Reland "[TypeProf][InstrPGO] Introduce raw and instr profile format change for type profiling." #82711

minglotus-6 · 2024-02-23T00:12:36Z

New change on top of reviewed patch are in commits after this one. Previous commits are restored from the remote branch with timestamps.

Fix build breakage for non-ELF platforms, by defining the missing functions {__llvm_profile_begin_vtables, __llvm_profile_end_vtables, __llvm_profile_begin_vtabnames , __llvm_profile_end_vtabnames} everywhere.
- Tested on mac laptop (for darwins) and Windows. Specifically, functions in InstrProfilingPlatformWindows.c returns NULL to make it more explicit that type prof isn't supported; see comments for the reason.
- For the rest (AIX, other), mostly follow existing examples (like this one)
Rename __llvm_prf_vtabnames -> __llvm_prf_vns for shorter section name, and make returned pointers const

Original Description

Raw profile format
- Header: records the byte size of compressed vtable names, and the number of profiled vtable entries (call it VTableProfData). Header also records padded bytes of each section.
- Payload: adds a section for compressed vtable names, and a section to store VTableProfData. Both sections are padded so the size is a multiple of 8.
Indexed profile format
- Header: records the byte offset of compressed vtable names.
- Payload: adds a section to store compressed vtable names. This section is used by llvm-profdata to show the list of vtables profiled for an instrumented site.

The originally reviewed patch will have profile reader/write change and llvm-profdata change.

To ensure this PR has all the necessary profile format change along with profile version bump, created a copy of the originally reviewed patch in A copy of https://github.com/llvm/llvm-project/pull/66825. Actually this is a superset (a copy plus the thinlto-import change) #80761. The copy doesn't have profile format change, but it has the set of tests which covers type profile generation, profile read and profile merge. Tests pass there.

rfc in https://discourse.llvm.org/t/rfc-dynamic-type-profiling-and-optimizations-in-llvm/74600

change, to get smaller PRs.

…a.inc

Co-authored-by: modiking <modiking213@gmail.com>

…er fields in raw profile header

profile version is outdated. - update llvm/test/tools/llvm-profdata/Inputs/compressed.profraw along with raw profile version bump - The test requires '!zlib', so I generate raw profile with DLLVM_ENABLE_ZLIB=1 and run the test with DLLVM_ENABLE_ZLIB=0.

github-actions · 2024-02-23T00:15:11Z

✅ With the latest revision this PR passed the C/C++ code formatter.

minglotus-6 · 2024-02-23T00:27:17Z

This is a link (https://lab.llvm.org/buildbot/#/builders/127/builds/62532/steps/8/logs/stdio) for previously broken builds.

minglotus-6 · 2024-02-23T00:43:13Z

compiler-rt/lib/profile/InstrProfilingPlatformWindows.c

 const char *__llvm_profile_begin_names(void) { return &NamesStart + 1; }
 const char *__llvm_profile_end_names(void) { return &NamesEnd; }

+const char *__llvm_profile_begin_vtabnames(void) { return NULL; }


https://gcc.godbolt.org/z/faMzTbGh3 as a compile test.

minglotus-6 · 2024-02-23T01:04:08Z

@w2yehia I wonder if you could help take a look at the changes to compiler-rt/lib/profile/InstrProfilingPlatformAIX.c? I expect them to be no-op in terms of new features, but would like a second pair of eyes. Thanks!

w2yehia

AIX changes lgtm.

…plement the linker magic. also remove one unused variable

minglotus-6 · 2024-02-26T23:34:37Z

Update: I got the Windows machine. Windows linker magic [1] works out without compile errors or bad side effects, but the current type profiling won't work under MSVC ABI. Thereby helper functions in compiler-rt/lib/profile/InstrProfilingPlatformWindows.c is made simple (just return NULL) such that it's very explicit that type profiling isn't supported.

To elaborate on why current type profiling implementation doesn't work under MSVC ABI

Under MSVC ABI, the name of the vtable variable is not necessarily the mangled name. Take https://gcc.godbolt.org/z/drs1n9qd5 as an example, @0, @1 and @2 are vtables.
For instrumentation && profile collection under Itanium ABI, the vtables' mangled names (and its MD5hash) are used to map profiled runtime address to actual vtable variables.

As the godbolt example shows, the MSVC mangled name is still in the IR (just not vtable's variable name). I added some comments.

