Forbid creation of non-faulting null-check nodes. #77078

JulieLeeMSFT · 2022-10-15T14:29:12Z

There are 2 methods that create GT_NULLCHECK: gtNewNullCheck and gtChangeOperToNullCheck.

Check that we don't create null checks on an address that is not null. #44059 added assert(fgAddrCouldBeNull(addr)) to gtNewNullCheck to avoid unnecessary null-checks early.
This PR added an assert assert(fgAddrCouldBeNull(tree->AsUnOp()->gtGetOp1())) to gtChangeOperToNullCheck using @sandreenko's prototype.

Edit: To delete an unnecessary node, we replaced the node with a GT_NULLCHECK or GT_NOP and allowed later phases to delete it.

ghost · 2022-10-15T14:29:23Z

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

Issue Details

Draft to test #44227.

Author:	JulieLeeMSFT
Assignees:	-
Labels:	`area-CodeGen-coreclr`
Milestone:	-

JulieLeeMSFT · 2022-10-27T03:28:18Z

Fixed all asserts, but there are size and throughput regressions on Arm64. Link

@AndyAyersMS @kunalspathak PTAL.

markples · 2022-10-27T19:25:23Z

src/coreclr/jit/lower.cpp

+        const bool isContainable = false;
+        if (ind->Addr() != nullptr)
+        {
+            const bool isContainable = IsSafeToContainMem(ind, ind->Addr());


This is hiding the other isContainable and then going out of scope and the value is lost.

Yep and this is the reason for so many diffs.

oh I forgot to remove it when moving it out of if. Thanks for spotting it.

just a nit here - you could write it as

const bool isContainable = (ind->Addr() != nullptr) && IsSafeToContainMem(ind, ind->Addr());

if you wanted to preserve the constness or shorten it. I don't think it's a big deal.

Thanks for the code snippet. Will make that change.

BruceForstall · 2022-11-28T20:07:08Z

@JulieLeeMSFT This was stale; I converted to Draft. Please convert back to non-Draft when it's ready for final code reviews.

ghost · 2022-12-28T23:01:56Z

Draft Pull Request was automatically closed for 30 days of inactivity. Please let us know if you'd like to reopen it.

JulieLeeMSFT · 2022-12-29T07:41:05Z

Diff result:

Small (~130 Bytes) code size improvements.
Minor throughput diff (0.01%) as shown in the second screenshot below.

JulieLeeMSFT · 2022-12-30T00:48:45Z

@dotnet/jit-contrib @BruceForstall, it is ready to review.
Failure is a knonw issue: #75244.

src/coreclr/jit/compiler.cpp

src/coreclr/jit/gentree.cpp

src/coreclr/jit/importer.cpp

BruceForstall · 2023-01-05T22:02:46Z

src/coreclr/jit/importer.cpp

@@ -8251,7 +8251,14 @@ void Compiler::impImportBlockCode(BasicBlock* block)
                            // via an underlying address, just null check the address.
                            if (op1->OperIs(GT_FIELD, GT_IND, GT_OBJ))
                            {
-                                gtChangeOperToNullCheck(op1, block);
+                                if (fgAddrCouldBeNull(op1->AsUnOp()->gtGetOp1()))


The op1 (addr) of a static GT_FIELD can be nullptr, but fgAddrCouldBeNull doesn't check for that. Could we AV here?

As you said, if the op1 address of GT_FIELD can be nullptr, we want true returned from fgAddrCouldBeNull(op1->gtGetOp1()). And, it falls through to the default case and returns true. Isn't it enough?

No, because fgAddrColdBeNull() has switch (addr->OperGet()) and in the case I'm worried about, addr will be nullptr, so that would AV. Now, maybe there's some reason the case I'm worried about doesn't happen.

Or, maybe you can rewrite it to protect against that, e.g.:

const GenTree* addr = op1->gtGetOp1(); if ((addr != nullptr) && fgAddrCouldBeNull(addr)) { gtChangeOperToNullCheck(op1, block); } else { op1 = gtNewNothingNode(); }

(presumably static fields don't need null checks)

src/coreclr/jit/lower.cpp

ghost · 2023-01-20T03:01:58Z

This pull request has been automatically marked no-recent-activity because it has not had any activity for 14 days. It will be closed if no further activity occurs within 14 more days. Any new comment (by anyone, not necessarily the author) will remove no-recent-activity.

