8314997: Missing optimization opportunities due to missing try_clean_mem_phi() calls #15445

chhagedorn · 2023-08-28T12:38:43Z

While working on a Valhalla bug, I've noticed that we sometimes miss RegionNode::try_clean_mem_phi() calls to remove a useless diamond

     If
True    False
   Region

with only a single memory phi. This blocks further optimizations like converting a loop into a counted one. The code in Valhalla looks slightly different but the problem is also reproducible in mainline.

Problem

In the test case, a region is transformed in IGVN such that it merges a diamond without any dependencies on both paths. The region has two phis. One of them is a memory phi which could be transformed by RegionNode::try_clean_mem_phi(). But when processing the region with its two phis in IGVN, we do not optimize the memory phi away because has_unique_phi() is false and we bail out:

jdk/src/hotspot/share/opto/cfgnode.cpp

Lines 450 to 471 in 725ec0c

    
           bool RegionNode::try_clean_mem_phi(PhaseGVN *phase) { 
        
             // Incremental inlining + PhaseStringOpts sometimes produce: 
        
             // 
        
             // cmpP with 1 top input 
        
             //           | 
        
             //          If 
        
             //         /  \ 
        
             //   IfFalse  IfTrue  /- Some Node 
        
             //         \  /      /    / 
        
             //        Region    / /-MergeMem 
        
             //             \---Phi 
        
             // 
        
             // 
        
             // It's expected by PhaseStringOpts that the Region goes away and is 
        
             // replaced by If's control input but because there's still a Phi, 
        
             // the Region stays in the graph. The top input from the cmpP is 
        
             // propagated forward and a subgraph that is useful goes away. The 
        
             // code below replaces the Phi with the MergeMem so that the Region 
        
             // is simplified. 
        
             PhiNode* phi = has_unique_phi(); 
        
             if (phi && phi->type() == Type::MEMORY && req() == 3 && phi->is_diamond_phi(true)) {

Later in IGVN, the second phi dies and we only have the single memory phi left. But the region will not be added to the IGVN worklist again to re-apply try_clean_mem_phi(). We therefore miss the removal of the diamond and we fail to apply further optimizations. In the test case, we fail to convert the loop into a counted loop.

Proposed Fix

The fix I propose is to try to apply try_clean_mem_phi() whenever a region is merging a diamond with the assumption that the transformation of a memory phi does not hurt when being applied without being able to remove the region with the diamond (because there are other phis left that cannot be removed). Another option would be to re-add the region to the IGVN worklist when the second last phi dies. But the first approach seems simpler and less invasive.

I've also applied some clean-ups and added an IR test.

Thanks,
Christian

Progress

Change must be properly reviewed (1 review required, with at least 1 Reviewer)
Change must not contain extraneous whitespace
Commit message must refer to an issue

Issue

JDK-8314997: Missing optimization opportunities due to missing try_clean_mem_phi() calls (Enhancement - P4)

Reviewers

Roland Westrelin (@rwestrel - Reviewer) ⚠️ Review applies to 8734fefa
Vladimir Kozlov (@vnkozlov - Reviewer)
Tobias Hartmann (@TobiHartmann - Reviewer)

Reviewing

Using git

Checkout this PR locally:
$ git fetch https://git.openjdk.org/jdk.git pull/15445/head:pull/15445
$ git checkout pull/15445

Update a local copy of the PR:
$ git checkout pull/15445
$ git pull https://git.openjdk.org/jdk.git pull/15445/head

Using Skara CLI tools

Checkout this PR locally:
$ git pr checkout 15445

View PR using the GUI difftool:
$ git pr show -t 15445

Using diff file

Download this PR as a diff file:
https://git.openjdk.org/jdk/pull/15445.diff

Webrev

Link to Webrev Comment

bridgekeeper · 2023-08-28T12:39:49Z

👋 Welcome back chagedorn! A progress list of the required criteria for merging this PR into master will be added to the body of your pull request. There are additional pull request commands available for use with this pull request.

openjdk · 2023-08-28T12:41:12Z

@chhagedorn The following label will be automatically applied to this pull request:

hotspot-compiler

When this pull request is ready to be reviewed, an "RFR" email will be sent to the corresponding mailing list. If you would like to change these labels, use the /label pull request command.

mlbridge · 2023-08-28T12:46:38Z

Webrevs

01: Full - Incremental (1788e3b4)
00: Full (8734fefa)

rwestrel

Looks reasonable to me.

openjdk · 2023-08-28T15:26:22Z

@chhagedorn This change now passes all automated pre-integration checks.