[1] diff in https://gist.github.com/minglotus-6/9df208e14f7c20b352cdb7f89d38c228

minglotus-6 · 2024-02-27T19:07:25Z

Given this is a reland to make the patch work on non-ELF platforms, I'll go ahead to commit this.

Several profdata tests pass the byte 012 to printf. This causes these tests to fail when using GnuWin32's version of printf because printf will detect that 012 is the LF character and will prepend the byte 015 (CR) in front of LF. This change is required after #82711 which bumped the version number.

…after #82711

This is a follow-up to the profile format change in #82711

whentojump · 2024-03-31T19:48:13Z

Hi @minglotus-6 , can you please also update this place:

llvm-project/llvm/include/llvm/ProfileData/InstrProf.h

Lines 1183 to 1198 in 82c6eee

    
           namespace RawInstrProf { 
        
           // Version 1: First version 
        
           // Version 2: Added value profile data section. Per-function control data 
        
           // struct has more fields to describe value profile information. 
        
           // Version 3: Compressed name section support. Function PGO name reference 
        
           // from control data struct is changed from raw pointer to Name's MD5 value. 
        
           // Version 4: ValueDataBegin and ValueDataSizes fields are removed from the 
        
           // raw header. 
        
           // Version 5: Bit 60 of FuncHash is reserved for the flag for the context 
        
           // sensitive records. 
        
           // Version 6: Added binary id. 
        
           // Version 7: Reorder binary id and include version in signature. 
        
           // Version 8: Use relative counter pointer. 
        
           // Version 9: Added relative bitmap bytes pointer and count used by MC/DC. 
        
           const uint64_t Version = INSTR_PROF_RAW_VERSION;

And may I ask whether you plan to back port this to branch 18? (I'm asking because we have some dependency on this version and would like to know in advance). Thanks!

- The raw profile format change is in #82711

minglotus-6 · 2024-04-01T00:33:47Z

can you please also update this place

Thanks for the catch! Updated in f87bde2

whether you plan to back port this to branch 18? (I'm asking because we have some dependency on this version and would like to know in advance)

Before seeing your comment, I don't have plans to back port the version update to LLVM version 18. Could you elaborate on how back porting a version bump to LLVM branch 18 helps? I have seen back-ports for bug fixes but haven't done it myself. I'm glad to back port if it helps.

whentojump · 2024-04-01T02:29:13Z

Updated in f87bde2

Thanks!

By the way, https://llvm.org/docs/InstrProfileFormat.html is a really informative page. Thanks for your contributions!

Before seeing your comment, I don't have plans to back port the version update to LLVM version 18. Could you elaborate on how back porting a version bump to LLVM branch 18 helps? I have seen back-ports for bug fixes but haven't done it myself. I'm glad to back port if it helps.

Sorry for not making it clear -- in fact, not backporting it can better help our use cases. :)) or :((

In short, we are instrumenting some OS kernels, like what this patch does.

In kernel space, it's impossible to dump the profiles using any "runtime". So instead, some kernel code itself has to do this job. As a result, this piece of kernel code has to keep synchronized with LLVM code, e.g. in terms of profile format. That's why a stable format is certainly more desirable for us. You may notice in the linked patch: they were using raw profile version 5 three years ago and now we are 10 already. (In user space, this should not be a problem, because whatever produces or consumes the profile usually comes from the same version of LLVM.)

That being said, please make the decision at your own discretion, or based on broader community feedback. :))

minglotus-6 · 2024-04-01T05:18:27Z

in fact, not backporting it can better help our use cases. :)) or :((

In short, we are instrumenting some OS kernels, like what this patch does.
...
That being said, please make the decision at your own discretion, or based on broader community feedback. :))

Ack, thanks for sharing!

minglotus-6 and others added 17 commits February 13, 2024 16:38

[TypeProf]Introduce value type in raw profiles

0dd267b

undo the changes that don't necessarily bundle with profile format

a4ec017

change, to get smaller PRs.

fix the alignment of struct

9c833bd

polish format

d56bb34

fix format

206415c

resolve review feedback

84570cd

add back one blank line in llvm/include/llvm/ProfileData/InstrProfDat…

0c9ed57

…a.inc

Apply suggestions from code review

67bbafb

Co-authored-by: modiking <modiking213@gmail.com>

return truncated error if any

c30a699

resolve review fee feedback

393d166

resolve review feedback

744b87f

Merge branch 'main' into users/minglotus-6/typeprofrawformat

4786174

restore __llvm_profile_get_num_vtable, and swap the order of two head…

36541d9

…er fields in raw profile header

Merge branch 'main' into users/minglotus-6/typeprofrawformat

d0757f4

port new object file sections and helper functions for non-ELF files

a825d2a

undo linker magic on Windows

90bdd0d

minglotus-6 requested review from david-xl, snehasish, modiking and teresajohnson February 23, 2024 00:20

minglotus-6 marked this pull request as ready for review February 23, 2024 00:20

minglotus-6 mentioned this pull request Feb 23, 2024

[TypeProf][InstrPGO] Introduce raw and instr profile format change for type profiling. #81691