JulieLeeMSFT · 2023-01-22T19:13:15Z

@BruceForstall, it is ready to review.

BruceForstall · 2023-01-25T18:15:33Z

src/coreclr/jit/gentree.cpp

+                        else
+                        {
+                            JITDUMP("Replace an unused OBJ/BLK node [%06d] with a NOTHING node\n", dspTreeID(node));
+                            node = m_compiler->gtNewNothingNode();


The problem is that Append(node) here happens before you set node = m_compiler->gtNewNothingNode();. The NothingNode gets created then never used. And for the non-faulting case, the original BLK node is still on the side-effect list.

I think you want code more like this:

if (m_compiler->gtNodeHasSideEffects(node, m_flags)) { if (node->OperIsBlk() && !node->OperIsStoreBlk()) { // Check for a guaranteed non-faulting IND, and create a NOP node instead of a NULLCHECK in that case. if (m_compiler->fgAddrCouldBeNull(node->AsBlk()->Addr())) { Append(node); JITDUMP("Replace an unused OBJ/BLK node [%06d] with a NULLCHECK\n", dspTreeID(node)); m_compiler->gtChangeOperToNullCheck(node, m_compiler->compCurBB); } else { JITDUMP("Dropping non-faulting OBJ/BLK node [%06d]\n", dspTreeID(node)); } } else { Append(node); } return Compiler::WALK_SKIP_SUBTREES; }

BruceForstall · 2023-01-25T18:32:12Z

src/coreclr/jit/importer.cpp

@@ -8251,7 +8251,14 @@ void Compiler::impImportBlockCode(BasicBlock* block)
                            // via an underlying address, just null check the address.
                            if (op1->OperIs(GT_FIELD, GT_IND, GT_OBJ))
                            {
-                                gtChangeOperToNullCheck(op1, block);
+                                if (fgAddrCouldBeNull(op1->AsUnOp()->gtGetOp1()))


No, because fgAddrColdBeNull() has switch (addr->OperGet()) and in the case I'm worried about, addr will be nullptr, so that would AV. Now, maybe there's some reason the case I'm worried about doesn't happen.

Or, maybe you can rewrite it to protect against that, e.g.:

const GenTree* addr = op1->gtGetOp1(); if ((addr != nullptr) && fgAddrCouldBeNull(addr)) { gtChangeOperToNullCheck(op1, block); } else { op1 = gtNewNothingNode(); }

(presumably static fields don't need null checks)

…des that was supposed to be deleted.

JulieLeeMSFT · 2023-01-26T20:02:26Z

@BruceForstall, it is ready to review.

Forbid creation of non-faulting null-check nodes.

d5e8d6a

dotnet-issue-labeler bot added the area-CodeGen-coreclr CLR JIT compiler in src/coreclr/src/jit and related components such as SuperPMI label Oct 15, 2022

ghost assigned JulieLeeMSFT Oct 15, 2022

JulieLeeMSFT marked this pull request as draft October 15, 2022 14:29

JulieLeeMSFT added this to the 8.0.0 milestone Oct 15, 2022

Skip null check if op1 cannot be null.

0a17d91

build-analysis bot mentioned this pull request Oct 20, 2022

Assertion failed: (*card_word)==0 in DynamicGenerics tests #76801

Closed

JulieLeeMSFT and others added 3 commits October 21, 2022 17:27

x86 bug fix for Skip gtChangeOperToNullCheck if it cannot be null.

37d8f34

Fix Arm64 AVE: Do not lower indir when NOP.

6434f20

formatting fix.

4c1fddd

JulieLeeMSFT marked this pull request as ready for review October 27, 2022 03:28

JulieLeeMSFT requested a review from AndyAyersMS October 27, 2022 03:29

markples reviewed Oct 27, 2022

View reviewed changes

Arm64 regression bug fix that removes const bool for a variable.

98b6b90

BruceForstall marked this pull request as draft November 28, 2022 20:06

ghost closed this Dec 28, 2022

JulieLeeMSFT reopened this Dec 29, 2022

Preserve const and shorten expression.

d9bb584

JulieLeeMSFT marked this pull request as ready for review December 29, 2022 04:00

Fix format.

