This is an archive of the discontinued LLVM Phabricator instance.

Change some addUsedGlobal to addUsedOrCompilerUsedGlobal
ClosedPublic

Authored by MaskRay on Feb 24 2021, 11:40 PM.

Details

Summary

An global value in the llvm.used list does not have GC root semantics on ELF targets.
This will be changed in a subsequent backend patch.

Change some llvm.used in the ELF code path to use llvm.compiler.used to
prevent undesired GC root semantics.

Change one extern "C" alias (due to __attribute__((used)) in extern "C") to use llvm.compiler.used on all targets.

GNU ld has a rule "__start_/__stop_ references from a live input section retain the associated C identifier name sections",
which LLD may drop entirely (currently refined to exclude SHF_LINK_ORDER/SHF_GROUP) in a future release (the rule makes it clumsy to GC metadata sections; D96914 added a way to try the potential future behavior).
For llvm.used global values defined in a C identifier name section, keep using llvm.used so that
the future LLD change will not affect them.

rnk kindly categorized the changes:

ObjC/blocks: this wants GC root semantics, since ObjC mainly runs on Mac.
MS C++ ABI stuff: wants GC root semantics, no change
OpenMP: unsure, but GC root semantics probably don't hurt
CodeGenModule: affected in this patch to *not* use GC root semantics so that __attribute__((used)) behavior remains the same on ELF, plus two other minor use cases that don't want GC semantics
Coverage: Probably want GC root semantics
CGExpr.cpp: refers to LTO, wants GC root
CGDeclCXX.cpp: one is MS ABI specific, so yes GC root, one is some other C++ init functionality, which should form GC roots (C++ initializers can have side effects and must run)
CGDecl.cpp: Changed in this patch for __attribute__((used))

Diff Detail

Event Timeline

MaskRay requested review of this revision.Feb 24 2021, 11:40 PM
MaskRay created this revision.
Herald added a project: Restricted Project. · View Herald TranscriptFeb 24 2021, 11:40 PM
Herald added a subscriber: cfe-commits. · View Herald Transcript

GNU ld has a rule "start_/stop_ references from a live input section retain the associated C identifier name sections",

which LLD may change in the future (D96914).

The phrasing "may change" implies LLD could eliminate the rule; in fact D96914 will only add a way to opt-out of the rule. You can't eliminate the rule entirely without breaking a lot of useful cases.

MaskRay edited the summary of this revision. (Show Details)Feb 25 2021, 9:22 AM

GNU ld has a rule "start_/stop_ references from a live input section retain the associated C identifier name sections",

which LLD may change in the future (D96914).

The phrasing "may change" implies LLD could eliminate the rule; in fact D96914 will only add a way to opt-out of the rule. You can't eliminate the rule entirely without breaking a lot of useful cases.

Reworded this sentence. D96914 does add a way to opt-out of the rule. In the future (when toolchains are more mature) we may try. My internal tests and Linux kernel tests are try to state that we may not have many dependent cases.

GNU ld has a rule "start_/stop_ references from a live input section retain the associated C identifier name sections", which LLD may change in the future (D96914).

The phrasing "may change" implies LLD could eliminate the rule; in fact D96914 will only add a way to opt-out of the rule. You can't eliminate the rule entirely without breaking a lot of useful cases.

Reworded this sentence. D96914 does add a way to opt-out of the rule. In the future (when toolchains are more mature) we may try. My internal tests and Linux kernel tests are try to state that we may not have many dependent cases.

My guess is that many use-cases for __start_/__stop_ and used sections are in test and debugging code; it might not be obvious where a dependency is, and it might well be the case that GC-ing these sections will make something useless without actively breaking it. The behavior of __start_/__stop_ has been in place for a very long time, IIUC, and it would be very hard to prove that nothing bad will happen if you eliminate that behavior. I don't think "Linux will still boot" is a particularly exhaustive test case.

Rather than just make vague worried noises, I will work on putting up a patch with my particular use-case, and we can see whether I am worried over nothing. If we're lucky I can do that tomorrow, or it might take another week.

zzheng added a subscriber: zzheng.Feb 25 2021, 10:18 AM
rnk added a comment.Feb 25 2021, 11:57 AM

I think the main user-facing feature here to think about is __attribute__((used)). I see that it's currently listed as Undocumented. We should fix that while we're rehashing how this is supposed to work, and clarify what it does on each platform. As I understand it, __attribute__((used)) globals will be GC roots on Mac/Win, but not ELF targets. The mismatch in behavior is unfortunate, but it matches the dominant system compilers on those platforms.

