8325821: [REDO] use "dmb.ishst+dmb.ishld" for release barrier #19278

kuaiwei · 2024-05-17T08:57:20Z

he origin patch for https://bugs.openjdk.org/browse/JDK-8324186 has 2 issues:
1 It show regression in some platform, like Apple silicon in mac os
2 Can not handle instruction sequence like "dmb.ishld; dmb.ishst; dmb.ishld; dmb.ishld"

It can be fixed by:
1 Enable AlwaysMergeDMB by default, only disable it in architecture we can see performance improvement (N1 or N2)
2 Check the special pattern and merge the subsequent dmb.

It also fix a bug when code buffer is expanding, st/ld/dmb can not be merged. I added unit tests for these.

This patch still has a unhandled case. Insts like "dmb.ishld; dmb.ishst; dmb.ish", it will merge the last 2 instructions and can not merge all three. Because when emitting dmb.ish, if merge all previous dmbs, the code buffer will shrink the size. I think it may break some resumption and think it's not a common pattern.

In previous PR #18467 , I tried an implementation to use state machine for merging. But it looks risky to pending instruction during emitting.

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-8325821: [REDO] use "dmb.ishst+dmb.ishld" for release barrier (Enhancement - P4)

Reviewers

Aleksey Shipilev (@shipilev - Reviewer) ⚠️ Review applies to 68c3018c
Andrew Haley (@theRealAph - Reviewer) ⚠️ Review applies to 68c3018c

Reviewing

Using git

Checkout this PR locally:
$ git fetch https://git.openjdk.org/jdk.git pull/19278/head:pull/19278
$ git checkout pull/19278

Update a local copy of the PR:
$ git checkout pull/19278
$ git pull https://git.openjdk.org/jdk.git pull/19278/head

Using Skara CLI tools

Checkout this PR locally:
$ git pr checkout 19278

View PR using the GUI difftool:
$ git pr show -t 19278

Using diff file

Download this PR as a diff file:
https://git.openjdk.org/jdk/pull/19278.diff

Webrev

Link to Webrev Comment

bridgekeeper · 2024-05-17T08:58:06Z

👋 Welcome back kuaiwei! 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 · 2024-05-17T08:58:53Z

@kuaiwei 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:

8325821: [REDO] use "dmb.ishst+dmb.ishld" for release barrier

Reviewed-by: shade, aph

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 81 new commits pushed to the master branch:

e7dc76b: 8333849: (dc) DatagramChannel send/receive fails with UOE if buffer backed by memory segment allocated from shared arena
ec1664e: 8333804: java/net/httpclient/ForbiddenHeadTest.java threw an exception with 0 failures
e22fc12: 8333775: Small improvement to outputStream auto-indentation mode
7b43a8c: 8333824: Unused ClassValue in VarHandles
5f9d3e3: 8333722: Fix CompilerDirectives for non-compiler JVM variants
83b3441: 8322811: jcmd System.dump_map help info has conflicting statements
8aa35ca: 8333833: Remove the use of ByteArrayLittleEndian from UUID::toString
de55db2: 8333522: JFR SwapSpace event might read wrong free swap space size
a941397: 8329031: CPUID feature detection for Advanced Performance Extensions (Intel® APX)
8d2f9e5: 8333749: Consolidate ConstantDesc conversion in java.base
... and 71 more: https://git.openjdk.org/jdk/compare/d826127970bd2ae8bf4cacc3c55634dc5af307c4...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.

As you do not have Committer status in this project an existing Committer must agree to sponsor your change. Possible candidates are the reviewers of this PR (@shipilev, @theRealAph) but any other Committer may sponsor as well.

➡️ To flag this PR as ready for integration with the above commit message, type /integrate in a new comment. (Afterwards, your sponsor types /sponsor in a new comment to perform the integration).

openjdk · 2024-05-17T08:59:24Z

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

hotspot

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 · 2024-05-17T09:02:12Z

Webrevs

theRealAph · 2024-05-20T13:59:31Z

src/hotspot/cpu/aarch64/aarch64.ad


  ins_encode %{
    __ block_comment("membar_release");
-    __ membar(Assembler::LoadStore|Assembler::StoreStore);
+    __ membar(Assembler::StoreStore);


Do we need to respect AlwaysMergeDMBhere?

Yes, usually they can be merged in macroAssembler. but it can help to reduce the possibility of unmerged case. Thanks to point it.

I checked code again. They will be merged if enable AlwaysMergeDMB. So we can skip the check.

Add a comment:

// These will be merged if AlwaysMergeDMB is enabled.