6bb0d26

JulieLeeMSFT force-pushed the 44227 branch from 8d7ad31 to 6bb0d26 Compare December 29, 2022 05:02

Merge branch 'main' into 44227

58a9b98

build-analysis bot mentioned this pull request Dec 29, 2022

Test failure Loader\\classloader\\DictionaryExpansion\\DictionaryExpansion\\DictionaryExpansion.cmd #75244

Closed

BruceForstall requested changes Jan 5, 2023

View reviewed changes

ghost added the needs-author-action An issue or pull request that requires more info or actions from the author. label Jan 5, 2023

ghost added the no-recent-activity label Jan 20, 2023

ghost removed no-recent-activity needs-author-action An issue or pull request that requires more info or actions from the author. labels Jan 21, 2023

Addressed code review feedback

7193ca3

BruceForstall requested changes Jan 25, 2023

View reviewed changes

ghost added the needs-author-action An issue or pull request that requires more info or actions from the author. label Jan 25, 2023

Forbid creation of non-faulting nullcheck node: Addressed feedback

c984a91

ghost removed the needs-author-action An issue or pull request that requires more info or actions from the author. label Jan 26, 2023

build-analysis bot mentioned this pull request Jan 26, 2023

Tracking issue for CI build timeouts #76454

Closed

runfoapp bot mentioned this pull request Jan 26, 2023

Infra improvements for Helix #68176

Closed

Forbid creation of non-faulting null-check nodes: Remove duplicate co…

e20a756

…des that was supposed to be deleted.

BruceForstall approved these changes Jan 27, 2023

View reviewed changes

JulieLeeMSFT merged commit 08f13dd into dotnet:main Jan 28, 2023

JulieLeeMSFT deleted the 44227 branch January 28, 2023 02:02

This was referenced Jan 31, 2023

JIT: Assertion failed '!tree->IsUnusedValue() || (dstCount != 0)' during 'Linear scan register alloc' #81356

Closed

Update the assert to correctly check if a tree with value has non-zero destination registers #81412

Merged

JulieLeeMSFT mentioned this pull request Feb 8, 2023

What's new in .NET 8 Preview 1 dotnet/core#8133

Closed

3 tasks

ghost locked as resolved and limited conversation to collaborators Feb 27, 2023

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Forbid creation of non-faulting null-check nodes. #77078

Forbid creation of non-faulting null-check nodes. #77078

JulieLeeMSFT commented Oct 15, 2022 •

edited

Loading

ghost commented Oct 15, 2022

JulieLeeMSFT commented Oct 27, 2022

markples Oct 27, 2022

kunalspathak Oct 27, 2022

JulieLeeMSFT Oct 27, 2022

markples Oct 28, 2022

JulieLeeMSFT Oct 31, 2022

BruceForstall commented Nov 28, 2022

ghost commented Dec 28, 2022

JulieLeeMSFT commented Dec 29, 2022

JulieLeeMSFT commented Dec 30, 2022

BruceForstall Jan 5, 2023

JulieLeeMSFT Jan 20, 2023

BruceForstall Jan 25, 2023

JulieLeeMSFT Jan 26, 2023

ghost commented Jan 20, 2023

JulieLeeMSFT commented Jan 22, 2023

BruceForstall Jan 25, 2023

BruceForstall Jan 25, 2023

JulieLeeMSFT commented Jan 26, 2023

Forbid creation of non-faulting null-check nodes. #77078

Forbid creation of non-faulting null-check nodes. #77078

Conversation

JulieLeeMSFT commented Oct 15, 2022 • edited Loading

ghost commented Oct 15, 2022

JulieLeeMSFT commented Oct 27, 2022

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

BruceForstall commented Nov 28, 2022

ghost commented Dec 28, 2022

JulieLeeMSFT commented Dec 29, 2022

JulieLeeMSFT commented Dec 30, 2022

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

ghost commented Jan 20, 2023

JulieLeeMSFT commented Jan 22, 2023

Choose a reason for hiding this comment

Choose a reason for hiding this comment

JulieLeeMSFT commented Jan 26, 2023

JulieLeeMSFT commented Oct 15, 2022 •

edited

Loading