This is an archive of the discontinued LLVM Phabricator instance.

[RFC][ELF] Refactor relocation processing
Needs ReviewPublic

Authored by MaskRay on Nov 4 2021, 4:46 PM.

Details

Reviewers
peter.smith
Summary

Current relocation processing does:

if (the relocation expr belongs to category A)
  do something;
if (the relocation expr belongs to category B)
  do something;
if (the relocation expr belongs to category C)
  do something;

This refactor flattens it into

switch (expr) {
case R_PC:
case R_...:
  do something;
case R_...:
  do something;
...
}

which may potentially improve performance.
In addition, if some architectures have strange logic, we can add a new RelExpr member and make it interrupt less to the generic code.

Diff Detail

Event Timeline

MaskRay created this revision.Nov 4 2021, 4:46 PM
MaskRay requested review of this revision.Nov 4 2021, 4:46 PM
Herald added a project: Restricted Project. · View Herald TranscriptNov 4 2021, 4:46 PM

Will take a look later this week, hopefully tomorrow. Just got back into the office and have a bit of a backlog of stuff to go through.

On a first look I like the way this is going. There is more information in a single place. I've made a few suggestions but none are particularly important for an RFC. If I get time I'll keep thinking to see if I can think of any other suggestions.

lld/ELF/Relocations.cpp
224

Is it worth moving closer to the only caller processRelocAux. Perhaps worth emphasisng that this only handles relocations that are expected in processRelocaAux. Could do with a comment or perhaps renaming to something like customRelocIsStaticLinkTimeConstant

If there were a way to know if a particular expr code could be CUSTOM outside of scanReloc then we could assert that which would make it clear, but I'm not sure if there is an easily maintainable way to do that.

227

I'm guessing that some of the entries have been taken out as the relocations were not CUSTOM so they would never expect to be called?

Perhaps worth changing the comment
// The following expressions always compute a constant. The list does not contain expressions marked RELOCATE in scanReloc.

952

suggest cantUseRelocationAgainstLocal or invalidRelocToLocal

1161

Perhaps worth renaming to handleTlsNonIE

1371

This is also in handleTlsRelocation, worth making an inline function? Something like toExecRelax()

1444

Would be a shame to lose the information in the comment, even if it had to be moved/split up.

MaskRay updated this revision to Diff 386053.Nov 9 2021, 9:07 PM
MaskRay marked 6 inline comments as done.

Thanks for the comments. Applies changes.

I put up this patch because I think our relocation processing still has some problems
wand want explore how we should organize relocations.

lld/ELF/Relocations.cpp
224

Thanks for the suggestion. I just pushed the isStaticLinkTimeConstant move and comment adding separately.

952

Used invalidRelocAgainstSym.

I was thinking of extracting this out as a separate function to be shared, but currently it is only used once. I may inline it again.

I feel like there is a lot going on in this patch. It may be easier if it was just done little by little.
For example: toExecRelax is now a static function. This can be separated out into a small NFC patch where the only thing that is done is the refactoring of that one static function.
Also, handleTlsRelocation returns a slightly different value (what should be skipped vs what was processed). This can also be done as a separate mini patch.

Anyway, those are my 2 cents.

lld/ELF/Relocations.cpp
1267–1269

Do we need this } and the return 0; ?
The if has been changed to an assert so now we just return return target->getTlsGdRelaxSkip(type) - 1;.

You can also remove the starting }.

MaskRay added a comment.EditedNov 10 2021, 2:28 PM

I feel like there is a lot going on in this patch. It may be easier if it was just done little by little.
For example: toExecRelax is now a static function. This can be separated out into a small NFC patch where the only thing that is done is the refactoring of that one static function.
Also, handleTlsRelocation returns a slightly different value (what should be skipped vs what was processed). This can also be done as a separate mini patch.

Anyway, those are my 2 cents.

Structurally it's difficult to do it piecemeal. This patch explores an alternative way handling relocations.
If we consider the current state and the exploration local optima, some changes (like moving toExecRelax outside) are actually strictly inferior (if we don't split handleTlsRelocations).

For minor details, some changes can be split and I've pushed such commits separately (e.g. moving isStaticLinkTimeConstant).

I have spent much time on this patch (some architectures like Mips and PPC64 add significant complexity) but not entirely decided this is the way to go.

It's about a year later; I'm curious if you've gained any insight into whether this is a good direction or not.

Herald added a project: Restricted Project. · View Herald TranscriptOct 19 2022, 4:46 PM

It's about a year later; I'm curious if you've gained any insight into whether this is a good direction or not.

I think there is value in optimizing relocation scanning. For some workload lld's pass takes 2.x time as mold's.
InputSectionBase::relocations takes some time but other checks contribute to the other slowdown.

We probably need to be more aggressive and possibly moving more relocation scanning into arch-specific ELF/Arch/*.cpp.