Comment added.

shipilev

Cursory review:

shipilev · 2024-05-22T10:25:40Z

test/hotspot/gtest/aarch64/test_assembler_aarch64.cpp

+  asm_check((const unsigned int *)code.insts()->start(), insns, sizeof insns / sizeof insns[0]);
+}
+
+TEST_VM(AssemblerAArch64, merge_ldst) {


This test seems to be irrelevant for the issue at hand? Tests ld/st -> ldp/stp merging, not the barrier merges?

In this patch, I fixed an issue, dmb/st/ld may not merge if CodeBuffer is expanding, I added some unit tests to check it.

Aha. So the logic is that the same issue that affects dmb merging also affects ldp merging? All right, it's fine to leave it here then.

src/hotspot/cpu/aarch64/globals_aarch64.hpp

shipilev · 2024-05-22T11:07:57Z

test/hotspot/gtest/aarch64/test_assembler_aarch64.cpp

+  printf("%s\n", ss.as_string());
+}
+
+TEST_VM(AssemblerAArch64, merge_dmb) {


Given the previous experience with barrier merges that prompted the backout, I would prefer to have a more comprehensive test here, maybe an additional one. I am thinking something like the exhaustive combination of 4 back-to-back barriers of each of 5 types. This gives us 5^4 = 625 test cases, which I think is still manageable.

Test is added as merge_dmb_all_kinds

Right. I was implicitly thinking that we can do this without coding the explicit patterns into the test. As it stands now, it is hard to check that generated patterns are actually correct. Let me see if I can whip up a sample of what I had in mind.

I was thinking about this: improve-tests.patch. Note how it uses the constants for better readability, and also runs the test in both AlwaysMergeDMB modes. You might want to adapt other tests to similar pattern.

Thanks for your patch. I patched and add few sanity check.

theRealAph · 2024-05-29T11:11:35Z

This looks ready to me. I think we need jcstress with C1 and C2, and we should be done. @shipilev , do you agree?

shipilev · 2024-05-29T11:12:38Z

This looks ready to me. I think we need jcstress with C1 and C2, and we should be done. @shipilev , do you agree?

Yes. Just run jcstress with defaults, maybe limiting the time budget to about 24 hours, and we are done. Default configuration would work through different combinations of C1/C2 compilations for all actors, which is what we want to check for this change: that we don't mess up the barrier emitting scheme in different compilers/interpreters.

shipilev · 2024-05-29T11:34:09Z

Note that current jcstress run would likely fail due to JDK-8332670.

kuaiwei · 2024-05-30T08:24:55Z

This looks ready to me. I think we need jcstress with C1 and C2, and we should be done. @shipilev , do you agree?

Yes. Just run jcstress with defaults, maybe limiting the time budget to about 24 hours, and we are done. Default configuration would work through different combinations of C1/C2 compilations for all actors, which is what we want to check for this change: that we don't mess up the barrier emitting scheme in different compilers/interpreters.

I can run the jcstress test. I will run fastdebug build with
java -jar jcstress-latest.jar -tb 24h
Is it the correct command ?

shipilev · 2024-05-30T08:26:56Z

I can run the jcstress test. I will run fastdebug build with java -jar jcstress-latest.jar -tb 24h Is it the correct command ?

Yes, I think so.

shipilev · 2024-05-31T08:23:51Z

I can run the jcstress test. I will run fastdebug build with java -jar jcstress-latest.jar -tb 24h Is it the correct command >
Yes, I think so.

FTR, I ran Linux AArch64 server release on Graviton 3 instance (ergonomics selects -AlwaysMergeDMB there) for 12 hours. Apart from failures from JDK-8332670, I see no other trouble. Scheduled a quick run with +AlwaysMergeDMB as well.

kuaiwei · 2024-05-31T10:06:06Z

I can run the jcstress test. I will run fastdebug build with java -jar jcstress-latest.jar -tb 24h Is it the correct command >
Yes, I think so.

FTR, I ran Linux AArch64 server release on Graviton 3 instance (ergonomics selects -AlwaysMergeDMB there) for 12 hours. Apart from failures from JDK-8332670, I see no other trouble. Scheduled a quick run with +AlwaysMergeDMB as well.