Here are all the existing addUsedGlobal call sites: https://reviews.llvm.org/P8257 I categorize them as:

  • ObjC/blocks: this wants GC root semantics, since ObjC mainly runs on Mac.
  • MS C++ ABI stuff: wants GC root semantics, no change
  • OpenMP: unsure, but GC root semantics probably don't hurt
  • CodeGenModule: affected in this patch to *not* use GC root semantics so that __attribute__((used)) behavior remains the same on ELF, plus two other minor use cases that don't want GC semantics
  • Coverage: Probably want GC root semantics
  • CGExpr.cpp: refers to LTO, wants GC root
  • CGDeclCXX.cpp: one is MS ABI specific, so yes GC root, one is some other C++ init functionality, which should form GC roots (C++ initializers can have side effects and must run)
  • CGDecl.cpp: Changed in this patch for __attribute__((used))

So, looks good, I think you spotted all the things that should change.

clang/lib/CodeGen/CodeGenModule.cpp
1829

Should -fkeep-static-consts be getting GC root semantics on Mac/Win? We can leave this as behavior preserving, I don't think -fkeep-static-consts is a very important feature.

5924

This probably doesn't need GC semantics, and should always used llvm.compiler.used on all platforms. This alias will be internal, this factory function it takes the linkage from the aliasee.

MaskRay updated this revision to Diff 326480.Feb 25 2021, 1:32 PM
MaskRay edited the summary of this revision. (Show Details)

Adopted rnk's list (Thanks!)
Address comments

rnk accepted this revision.Feb 25 2021, 1:55 PM

lgtm

I see that it's currently listed as Undocumented. We should fix that while we're rehashing how this is supposed to work, and clarify what it does on each platform. As I understand it, attribute((used)) globals will be GC roots on Mac/Win, but not ELF targets. The mismatch in behavior is unfortunate, but it matches the dominant system compilers on those platforms.

I found the doc update in the other patch:
https://reviews.llvm.org/D97447#change-3MBnnJ2r7Yn8

clang/lib/CodeGen/CodeGenModule.cpp
2093

nit: CGM has a getTriple() helper to shorten this.

This revision is now accepted and ready to land.Feb 25 2021, 1:55 PM
MaskRay updated this revision to Diff 326739.Feb 26 2021, 10:35 AM

Simplify with CodeGenModule::getTriple()

This revision was landed with ongoing or failed builds.Feb 26 2021, 10:42 AM
This revision was automatically updated to reflect the committed changes.
JonChesterfield reopened this revision.EditedJan 13 2022, 4:15 AM
JonChesterfield added a subscriber: JonChesterfield.

I don't know the context of this patch but changing attribute((used)) to put things under compiler.used is definitely a breaking change. Please introduce a new attribute if necessary for garbage collection purposes instead of breaking this.

If I had a red buildbot to reference I would have reverted - the commit message does not make it clear that this is a breaking change. This broke a debugging hook in openmp, which is apparently not tested, and will break any application code that uses compiler.used on a target that uses elf.

Linked docs at https://reviews.llvm.org/D97447 suggest applications can get back to a working state by marking things as attribute((used)) attribute((retain)), presumably guarded by a test for whether retain exists. I think this would be another point in a codebase that has to distinguish between gcc and clang versions when picking attributes.

edit: further reading suggests retain turned up in gcc 11 and requires binutils 2.36, with semantics similar to but distinct from used. It's a linker section discard directive, so the 'garbage collection roots' in this context may refer to bits of an elf instead of the language runtime feature.

I have fixed openmp by marking variables as retain but breaking applications that are relying on used (by discarding the variables) remains bad. Is this breakage already shipping with gcc, thus the ship has sailed, or can we keep backwards compat here?

edit: Would making attribute((used)) imply attribute((retain)) on elf targets achieve the objective of this patch without breaking code that expects 'used' to mean "don't throw this away"?

clang/lib/CodeGen/CGDecl.cpp
445

I think this (and the corresponding line in codgen) is incorrect.

Previously, attribute((used)) marked something as 'used', so it makes it all the way to the linker.

