[IR] Ignore bitcasts of function pointers which are only used as callees in callbase instruction

[madhur13490]

Summary

This patch enhances hasAddressTaken() to ignore bitcasts as a
callee in callbase instruction. Such bitcast usage doesn't really take
the address in a useful meaningful way.

[madhur13490]

llvm/include/llvm/IR/Function.h

879 ↗

(On Diff #331651)

Done by lint.

llvm/lib/IR/Function.cpp

1628

Done by lint

[lebedev.ri]

What problem are you trying to solve?

[rampitec]

llvm/lib/IR/Function.cpp

1643

Should add test that is a callee operand. Also need to add test where it is an argument in a call.

[rampitec]

llvm/test/Analysis/CallGraph/ignore-bitcast-callees.ll

12

remove #1, here and below.

[arsenm]

Users can just call stripPointerCasts first.No need to add a special case here(which is also incoplete)

[rampitec]

This test also does not really work until you add true argument to the call graph builder.

[arsenm]

I'm not sure it's worth having this be an option. You can just always do this, or not

llvm/lib/IR/Function.cpp

1641

Also should restrict to constant expression case. Also need to verify it is a call operand

[madhur13490]

In D98884#2635456, @lebedev.ri wrote:

What problem are you trying to solve?

For AMDGPU target, we want to add a bunch of attributes to those functions which are indirectly called. To call a function indirectly, the programmer first needs to capture its address in a function pointer and use the function pointer in a call instruction. Many times the address is taken but not really stored anywhere and thus not used meaningfully or in any way which can be de-referenced. One such example is usage of bitcasts in call instruction as shown in the test case. hasAddressTaken() API is useful for the purpose but we want ignore such “false” address taken cases whose the address is not really used. This patch adds an optional flag to ignore those cases. As it stands today, hasAddressTaken() returns "true" for the test case, which is correct but skipping the test scenario is useful for us.

I hope this clarifies.

[madhur13490]

address review comments

[madhur13490]

In D98884#2635600, @rampitec wrote:

This test also does not really work until you add true argument to the call graph builder.

I am not sure if I follow. I don't intend do any changes (at least for current patch) in the way call graph builder works because that will break a lot of things. (I tried adding "true" for call graph builder and call changed which was obvious.) FileCheck commands are just inline with what this opt with this patch would emit. Referring to https://reviews.llvm.org/D96087, would those test cases too won't work?

[madhur13490]

remove unnecessary changes

[jdoerfert]

Idea is fine. We don't want a flag here though.

Bitcasts of a function pointer that are only used as *callees* do not take the address of the function pointer.

[rampitec]

In D98884#2638318, @madhur13490 wrote:

In D98884#2635600, @rampitec wrote:

This test also does not really work until you add true argument to the call graph builder.

I am not sure if I follow. I don't intend do any changes (at least for current patch) in the way call graph builder works because that will break a lot of things. (I tried adding "true" for call graph builder and call changed which was obvious.) FileCheck commands are just inline with what this opt with this patch would emit. Referring to https://reviews.llvm.org/D96087, would those test cases too won't work?

The problem is that test does not exercise your code unless it is enabled, so it does not test anything.
Remove the option though and do it always, a direct call through bitcast does not take function's address. That will make test functional as well.
Also a second call to the same bitcast which will test that all_of() in case it has more than one use.

[madhur13490]

rebase, address review comments to remove the optiona flag, update summary, rename and fix tests.

[madhur13490]

revert unnecessary changes to comments.

[rampitec]

llvm/include/llvm/IR/Function.h

887 ↗

(On Diff #332208)

Revert this change, you are not touching this file anymore.

llvm/test/Analysis/CallGraph/ignore-bitcast-call-argument-callee.ll

13

I do not like that CG does not see a call to foo, but that's probably a separate change.

llvm/test/Analysis/CallGraph/ignore-bitcast-callees.ll

17

Can you add a second identical call to the same bitcast? So that you have test where the same bitcast is used twice.

[madhur13490]

rebase, remove function.h, add one more test

[rampitec]

LGTM, thanks!

[arsenm]

Commit message needs update

[madhur13490]

In D98884#2644969, @arsenm wrote:

Commit message needs update

I updated it but I don't why phab does not show it. Here is my new commit message:

[IR] Ignore bitcasts of function pointers which are only used as callees in callbase instruction

This patch enhances hasAddressTaken() to ignore bitcasts as a
callee in callbase instruction. Such bitcast usage don't really take
the address in a useful meaningful way.