Thanks for testing. I'm running jcstress on a neoverse-n2 instance. I got some "soft errs" in console output. Are they real error?

shipilev · 2024-05-31T10:08:47Z

test/hotspot/gtest/aarch64/test_assembler_aarch64.cpp

+constexpr uint32_t test_encode_dmb_ld = 0xd50339bf;
+constexpr uint32_t test_encode_dmb_st = 0xd5033abf;
+constexpr uint32_t test_encode_dmb    = 0xd5033bbf;


Can you maybe move these to the top, and use these constants across the test? You would not need the comments like 0xd5033abf, // dmb.ishst then.

shipilev · 2024-05-31T10:09:30Z

test/hotspot/gtest/aarch64/test_assembler_aarch64.cpp

+  asm_check((const unsigned int *)code.insts()->start(), insns, sizeof insns / sizeof insns[0]);
+}
+
+TEST_VM(AssemblerAArch64, merge_ldst) {


Aha. So the logic is that the same issue that affects dmb merging also affects ldp merging? All right, it's fine to leave it here then.

shipilev · 2024-05-31T11:02:31Z

src/hotspot/cpu/aarch64/macroAssembler_aarch64.hpp

@@ -150,6 +150,7 @@ class MacroAssembler: public Assembler {
  void bind(Label& L) {
    Assembler::bind(L);
    code()->clear_last_insn();
+    code()->set_last_label(pc());


OK, so we have added _last_label to shared code in codeBuffer, but only update it in aarch64. This would be surprising for other platforms. On the other hand, this is what we already do with _last_insn -- only implementing it for specific platforms. Probably fine, but it would be nice to strengthen this with asserts, maybe in separate PR.

It reminds me it could be applied to riscv. It also need merge membar. I will move this part to a new PR.

I need _last_label in this patch. I need it to check previous 2 instructions and they are not cross block boundary. I will create a new PR for RISCV only.

shipilev · 2024-05-31T11:04:49Z

Thanks for testing. I'm running jcstress on a neoverse-n2 instance. I got some "soft errs" in console output. Are they real error?

Soft errors are API inconsistencies that jcstress harness handled. As long as you don't have new hard errors, you are fine.

kuaiwei · 2024-06-03T06:42:18Z

I can run the jcstress test. I will run fastdebug build with java -jar jcstress-latest.jar -tb 24h Is it the correct command >
Yes, I think so.

FTR, I ran Linux AArch64 server release on Graviton 3 instance (ergonomics selects -AlwaysMergeDMB there) for 12 hours. Apart from failures from JDK-8332670, I see no other trouble. Scheduled a quick run with +AlwaysMergeDMB as well.

Thanks for testing. I'm running jcstress on a neoverse-n2 instance. I got some "soft errs" in console output. Are they real error?

My jcstress test is done and no error is reported.

theRealAph

I think we're done here. Thank you for your patience and hard work.
@shipilev, are you happy too?

shipilev

Yes, I believe we are done. Good job!

kuaiwei · 2024-06-04T02:12:59Z

Thanks for your review.

kuaiwei · 2024-06-04T02:13:11Z

/integrate

openjdk · 2024-06-04T02:13:36Z

@kuaiwei
Your change (at version 68c3018) is now ready to be sponsored by a Committer.

kuaiwei · 2024-06-04T11:00:42Z

@theRealAph @shipilev Could you help sponsor it ?
Thanks.

shipilev · 2024-06-04T17:04:19Z

GH says the branch has conflicts that must be resolved. Merging from master is advisable anyway to avoid merge surprises. Let's do that before integrating.

TobiHartmann · 2024-06-05T05:18:22Z

Given that the first version of the patch had some issues, I would recommend waiting for the fork tomorrow and only integrating this into JDK 24.

TobiHartmann · 2024-06-05T05:19:55Z

I'll run this through our correctness and performance testing and report back once it passed.

shipilev · 2024-06-07T09:06:12Z

I'll run this through our correctness and performance testing and report back once it passed.

How was it, Tobias? Any surprises?

TobiHartmann · 2024-06-10T09:30:25Z

Sorry for the delay, I had to re-run a subset of the benchmarks due to high variance. All green now.

shipilev · 2024-06-10T09:46:15Z

Sorry for the delay, I had to re-run a subset of the benchmarks due to high variance. All green now.

Great, thanks for testing.

I think we are ready to integrate this now.

kuaiwei · 2024-06-10T12:47:07Z

/integrate

openjdk · 2024-06-10T12:48:48Z

@kuaiwei
Your change (at version 0afa693) is now ready to be sponsored by a Committer.