After this change, anything that answers getTriple().isOSBinFormatELF() with true will emit ((used)) as compiler.used, which means it gets deleted earlier. In particular, amdgpu uses elf and the openmp runtime marks some symbols used, which are now getting deleted by clang during internalise.

Lots of targets presumably use 'elf' as the target binary format though, so I expect this to have broken user facing code on all of them.

This revision is now accepted and ready to land.Jan 13 2022, 4:15 AM
rnk added a comment.Jan 13 2022, 10:10 AM

I don't know the context of this patch but changing attribute((used)) to put things under compiler.used is definitely a breaking change. Please introduce a new attribute if necessary for garbage collection purposes instead of breaking this.

This change was implemented so that llvm.used could prevent section GC on ELF, to make its semantics consistent with llvm.used on COFF and MachO. Essentially, llvm.used behaves differently on ELF now so to prevent behavior changes for users of __attribute__((used)), it was migrated to llvm.compiler.used on ELF targets. This is consistent with GCC's behavior.

This should only change behavior for you if you depend on the details of the LLVM IR being emitted, or perhaps if you use LTO, where GlobalOpt will behave differently. I don't think these use cases are as important as consistent ELF section GC behavior.

So, apologies for changing the LLVM IR emitted for this attribute, but I think it's unlikely we will change our minds and revert this a year later.

If I had a red buildbot to reference I would have reverted - the commit message does not make it clear that this is a breaking change. This broke a debugging hook in openmp, which is apparently not tested, and will break any application code that uses compiler.used on a target that uses elf.

Linked docs at https://reviews.llvm.org/D97447 suggest applications can get back to a working state by marking things as attribute((used)) attribute((retain)), presumably guarded by a test for whether retain exists. I think this would be another point in a codebase that has to distinguish between gcc and clang versions when picking attributes.

edit: further reading suggests retain turned up in gcc 11 and requires binutils 2.36, with semantics similar to but distinct from used. It's a linker section discard directive, so the 'garbage collection roots' in this context may refer to bits of an elf instead of the language runtime feature.

I have fixed openmp by marking variables as retain but breaking applications that are relying on used (by discarding the variables) remains bad. Is this breakage already shipping with gcc, thus the ship has sailed, or can we keep backwards compat here?

edit: Would making attribute((used)) imply attribute((retain)) on elf targets achieve the objective of this patch without breaking code that expects 'used' to mean "don't throw this away"?

clang/lib/CodeGen/CGDecl.cpp
445

This is the same behavior they would get in a native link if they used -ffunction-sections / --gc-sections, so you have described the desired behavior change: it makes LTO internalize+globalopt consistent with native links.

MaskRay added a comment.EditedJan 13 2022, 10:27 AM
In D97446#3241142, @rnk wrote:

If I had a red buildbot to reference I would have reverted - the commit message does not make it clear that this is a breaking change. This broke a debugging hook in openmp, which is apparently not tested, and will break any application code that uses compiler.used on a target that uses elf.

llvm.compiler.used hasn't been changed.

Linked docs at https://reviews.llvm.org/D97447 suggest applications can get back to a working state by marking things as attribute((used)) attribute((retain)), presumably guarded by a test for whether retain exists. I think this would be another point in a codebase that has to distinguish between gcc and clang versions when picking attributes.

edit: further reading suggests retain turned up in gcc 11 and requires binutils 2.36, with semantics similar to but distinct from used. It's a linker section discard directive, so the 'garbage collection roots' in this context may refer to bits of an elf instead of the language runtime feature.

The text focuses on the semantics, and for practical reasons refers to the toolchain support.
Before GCC 11/binutils 2.36 there was just no portable way making a definition not discarded by ld --gc-sections.
The patch series added the support but obviously cannot alter the fact that the toolchain was catching up.

I have fixed openmp by marking variables as retain but breaking applications that are relying on used (by discarding the variables) remains bad. Is this breakage already shipping with gcc, thus the ship has sailed, or can we keep backwards compat here?

edit: Would making attribute((used)) imply attribute((retain)) on elf targets achieve the objective of this patch without breaking code that expects 'used' to mean "don't throw this away"?

This would make semantics less orthogonal and incompatible with GCC.
On COFF and Mach-O, there have been Clang-specific (not GCC) use cases relying on attribute((used)) implying GC roots.
On ELF, there was none before the toolchain support.

