Free>();


I believe you can either do:
effectOp.hasEffect<MemoryEffects::Write, MemoryEffects::Allocate, MemoeryEffects::Free>(); or at least hasEffect<MemoryEffects::Write, MemoryEffects::Allocate, MemoeryEffects::Free>(effectOp);

iangneal · 2026-07-15T17:48:44Z

+static std::optional<llvm::StringRef> getProductSource(Operation *op) {
+  if (StringAttr source = op->getAttrOfType<StringAttr>(PRODUCT_SOURCE)) {
+    return source.getValue();
+  }
+  return std::nullopt;
+}
+
+static inline bool hasProductSource(Operation *op, llvm::StringRef source) {
+  std::optional<llvm::StringRef> productSource = getProductSource(op);
+  return productSource && *productSource == source;
+}


I think we should move these functions to a shared helper file so they can be reused, e.g. by llzk-compute-constrain-to-product-pass

iangneal · 2026-07-15T17:52:03Z

+         effectOp.hasEffect<MemoryEffects::Allocate>() || effectOp.hasEffect<MemoryEffects::Free>();
+}
+
+static bool hasUnsafeMovedConstrainOp(scf::IfOp constrainIf) {


Add documentation to functions, especially here, since it's not clear what you're trying to do.

iangneal · 2026-07-15T17:52:14Z

+    if (isa<MemberReadOp, MemberWriteOp, llzk::global::GlobalWriteOp, llzk::array::WriteArrayOp,
+            llzk::array::InsertArrayOp, llzk::pod::WritePodOp, llzk::ram::StoreOp>(op)) {
+      return WalkResult::interrupt();
+    }
+
+    if (op->hasTrait<llzk::function::WitnessGen>() || hasWriteLikeMemoryEffect(op)) {
+      return WalkResult::interrupt();
+    }
+
+    if (isa<CallOpInterface>(op)) {
+      return WalkResult::interrupt();
+    }


This could all be one if statement.

iangneal · 2026-07-15T17:53:52Z

+) {
+  // `scf.if` fusion moves the constrain branch before intervening member writes of compute-if
+  // results. The moved branch must not read members, write storage, call functions, or have
+  // write-like effects.


Currently this loop only checks member reads/writes, but ignores RAM/globals

iangneal · 2026-07-15T17:56:56Z

+// CHECK:           struct.writem %{{[0-9a-zA-Z_\.]+}}[@foo] = %[[RESULT]] : <@A>, !felt.type
+// CHECK:           %[[FOO:[0-9a-zA-Z_\.]+]] = struct.readm %{{[0-9a-zA-Z_\.]+}}[@foo] : <@A>, !felt.type
+// CHECK:           scf.if %[[B]] {
+// CHECK:             constrain.eq %[[FOO]]


Lit test output should show the whole module, like all other lit tests in this repo.

iangneal requested a review from a team May 28, 2026 01:26

chatgpt-codex-connector Bot reviewed May 29, 2026

View reviewed changes

Comment thread lib/Transforms/LLZKFuseProductLoopsPass.cpp Outdated

raghav198 reviewed May 29, 2026

View reviewed changes

tim-hoffman requested review from Copilot and raghav198 June 1, 2026 14:59

Copilot started reviewing on behalf of tim-hoffman June 1, 2026 14:59 View session

Copilot AI reviewed Jun 1, 2026

View reviewed changes

Comment thread lib/Transforms/LLZKFuseProductLoopsPass.cpp Outdated

Comment thread lib/Transforms/LLZKFuseProductLoopsPass.cpp

1sgtpepper mentioned this pull request Jun 16, 2026

Implement compute/constrain fusion for scf.if ops in the product program #300

Open

tim-hoffman reviewed Jul 2, 2026

View reviewed changes

Comment thread CHANGELOG.md Outdated

iangneal requested changes Jul 15, 2026

View reviewed changes

Fuse matching product scf.if regions

6719d38

1sgtpepper force-pushed the fuse-product-if-regions branch from 53ba923 to 6719d38 Compare July 16, 2026 14:11

		return effectOp.hasEffect<MemoryEffects::Write>() \|\|
		effectOp.hasEffect<MemoryEffects::Allocate>() \|\| effectOp.hasEffect<MemoryEffects::Free>();

Uh oh!

Conversation

1sgtpepper commented May 26, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Testing

Uh oh!

iangneal commented May 29, 2026

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

raghav198 left a comment

Choose a reason for hiding this comment

Uh oh!

1sgtpepper commented May 29, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

Uh oh!

iangneal commented Jun 25, 2026

Uh oh!

chatgpt-codex-connector Bot commented Jun 25, 2026

Uh oh!

1sgtpepper commented Jun 26, 2026

Uh oh!

1sgtpepper commented Jun 26, 2026

Uh oh!

Uh oh!

iangneal commented Jul 14, 2026

Uh oh!

chatgpt-codex-connector Bot commented Jul 14, 2026

Uh oh!

Uh oh!

iangneal Jul 14, 2026

Choose a reason for hiding this comment

Uh oh!

iangneal Jul 15, 2026

Choose a reason for hiding this comment

Uh oh!

iangneal Jul 15, 2026

Choose a reason for hiding this comment

Uh oh!

iangneal Jul 15, 2026

Choose a reason for hiding this comment

Uh oh!

iangneal Jul 15, 2026

Choose a reason for hiding this comment

Uh oh!

iangneal Jul 15, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

1sgtpepper commented May 26, 2026 •

edited

Loading