ℹ️ This project also has non-automated pre-integration requirements. Please see the file CONTRIBUTING.md for details.

After integration, the commit message for the final commit will be:

8314997: Missing optimization opportunities due to missing try_clean_mem_phi() calls

Reviewed-by: roland, kvn, thartmann

You can use pull request commands such as /summary, /contributor and /issue to adjust it as needed.

At the time when this comment was updated there had been 93 new commits pushed to the master branch:

764f65c: 8315278: Patch 'print-targets' target to print targets separated by new line
7754ac0: 8314925: ChoiceFormat does not specify IllegalArgumentExceptions
2264667: 8315422: getSoTimeout() would be in try block in SSLSocketImpl
d5973bd: 6333341: [BI] Doc: java.text.BreakIterator class specification is unclear
63f561f: 8306882: (fs) Path.toRealPath(LinkOption.NOFOLLOW_LINKS) fails when "../../" follows a link
2436fb0: 8312306: Add more Reference.reachabilityFence() calls to the security classes using Cleaner
351c31e: 8315378: [BACKOUT] runtime/NMT/SummarySanityCheck.java failed with "Total committed (MMMMMM) did not match the summarized committed (NNNNNN)"
c12ca88: 8312521: Unused field LocaleProviderAdapter#defaultLocaleProviderAdapter could be removed
b38bcae: 8313656: assert(!JvmtiExport::can_support_virtual_threads()) with -XX:-DoJVMTIVirtualThreadTransitions
c8acab1: 8315413: Remove special filtering of Continuation.yield0 in StackWalker
... and 83 more: https://git.openjdk.org/jdk/compare/de0e46c2f6dfa817bfa745195573810d250640de...master

As there are no conflicts, your changes will automatically be rebased on top of these commits when integrating. If you prefer to avoid this automatic rebasing, please check the documentation for the /integrate command for further details.

➡️ To integrate this PR with the above commit message to the master branch, type /integrate in a new comment.

chhagedorn · 2023-08-29T09:11:04Z

Thanks Roland for your review!

vnkozlov

Looks good. Only few small comments.

vnkozlov · 2023-08-29T16:43:44Z

src/hotspot/share/opto/cfgnode.cpp

+  if (left_path == nullptr || right_path == nullptr) {
+    return false;
+  }


So the TOP input will fail next check. May be add comment about that.

I've added a comment at the next if.

vnkozlov · 2023-08-29T16:50:44Z

src/hotspot/share/opto/cfgnode.cpp

+  if (_type != Type::MEMORY) {
+    return false;
+  }
+  assert(is_diamond_phi(), "sanity");


Add explicit check > 0

Good point, added.

chhagedorn · 2023-08-30T06:33:10Z

Thanks Vladimir for your review! I've pushed an update.

vnkozlov

Good.

TobiHartmann

Nice analysis. The fix looks good to me!

chhagedorn · 2023-09-01T07:43:01Z

Thanks Roland, Vladimir, and Tobias for your reviews!

chhagedorn · 2023-09-04T06:51:43Z

/integrate

openjdk · 2023-09-04T06:53:21Z

Going to push as commit 2dc930d.
Since your change was applied there have been 109 commits pushed to the master branch:

ab12c5d: 8315549: CITime misreports code/total nmethod sizes
ba8d5c2: 8315582: Exclude compiler/codecache/CodeCacheFullCountTest.java with Xcomp
75d4ac2: 8311775: [TEST] duplicate verifyHeapDump in several tests
4f90aba: 8311585: Add JRadioButtonMenuItem to bug8031573.java
2a11bc4: 8314837: 5 compiled/codecache tests ignore VM flags
2f7c65e: 8303427: Fixpath confused if unix root contains "/jdk"
e9e0c56: 8314319: LogCompilation doesn't reset lateInlining when it encounters a failure.
56b8db1: 8258970: Disabled JPasswordField foreground color is wrong with GTK LAF
0d4cadb: 8315195: RISC-V: Update hwprobe query for new extensions
b4f7069: 8315446: G1: Remove unused G1AllocRegion::attempt_allocation
... and 99 more: https://git.openjdk.org/jdk/compare/de0e46c2f6dfa817bfa745195573810d250640de...master

Your commit was automatically rebased without conflicts.

openjdk · 2023-09-04T06:53:31Z

@chhagedorn Pushed as commit 2dc930d.

💡 You may see a message that your pull request was closed with unmerged commits. This can be safely ignored.