Every llvm.used usage I can find in the wild does intend to have the GC semantics, and not having it on ELF was actually a bug and has been fixed by the patch series.

clang/lib/CodeGen/CGDecl.cpp
445

llvm.used traditionally did not have GC root semantics on ELF platforms, actually it was identical to llvm.compiler.used (modulo possible bugs). This is the intended change.

JonChesterfield added a comment.EditedJan 13 2022, 11:43 AM

It looks deeply wrong to be marking globals as either llvm.used or llvm.compiler.used based on the output file format. It should be based on the (purpose of) the global.

In D97446#3241142, @rnk wrote:

This change was implemented so that llvm.used could prevent section GC on ELF, to make its semantics consistent with llvm.used on COFF and MachO. Essentially, llvm.used behaves differently on ELF now so to prevent behavior changes for users of __attribute__((used)), it was migrated to llvm.compiler.used on ELF targets. This is consistent with GCC's behavior.

Is this sentence inverted? llvm.used should prevent sections from being discarded. If it doesn't at present, that's a bug in the linker. llvm.compiler.used should generally be discarded by whatever part of the compiler wanted the variable, but if it makes it to the linker, the linker should throw it away. Because it was only used by the compiler. It's possible some users marked things as 'used' but wanted them thrown away, but it seems more likely that users weren't using gc-sections if it broke their application by throwing away things they asked to keep.

This should only change behavior for you if you depend on the details of the LLVM IR being emitted, or perhaps if you use LTO, where GlobalOpt will behave differently. I don't think these use cases are as important as consistent ELF section GC behavior.

So, apologies for changing the LLVM IR emitted for this attribute, but I think it's unlikely we will change our minds and revert this a year later.

I'm hopeful I can change your minds now. The current modelling in the compiler doesn't match the stated intent.

llvm.compiler.used hasn't been changed.

True, but attribute((used)) has. I only noticed this patch because it broke some test cases in rocm, but it'll break some of my own code too. Moving a variable from llvm.used to llvm.compiler.used changes whether internalise skips the variable.

The text focuses on the semantics, and for practical reasons refers to the toolchain support.
Before GCC 11/binutils 2.36 there was just no portable way making a definition not discarded by ld --gc-sections.

Sure there was. Don't pass gc-sections to the linker, or don't compile with ffunction-sections to get a close approximation.

edit: Would making attribute((used)) imply attribute((retain)) on elf targets achieve the objective of this patch without breaking code that expects 'used' to mean "don't throw this away"?

This would make semantics less orthogonal and incompatible with GCC.
On COFF and Mach-O, there have been Clang-specific (not GCC) use cases relying on attribute((used)) implying GC roots.
On ELF, there was none before the toolchain support.

Every llvm.used usage I can find in the wild does intend to have the GC semantics, and not having it on ELF was actually a bug and has been fixed by the patch series.

I'm absolutely sure that people mark things as attribute((used)) to stop the toolchain discarding them. I think we're in agreement there, but differ in our assessment of popularity of gc-sections.

Are we missing a category here?

llvm.compiler.used <- the compiler uses the global, and may discard it. If it doesn't, the linker should discard it
llvm.linker.used <- the linker uses this global, and may discard it. The compiler should leave it alone aside from passing it to the linker
llvm.used <- some unspecified thing uses the global, the compiler and linker should leave it alone aside from embedding it in the linked output

If we have to map 'attribute((used))' onto the new llvm.linker.used and 'attribute((retain))' onto llvm.used that's a shame, but at least it keeps the naming weirdness localised to the language front end.

MaskRay added a comment.EditedJan 13 2022, 12:39 PM

It looks deeply wrong to be marking globals as either llvm.used or llvm.compiler.used based on the output file format. It should be based on the (purpose of) the global.

In D97446#3241142, @rnk wrote:

This change was implemented so that llvm.used could prevent section GC on ELF, to make its semantics consistent with llvm.used on COFF and MachO. Essentially, llvm.used behaves differently on ELF now so to prevent behavior changes for users of __attribute__((used)), it was migrated to llvm.compiler.used on ELF targets. This is consistent with GCC's behavior.

Is this sentence inverted? llvm.used should prevent sections from being discarded. If it doesn't at present, that's a bug in the linker.