shipilev · 2024-06-10T12:55:29Z

/sponsor

openjdk · 2024-06-10T12:57:05Z

Going to push as commit 2a242db.
Since your change was applied there have been 81 commits pushed to the master branch:

e7dc76b: 8333849: (dc) DatagramChannel send/receive fails with UOE if buffer backed by memory segment allocated from shared arena
ec1664e: 8333804: java/net/httpclient/ForbiddenHeadTest.java threw an exception with 0 failures
e22fc12: 8333775: Small improvement to outputStream auto-indentation mode
7b43a8c: 8333824: Unused ClassValue in VarHandles
5f9d3e3: 8333722: Fix CompilerDirectives for non-compiler JVM variants
83b3441: 8322811: jcmd System.dump_map help info has conflicting statements
8aa35ca: 8333833: Remove the use of ByteArrayLittleEndian from UUID::toString
de55db2: 8333522: JFR SwapSpace event might read wrong free swap space size
a941397: 8329031: CPUID feature detection for Advanced Performance Extensions (Intel® APX)
8d2f9e5: 8333749: Consolidate ConstantDesc conversion in java.base
... and 71 more: https://git.openjdk.org/jdk/compare/d826127970bd2ae8bf4cacc3c55634dc5af307c4...master

Your commit was automatically rebased without conflicts.

openjdk · 2024-06-10T12:57:12Z

@shipilev @kuaiwei Pushed as commit 2a242db.

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

8325821: [REDO] use "dmb.ishst+dmb.ishld" for release barrier

Loading
Loading status checks…

b71a1b3

openjdk bot added the rfr label May 17, 2024

openjdk bot added the hotspot label May 17, 2024

theRealAph reviewed May 20, 2024

View reviewed changes

Make MacroAssembler::merge more clear

Loading
Loading status checks…

8767e8f

shipilev reviewed May 22, 2024

View reviewed changes

kuaiwei added 2 commits May 23, 2024 13:46

Add more unit tests

Loading
Loading status checks…

6214b43

Refine merge dmb test cases

Loading
Loading status checks…

00262c4

openjdk bot removed the rfr label May 24, 2024

Remove tailing white space

Loading
Loading status checks…

8ef3e03

openjdk bot added the rfr label May 27, 2024

Add comment in aarch64.ad

Loading
Loading status checks…

7df2103

shipilev reviewed May 31, 2024

View reviewed changes

Use constexpr for test encoding

Loading
Loading status checks…

68c3018

theRealAph approved these changes Jun 3, 2024 •

edited

Loading

View reviewed changes

theRealAph reviewed Jun 3, 2024

View reviewed changes

openjdk bot added the ready label Jun 3, 2024

shipilev approved these changes Jun 3, 2024

View reviewed changes

openjdk bot added the sponsor label Jun 4, 2024

Merge master

Loading
Loading status checks…

0afa693

openjdk bot removed the sponsor label Jun 5, 2024

openjdk bot added the sponsor label Jun 10, 2024

openjdk bot added the integrated label Jun 10, 2024

openjdk bot closed this Jun 10, 2024

openjdk bot removed ready rfr sponsor labels Jun 10, 2024

8325821: [REDO] use "dmb.ishst+dmb.ishld" for release barrier #19278

8325821: [REDO] use "dmb.ishst+dmb.ishld" for release barrier #19278

Conversation

kuaiwei commented May 17, 2024 • edited by openjdk bot Loading

Progress

Issue

Reviewers

Reviewing

Webrev

bridgekeeper bot commented May 17, 2024

openjdk bot commented May 17, 2024 • edited Loading

openjdk bot commented May 17, 2024

mlbridge bot commented May 17, 2024 • edited Loading

Webrevs

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

shipilev left a 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

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

shipilev May 23, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

theRealAph commented May 29, 2024

shipilev commented May 29, 2024 • edited Loading

shipilev commented May 29, 2024

kuaiwei commented May 30, 2024

shipilev commented May 30, 2024

shipilev commented May 31, 2024

kuaiwei commented May 31, 2024

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

Choose a reason for hiding this comment

shipilev commented May 31, 2024

kuaiwei commented Jun 3, 2024

theRealAph left a comment

Choose a reason for hiding this comment

shipilev left a comment

Choose a reason for hiding this comment

kuaiwei commented Jun 4, 2024

kuaiwei commented Jun 4, 2024

openjdk bot commented Jun 4, 2024

kuaiwei commented Jun 4, 2024

shipilev commented Jun 4, 2024

TobiHartmann commented Jun 5, 2024

TobiHartmann commented Jun 5, 2024

shipilev commented Jun 7, 2024

TobiHartmann commented Jun 10, 2024

shipilev commented Jun 10, 2024

kuaiwei commented Jun 10, 2024

openjdk bot commented Jun 10, 2024

shipilev commented Jun 10, 2024

openjdk bot commented Jun 10, 2024

openjdk bot commented Jun 10, 2024

kuaiwei commented May 17, 2024 •

edited by openjdk bot

Loading

openjdk bot commented May 17, 2024 •

edited

Loading

mlbridge bot commented May 17, 2024 •

edited

Loading

shipilev May 23, 2024 •

edited

Loading

shipilev commented May 29, 2024 •

edited

Loading