Differential Revision: https://reviews.llvm.org/D98884/

[madhur13490]

@rampitec, @jdoerfert @arsenm

There are a lot of tests failing as shown by the buildbot on this page. I tried to fix a couple of them but to fix some of them I need to change the code as well. For example, to fix Transforms/GlobalOpt/evaluate-call.ll, I should fix casting in ChangeCalleesToFastCall in GlobalOpt.cpp because cast<CallBase>(U) fails as the user is now BitCastOperator. For more details, you can see the backtrace reported by the buildbot.

I am thinking of reintroducing the flag again because this change may lead to unknown territory and surprising behaviour for some users. Flag would minimize the impact. We should remove the flag as a separate change.

What do you think?

[rampitec]

In D98884#2645382, @madhur13490 wrote:

@rampitec, @jdoerfert @arsenm

There are a lot of tests failing as shown by the buildbot on this page. I tried to fix a couple of them but to fix some of them I need to change the code as well. For example, to fix Transforms/GlobalOpt/evaluate-call.ll, I should fix casting in ChangeCalleesToFastCall in GlobalOpt.cpp because cast<CallBase>(U) fails as the user is now BitCastOperator. For more details, you can see the backtrace reported by the buildbot.

I am thinking of reintroducing the flag again because this change may lead to unknown territory and surprising behaviour for some users. Flag would minimize the impact. We should remove the flag as a separate change.

What do you think?

Maybe just strip bitcasts there?

[arsenm]

In D98884#2645382, @madhur13490 wrote:

@rampitec, @jdoerfert @arsenm

There are a lot of tests failing as shown by the buildbot on this page. I tried to fix a couple of them but to fix some of them I need to change the code as well. For example, to fix Transforms/GlobalOpt/evaluate-call.ll, I should fix casting in ChangeCalleesToFastCall in GlobalOpt.cpp because cast<CallBase>(U) fails as the user is now BitCastOperator. For more details, you can see the backtrace reported by the buildbot.

I am thinking of reintroducing the flag again because this change may lead to unknown territory and surprising behaviour for some users. Flag would minimize the impact. We should remove the flag as a separate change.

No, don't add cruft due to FUD

[jdoerfert]

In D98884#2645401, @arsenm wrote:

In D98884#2645382, @madhur13490 wrote:

@rampitec, @jdoerfert @arsenm

There are a lot of tests failing as shown by the buildbot on this page. I tried to fix a couple of them but to fix some of them I need to change the code as well. For example, to fix Transforms/GlobalOpt/evaluate-call.ll, I should fix casting in ChangeCalleesToFastCall in GlobalOpt.cpp because cast<CallBase>(U) fails as the user is now BitCastOperator. For more details, you can see the backtrace reported by the buildbot.

I am thinking of reintroducing the flag again because this change may lead to unknown territory and surprising behaviour for some users. Flag would minimize the impact. We should remove the flag as a separate change.

No, don't add cruft due to FUD

I'm totally with @arsenm here. You will have to adjust the users to cope with this, arguably it improves our analysis and optimization capabilities and is the entire point of this patch.
Actually looking at the effect also minimizes the risk of corner cases we forgot.

[madhur13490]

fix failing tests reported by bot, handle globalopt when bitcast callees present

[madhur13490]

In D98884#2645533, @jdoerfert wrote:

In D98884#2645401, @arsenm wrote:

In D98884#2645382, @madhur13490 wrote:

@rampitec, @jdoerfert @arsenm

There are a lot of tests failing as shown by the buildbot on this page. I tried to fix a couple of them but to fix some of them I need to change the code as well. For example, to fix Transforms/GlobalOpt/evaluate-call.ll, I should fix casting in ChangeCalleesToFastCall in GlobalOpt.cpp because cast<CallBase>(U) fails as the user is now BitCastOperator. For more details, you can see the backtrace reported by the buildbot.

I am thinking of reintroducing the flag again because this change may lead to unknown territory and surprising behaviour for some users. Flag would minimize the impact. We should remove the flag as a separate change.

No, don't add cruft due to FUD

I'm totally with @arsenm here. You will have to adjust the users to cope with this, arguably it improves our analysis and optimization capabilities and is the entire point of this patch.
Actually looking at the effect also minimizes the risk of corner cases we forgot.

Okay, please review further changes.