8314997: Missing optimizations due to missing try_clean_mem_phi() calls

8734fef

chhagedorn changed the title ~~8314997: Missing optimizations due to missing try_clean_mem_phi() calls~~ 8314997: Missing optimization opportunities due to missing try_clean_mem_phi() calls Aug 28, 2023

openjdk bot added the rfr Pull request is ready for review label Aug 28, 2023

openjdk bot added the hotspot-compiler hotspot-compiler-dev@openjdk.org label Aug 28, 2023

rwestrel approved these changes Aug 28, 2023

View reviewed changes

openjdk bot added the ready Pull request is ready to be integrated label Aug 28, 2023

vnkozlov approved these changes Aug 29, 2023

View reviewed changes

Vladimir's review

1788e3b

vnkozlov approved these changes Aug 30, 2023

View reviewed changes

TobiHartmann approved these changes Sep 1, 2023

View reviewed changes

openjdk bot added the integrated Pull request has been integrated label Sep 4, 2023

openjdk bot closed this Sep 4, 2023

openjdk bot removed ready Pull request is ready to be integrated rfr Pull request is ready for review labels Sep 4, 2023

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

8314997: Missing optimization opportunities due to missing try_clean_mem_phi() calls #15445

8314997: Missing optimization opportunities due to missing try_clean_mem_phi() calls #15445

chhagedorn commented Aug 28, 2023 •

edited by openjdk bot

Loading

bridgekeeper bot commented Aug 28, 2023

openjdk bot commented Aug 28, 2023

mlbridge bot commented Aug 28, 2023 •

edited

Loading

rwestrel left a comment

openjdk bot commented Aug 28, 2023 •

edited

Loading

chhagedorn commented Aug 29, 2023

vnkozlov left a comment

vnkozlov Aug 29, 2023

chhagedorn Aug 30, 2023

vnkozlov Aug 29, 2023

chhagedorn Aug 30, 2023

chhagedorn commented Aug 30, 2023

vnkozlov left a comment

TobiHartmann left a comment

chhagedorn commented Sep 1, 2023

chhagedorn commented Sep 4, 2023

openjdk bot commented Sep 4, 2023

openjdk bot commented Sep 4, 2023

	bool RegionNode::try_clean_mem_phi(PhaseGVN *phase) {
	// Incremental inlining + PhaseStringOpts sometimes produce:
	//
	// cmpP with 1 top input
	// \|
	// If
	// / \
	// IfFalse IfTrue /- Some Node
	// \ / / /
	// Region / /-MergeMem
	// \---Phi
	//
	//
	// It's expected by PhaseStringOpts that the Region goes away and is
	// replaced by If's control input but because there's still a Phi,
	// the Region stays in the graph. The top input from the cmpP is
	// propagated forward and a subgraph that is useful goes away. The
	// code below replaces the Phi with the MergeMem so that the Region
	// is simplified.

	PhiNode* phi = has_unique_phi();
	if (phi && phi->type() == Type::MEMORY && req() == 3 && phi->is_diamond_phi(true)) {

8314997: Missing optimization opportunities due to missing try_clean_mem_phi() calls #15445

8314997: Missing optimization opportunities due to missing try_clean_mem_phi() calls #15445

Conversation

chhagedorn commented Aug 28, 2023 • edited by openjdk bot Loading

Progress

Issue

Reviewers

Reviewing

Webrev

bridgekeeper bot commented Aug 28, 2023

openjdk bot commented Aug 28, 2023

mlbridge bot commented Aug 28, 2023 • edited Loading

Webrevs

rwestrel left a comment

Choose a reason for hiding this comment

openjdk bot commented Aug 28, 2023 • edited Loading

chhagedorn commented Aug 29, 2023

vnkozlov left a comment

Choose a reason for hiding this comment

vnkozlov Aug 29, 2023

Choose a reason for hiding this comment

chhagedorn Aug 30, 2023

Choose a reason for hiding this comment

vnkozlov Aug 29, 2023

Choose a reason for hiding this comment

chhagedorn Aug 30, 2023

Choose a reason for hiding this comment

chhagedorn commented Aug 30, 2023

vnkozlov left a comment

Choose a reason for hiding this comment

TobiHartmann left a comment

Choose a reason for hiding this comment

chhagedorn commented Sep 1, 2023

chhagedorn commented Sep 4, 2023

openjdk bot commented Sep 4, 2023

openjdk bot commented Sep 4, 2023

chhagedorn commented Aug 28, 2023 •

edited by openjdk bot

Loading

mlbridge bot commented Aug 28, 2023 •

edited

Loading

openjdk bot commented Aug 28, 2023 •

edited

Loading