I agree with rnk.
The bug can be on the compiler side which does not give enough information to the linker.
Before retain was added to GCC and binutils supported SHF_GNU_RETAIN, the ELF world was in that state.

llvm.compiler.used should generally be discarded by whatever part of the compiler wanted the variable

The compiler cannot discard it per LangRef https://llvm.org/docs/LangRef.html#the-llvm-compiler-used-global-variable

but if it makes it to the linker, the linker should throw it away. Because it was only used by the compiler. It's possible some users marked things as 'used' but wanted them thrown away, but it seems more likely that users weren't using gc-sections if it broke their application by throwing away things they asked to keep.

Correct. There are users not wanting __attribute__((used)) to remain a symbol. That was why retain was designed.

This should only change behavior for you if you depend on the details of the LLVM IR being emitted, or perhaps if you use LTO, where GlobalOpt will behave differently. I don't think these use cases are as important as consistent ELF section GC behavior.

So, apologies for changing the LLVM IR emitted for this attribute, but I think it's unlikely we will change our minds and revert this a year later.

I'm hopeful I can change your minds now. The current modelling in the compiler doesn't match the stated intent.

llvm.compiler.used hasn't been changed.

True, but attribute((used)) has. I only noticed this patch because it broke some test cases in rocm, but it'll break some of my own code too. Moving a variable from llvm.used to llvm.compiler.used changes whether internalise skips the variable.

Modulo optimizer bugs, __attribute__((used)) hasn't changed semantics.
If your downstream project does not handle llvm.compiler.used, maybe handle it now :)

I apologize that previous Clang probably very rarely emitted llvm.compiler.used and it started to do it more often now.

The text focuses on the semantics, and for practical reasons refers to the toolchain support.
Before GCC 11/binutils 2.36 there was just no portable way making a definition not discarded by ld --gc-sections.

Sure there was. Don't pass gc-sections to the linker, or don't compile with ffunction-sections to get a close approximation.

edit: Would making attribute((used)) imply attribute((retain)) on elf targets achieve the objective of this patch without breaking code that expects 'used' to mean "don't throw this away"?

This would make semantics less orthogonal and incompatible with GCC.
On COFF and Mach-O, there have been Clang-specific (not GCC) use cases relying on attribute((used)) implying GC roots.
On ELF, there was none before the toolchain support.

Every llvm.used usage I can find in the wild does intend to have the GC semantics, and not having it on ELF was actually a bug and has been fixed by the patch series.

I'm absolutely sure that people mark things as attribute((used)) to stop the toolchain discarding them. I think we're in agreement there, but differ in our assessment of popularity of gc-sections.

Are we missing a category here?

llvm.compiler.used <- the compiler uses the global, and may discard it. If it doesn't, the linker should discard it
llvm.linker.used <- the linker uses this global, and may discard it. The compiler should leave it alone aside from passing it to the linker
llvm.used <- some unspecified thing uses the global, the compiler and linker should leave it alone aside from embedding it in the linked output

I think the llvm.compiler.used interpretation diverges from LangRef and how we teach optimizers to respect llvm.compiler.used.
Then llvm.linker.used shall not be needed.

If we have to map 'attribute((used))' onto the new llvm.linker.used and 'attribute((retain))' onto llvm.used that's a shame, but at least it keeps the naming weirdness localised to the language front end.

I have checked the documentation of many GCC releases (including very ancient ones).
Its __attribute__((used)) never suggests that it has the linker GC semantics.
In LLVM/Clang, Mach-O somehow chose to overload used with the additional linker dead stripping semantics.

In my opinion, the ideal semantics is that COFF/Mach-O uses llvm.compiler.used for __attribute__((used)) as well.
But downgrading llvm.used to llvm.compiler.used may be a regression for some Mach-O code, so I chose not to do that for Mach-O.
(COFF has an extra limitation that internal linkage symbols currently don't work with llvm.used.)

I agree that there is unfortunate binary format inconsistency, but the ELF semantics as implemented in the patch series are ideal.

JonChesterfield added a comment.EditedJan 13 2022, 1:12 PM

Modulo optimizer bugs, attribute((used)) hasn't changed semantics.
If your downstream project does not handle llvm.compiler.used, maybe handle it now :)

There seems to be some confusion here. The 'downstream project' is openmp, which has worked around this regression in D117211 and D117231.

Before this patch, __attribute__((used)) mapped onto llvm.used, and the variables so annotated made it all the way to the compiled artefact. After this patch, it is mapped onto llvm.compiler.used, gets hit by an internalisation pass and ends up in the compiled output but missing from the symbol table. That is itself presumably a bug, as the linker should have completely discarded it, with much the same effect.