[madhur13490]

In D98884#2645239, @madhur13490 wrote:

In D98884#2644969, @arsenm wrote:

Commit message needs update

I updated it but I don't why phab does not show it. Here is my new commit message:

[IR] Ignore bitcasts of function pointers which are only used as callees in callbase instruction

This patch enhances hasAddressTaken() to ignore bitcasts as a
callee in callbase instruction. Such bitcast usage don't really take
the address in a useful meaningful way.

Differential Revision: https://reviews.llvm.org/D98884/

Updated summary manually on phab.

[jdoerfert]

llvm/lib/Transforms/IPO/GlobalOpt.cpp

2161

Now we could even have a helper of some sort. Unsure what design is best.

llvm/test/Transforms/Attributor/IPConstantProp/arg-count-mismatch.ll

6

Hm, if you just run the update script the tests should be properly updated, did you do that?

llvm/test/Transforms/Attributor/IPConstantProp/arg-type-mismatch.ll

5

Same as above but I do not understand why this change would cause us to loose the function attributes if no AA is provided.

[madhur13490]

use script to update the test, add utility function and address other parts of global opt to fix clang tests

[madhur13490]

llvm/lib/Transforms/IPO/GlobalOpt.cpp

2161

Sounds fine. Outlined in a separate function.

llvm/test/Transforms/Attributor/IPConstantProp/arg-count-mismatch.ll

6

Done now, Didn't know the script is so powerful!

[madhur13490]

@jdoerfert This patch is blocking https://reviews.llvm.org/D99347 for AMDGPU. Please have a look.

[arsenm]

llvm/lib/Transforms/IPO/GlobalOpt.cpp

2135–2139

Just use cast<>

2166–2170

Just use cast<>

[madhur13490]

llvm/lib/Transforms/IPO/GlobalOpt.cpp

2166–2170

Are you suggesting to remove if (!CB) code block?

If so, how would it work in release builds? If cast<> fails, in debug builds, the assert would fire but in release builds, the code will try to access the member function because we don't have "return" statement. Does failure on cast<> return?

[arsenm]

llvm/lib/Transforms/IPO/GlobalOpt.cpp

2166–2170

cast<> asserts in a debug build. dyn_cast is when you are trying to handle the null case

[jdoerfert]

In D98884#2650715, @madhur13490 wrote:

@jdoerfert This patch is blocking https://reviews.llvm.org/D99347 for AMDGPU. Please have a look.

I'm looking and it seems there is more content added every revision which makes it a moving target.

Also, if you don't get feedback, it is custom to wait for a week to ping: https://www.llvm.org/docs/Contributing.html#how-to-submit-a-patch

llvm/lib/Transforms/IPO/GlobalOpt.cpp

2135–2139

This can at some point error out. Why would the user of ChangeCalleesToFastCall and others need to require this property?

2155

This part was not in the original patch and is now tagged along. At the very least add it to the commit message. It is also unclear why you chose these two functions, are there no other users() loops?

2187

The function is called remove... and the helper set...

llvm/test/Analysis/CallGraph/ignore-bitcast-call-argument-callee.ll

13

I run this with trunk and the output looks the same. What exactly is this testing?

llvm/test/Transforms/GlobalOpt/bitcast-callees-fastcc.ll

18

Also unclear what is tested here? Looks like the trunk output.

[madhur13490]

use cast<> instead of dyn_cast<>

[rampitec]

Since it keeps changing revoking the accept vote for now.

[madhur13490]

llvm/lib/Transforms/IPO/GlobalOpt.cpp

2135–2139

I think I am missing something here. This function is capturing common logic used by ChangeCalleesToFastCall.

2155

This fix is needed because GlobalOpt/evaluate-call-errors.ll failed with an assert as now OptimizeFunctions() is making a call to this function because of the change in return value by hasAddressTaken().

Apologies, I should have handled this earlier.

2187

Well, the RemoveAttribute is indeed setting attributes by making call to setAttributes() and the utility function is just wrapping the commonly used code. Should we rename the function?

llvm/test/Analysis/CallGraph/ignore-bitcast-call-argument-callee.ll

13

Yeah, output is same. But the test case shows that hasAddressTaken() return value is indeed correct when there is a direct call as well as bitcast callees is used as an argument as is used by call graph builder. The similar tests in Analysis/CallGraph,check only the call graph.

llvm/test/Transforms/GlobalOpt/bitcast-callees-fastcc.ll

18

This test demonstrates that the amendments made in this patch would actually set "fastcc" on bitcast callees. With trunk, "fastcc" would not be set on "foo" because hasAddressTaken() is true when queried by OptimizeFunctions() in GlobalOpt.cpp

[madhur13490]

The bot shows failure of bad-signature.c at https://reviews.llvm.org/B95560. When ran locally it does report the correct values and thus pass. I am not sure how to fix the test failure.
Failure in tile-and-distribute.mlir seems unrelated to this change.