Merged

run clang-format

8b35f49

minglotus-6 added 2 commits February 22, 2024 16:32

Return const pointer (not pointer) for *Linux.c

ac0515a

#include<stddef.h> for Windows

0de9631

minglotus-6 commented Feb 23, 2024

View reviewed changes

run clang-format

13a7e78

minglotus-6 requested a review from w2yehia February 23, 2024 01:01

minglotus-6 added 2 commits February 22, 2024 17:06

fix a type for *Other.c

84e702e

Add COMPILER_RT_VISIBLITY to not leak the symbols

51fca4a

w2yehia reviewed Feb 23, 2024

View reviewed changes

add comment to explain why InstrProfilingPlatformWindows.c doesn't im…

8cad3bd

…plement the linker magic. also remove one unused variable

minglotus-6 merged commit 16e74fd into main Feb 27, 2024
4 checks passed

minglotus-6 deleted the users/minglotus-6/typeprofrawformat branch February 27, 2024 19:07

minglotus-6 mentioned this pull request Feb 28, 2024

[docs][TypeProf]Update instrumentation file format document #83309

Merged

alanzhao1 mentioned this pull request Mar 4, 2024

[profdata][nfc] Disable several tests on Windows #83907

Merged

wangleiat added a commit that referenced this pull request Mar 6, 2024

[InstrProf][NFC] Fix -Wimplicit-fallthrough warning in InstrProf.cpp …

f439c71

…after #82711

minglotus-6 added a commit that referenced this pull request Mar 28, 2024

[docs][TypeProf]Update instrumentation file format document (#83309)

07a1fbe

This is a follow-up to the profile format change in #82711

minglotus-6 mentioned this pull request Mar 31, 2024

Add comment for raw profile version 10. #87207

Merged

minglotus-6 added a commit that referenced this pull request Apr 1, 2024

[nfc] Update comment for raw profile version 10. (#87207)

f87bde2

- The raw profile format change is in #82711

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Reland "[TypeProf][InstrPGO] Introduce raw and instr profile format change for type profiling." #82711

Reland "[TypeProf][InstrPGO] Introduce raw and instr profile format change for type profiling." #82711

minglotus-6 commented Feb 23, 2024 •

edited

github-actions bot commented Feb 23, 2024 •

edited

minglotus-6 commented Feb 23, 2024

minglotus-6 Feb 23, 2024 •

edited

minglotus-6 commented Feb 23, 2024 •

edited

w2yehia left a comment

minglotus-6 commented Feb 26, 2024 •

edited

minglotus-6 commented Feb 27, 2024

whentojump commented Mar 31, 2024

minglotus-6 commented Apr 1, 2024

whentojump commented Apr 1, 2024

minglotus-6 commented Apr 1, 2024

Reland "[TypeProf][InstrPGO] Introduce raw and instr profile format change for type profiling." #82711

Reland "[TypeProf][InstrPGO] Introduce raw and instr profile format change for type profiling." #82711

Conversation

minglotus-6 commented Feb 23, 2024 • edited

github-actions bot commented Feb 23, 2024 • edited

minglotus-6 commented Feb 23, 2024

minglotus-6 Feb 23, 2024 • edited

Choose a reason for hiding this comment

minglotus-6 commented Feb 23, 2024 • edited

w2yehia left a comment

Choose a reason for hiding this comment

minglotus-6 commented Feb 26, 2024 • edited

minglotus-6 commented Feb 27, 2024

whentojump commented Mar 31, 2024

minglotus-6 commented Apr 1, 2024

whentojump commented Apr 1, 2024

minglotus-6 commented Apr 1, 2024

minglotus-6 commented Feb 23, 2024 •

edited

github-actions bot commented Feb 23, 2024 •

edited

minglotus-6 Feb 23, 2024 •

edited

minglotus-6 commented Feb 23, 2024 •

edited

minglotus-6 commented Feb 26, 2024 •

edited