With this patch applied, what's the remaining use case for __attribute__((used))? It can no longer be used to keep something in the final executable, so it seems s/used/retain/g is the recommendation for programs that previously used that attribute.

It's possible that the check in internalize that skips over llvm.used and not llvm.compiler.used is itself a bug, and the intent is for llvm.used to be identical to llvm.compiler.used, but if that's the case we should delete the llvm.compiler.used array.

MaskRay added a comment.EditedJan 13 2022, 1:43 PM

Modulo optimizer bugs, attribute((used)) hasn't changed semantics.
If your downstream project does not handle llvm.compiler.used, maybe handle it now :)

There seems to be some confusion here. The 'downstream project' is openmp, which has worked around this regression in D117211 and D117231.

Before this patch, __attribute__((used)) mapped onto llvm.used, and the variables so annotated made it all the way to the compiled artefact. After this patch, it is mapped onto llvm.compiler.used, gets hit by an internalisation pass and ends up in the compiled output but missing from the symbol table. That is itself presumably a bug, as the linker should have completely discarded it, with much the same effect.

For your comment it appears an issue in the internalisation pass. It is orthogonal to this patch.
Do you have a reduced example with reproduce instructions for the issue? I know very little about OpenMP.
Well, I assume we can continuation the discussion in that OpenMP thread.

With this patch applied, what's the remaining use case for __attribute__((used))? It can no longer be used to keep something in the final executable, so it seems s/used/retain/g is the recommendation for programs that previously used that attribute.

It's possible that the check in internalize that skips over llvm.used and not llvm.compiler.used is itself a bug, and the intent is for llvm.used to be identical to llvm.compiler.used, but if that's the case we should delete the llvm.compiler.used array.

A __attribute__((used)) object can be referenced by a relocation. The relocation (from A to B) establishes a dependency relation in the linker: if the section defining A is retained, the section defining B should be retained as well.
This can be used by inline assembly (the compiler may not know the symbol references in assembly), and some code doing manual relocations.
I agree that in many cases user don't need more fine-grained GC precision and probably just add both used/retain to not bother think about the problem.
These people may find the split strange.

llvm.compiler.used (the underlying mechanism of __attribute__((used))) is useful in instrumentations.
There are quite few cases that the compiler does not fully discard definitions and has to defer it to the linker.
I have changed some instrumentations (PGO/SanitizerCoverage/other sanitizers) to downgrade llvm.used to llvm.compiler.used to improve section based linker garbage collection for all of PE-COFF, Mach-O, and ELF.
There has been decent object size decrease (at least single digit percent, 10+% for some).

I can understand how we got here, but it's a bad place to have ended up.

Toolchains that do dead-stripping need a way to prevent it for specific objects and functions. Windows and Darwin toolchains have historically done aggressive dead-stripping in both the compiler and linker, so they've always had this need. ELF toolchains have historically emitted code and restricted the linker in ways that, together, largely inhibit link-time dead-stripping (except of specific things like C++ entities with vague linkage). As a result, ELF has gotten away with not providing a fine-grained way to prevent link-time dead-stripping for a long time, when otherwise that would have been a major blocker.

There are basically three broad categories of reasons why programs need to use features like this:

  1. Some system at runtime needs to be able to find the entity "reflectively" (generally, because it's in a section with a special name or will be looked up by its symbol name). Usually this is intended as a kind of passive global registration, and so dead-stripping needs to be completely blocked for the entity.
  2. Some system at runtime that finds the entity "reflectively" will access it in weird ways that the compiler can't hope to understand. The attribute is not being used to prevent dead-stripping (unless the entity is *also* a passive registration), but to block compiler analysis and optimization if the value *does* end up being used. This tends to be a language-implementation thing more than something that a library would do.
  3. The entity isn't actually unreferenced, but some portion of the toolchain is just unable to see that. The most important example of this is a reference from inline assembly.

If I were a GCC developer reacting to the addition of retain to ELF, I would have given __attribute__((used)) the stronger semantics, forcing the entity to be retained at link-time, and I would have introduced a weaker attribute (or attributes) that people could use selectively in the second and third cases if they wanted better dead-stripping. To me, that's conservatively living up to user expectations about what they're trying to achieve with __attribute__((used)) while giving them an opportunity to be more precise if it's useful; and it also unifies the semantics across object formats now that it's possible to do so. But that's not generally how GCC developers do things; they tend to treat the current compiler behavior as an inviolate contract down to a very low level, and if you want different behavior, you need to use different options. I can understand and sympathize with that approach, even if I think it also tends to create situations like this one.

At the end of the day, I don't think we have much choice but to follow GCC's lead on ELF platforms. They get to define what these attributes mean, and if they want to make weaker guarantees on ELF, that's their decision.

JonChesterfield added a comment.EditedJan 13 2022, 2:17 PM

For your comment it appears an issue in the internalisation pass. It is orthogonal to this patch.
Do you have a reduced example with reproduce instructions for the issue? I know very little about OpenMP.
Well, I assume we can continuation the discussion in that OpenMP thread.

Internalize.cpp is fairly clear that it treats the two arrays differently, copying the corresponding part. Perhaps orthogonal to but exposed by this patch.

// We must assume that globals in llvm.used have a reference that not even
// the linker can see, so we don't internalize them.
// For llvm.compiler.used the situation is a bit fuzzy. The assembler and
// linker can drop those symbols. If this pass is running as part of LTO,
// one might think that it could just drop llvm.compiler.used. The problem
// is that even in LTO llvm doesn't see every reference. For example,
// we don't see references from function local inline assembly. To be
// conservative, we internalize symbols in llvm.compiler.used, but we
// keep llvm.compiler.used so that the symbol is not deleted by llvm.
for (GlobalValue *V : Used) {
  AlwaysPreserved.insert(V->getName());
}

It's possible that the check in internalize that skips over llvm.used and not llvm.compiler.used is itself a bug

^ probably is a bug now, even if it wasn't before

...
I agree that in many cases user don't need more fine-grained GC precision and probably just add both used/retain to not bother think about the problem.
These people may find the split strange.

llvm.compiler.used (the underlying mechanism of __attribute__((used))) is useful in instrumentations.
There are quite few cases that the compiler does not fully discard definitions and has to defer it to the linker.
I have changed some instrumentations (PGO/SanitizerCoverage/other sanitizers) to downgrade llvm.used to llvm.compiler.used to improve section based linker garbage collection for all of PE-COFF, Mach-O, and ELF.
There has been decent object size decrease (at least single digit percent, 10+% for some).

Uh, yes. Discarding things that previously were not discarded will make things smaller. The users I'm advocating for here are the ones who would have preferred we not discard the things that they asked us to keep and that we used to keep. Perhaps they will be few in number, and at least there's a workaround to be discovered.

I can understand how we got here, but it's a bad place to have ended up.
...elided...
At the end of the day, I don't think we have much choice but to follow GCC's lead on ELF platforms. They get to define what these attributes mean, and if they want to make weaker guarantees on ELF, that's their decision.

That's compelling, thank you for articulating it.

I think we have an outstanding issue to resolve in internalize, where the assumption behind this patch that the two forms are equivalent on elf does not hold.

I'd much refer we put variables in arrays corresponding to their intended lifespan instead of the binary format they're destined for. I don't know if the OS in question is certain to match the linker output either, seems possible to compile on an OS that usually uses one format but emit code in a different one.

clang/lib/CodeGen/CodeGenModule.cpp
2093

^ this specifically looks wrong, which array the variable goes in should be based on what the variable is used for or what the programmer asked for, not on the binary format used by the OS (is that even a unique test? One can run elf or coff on windows iirc)

MaskRay marked 2 inline comments as done.Jan 13 2022, 2:54 PM
MaskRay added inline comments.
clang/lib/CodeGen/CodeGenModule.cpp
2093

addUsedOrCompilerUsedGlobal is carefully used on __attribute__((used)) and the related keep-static-consts.cpp.

__attribute__((used)) has a semantic split on ELF and non-ELF, so the difference is unavoidable. For other constructs we try to use the precise LLVMCompilerUsed or LLVMUsed.

I'd much refer we put variables in arrays corresponding to their intended lifespan instead of the binary format they're destined for.

I agree that LLVM features should generally mean something consistent across targets, and frontends should use the appropriate feature for the semantics they require. Here the problem is that, for better or worse, the source language has different semantics on different targets. Historically, that really has been driven by object format and the corresponding presumed linker behavior; I don't know where to begin answering the question of whether we should treat it as a platform difference instead when targeting a non-standard object format.

MaskRay closed this revision.Jan 13 2022, 4:37 PM
MaskRay marked an inline comment as done.