@jdoerfert/ @arsenm / @rampitec , please advise.

[arsenm]

In D98884#2651396, @madhur13490 wrote:

The bot shows failure of bad-signature.c at https://reviews.llvm.org/B95560. When ran locally it does report the correct values and thus pass. I am not sure how to fix the test failure.
Failure in tile-and-distribute.mlir seems unrelated to this change.

@jdoerfert/ @arsenm / @rampitec , please advise.

The bots are imperfect, I would say it's unrelated

[madhur13490]

In D98884#2651450, @arsenm wrote:

In D98884#2651396, @madhur13490 wrote:

The bot shows failure of bad-signature.c at https://reviews.llvm.org/B95560. When ran locally it does report the correct values and thus pass. I am not sure how to fix the test failure.
Failure in tile-and-distribute.mlir seems unrelated to this change.

@jdoerfert/ @arsenm / @rampitec , please advise.

The bots are imperfect, I would say it's unrelated

Thanks. This patch now addresses all known failures and thus I don't expect it to be a moving target from now.

[rampitec]

llvm/lib/Transforms/IPO/GlobalOpt.cpp

2154

I do not see why a user of bitcast must be necessarily a call.

2184

Same here, it is not necessarily a call.

[arsenm]

llvm/lib/Transforms/IPO/GlobalOpt.cpp

2154

This could be a function in a constant initializer for a global

[rampitec]

llvm/lib/Transforms/IPO/GlobalOpt.cpp

2154

Right, that's what I mean. It could be anything, even a compare.

[madhur13490]

go back to dyn_cast to handle all users

[madhur13490]

llvm/lib/Transforms/IPO/GlobalOpt.cpp

2154

Right, now I remember why I had dyn_cast and return in previous diff.

[fhahn]

llvm/lib/Transforms/IPO/GlobalOpt.cpp

2154

Is there a test where the user is not a bit cast ? If not, please add one, similar for the other places that were changed back to dyn_cast

[rampitec]

llvm/lib/Transforms/IPO/GlobalOpt.cpp

2154

You do not check that user is a callee.

[rampitec]

I think this GlobalOpt patch is a separate change from the original one and needs to go first.

llvm/lib/Transforms/IPO/GlobalOpt.cpp

2154

You do not check that user is a callee.

And the original code does not check it either. This also warrants a test.

[madhur13490]

specifically check for callee

[madhur13490]

llvm/lib/Transforms/IPO/GlobalOpt.cpp

2154

You do not check that user is a callee.

Done. I think we can introduce a utility function `CallBase* isCallBaseCallee(Use &U)` which can outline the below checks

User *UU = U.getUser();
 CallBase *CB = dyn_cast<CallBase>(UU);
 if (!CB)
   return nullptr;
 if (!CB->isCallee(&U))
   return nullptr;
 return CB

Not sure if this is an overkill.

[rampitec]

This should work I guess. It needs tests for non-callee use and non call base uses. I would also prefer to GlobalOpt change to be extracted into a separate patch.

[rampitec]

llvm/lib/Transforms/IPO/GlobalOpt.cpp

2154

You now have it in 3 different places, so it may make sense. It would look better if you combine conditions like '!CB || !CB->isCallee(&U))' and also 'isSomething' function shall return bool, not a pointer.

[madhur13490]

In D98884#2653291, @rampitec wrote:

This should work I guess. It needs tests for non-callee use and non call base uses. I would also prefer to GlobalOpt change to be extracted into a separate patch.

non call base users - What all such cases possible?We can use invoke but that is also CallBase.
non callee - ignore-bitcast-call-argument-callee.ll is similar to what to suggest.

[rampitec]

In D98884#2653361, @madhur13490 wrote:

In D98884#2653291, @rampitec wrote:

This should work I guess. It needs tests for non-callee use and non call base uses. I would also prefer to GlobalOpt change to be extracted into a separate patch.

non call base users - What all such cases possible?We can use invoke but that is also CallBase.

That is more than possible. Take a pointer and store it, compare to another pointer, bitcast and store or compare.

non callee - ignore-bitcast-call-argument-callee.ll is similar to what to suggest.

Yes, but this is not for global opt.

[madhur13490]

In D98884#2653366, @rampitec wrote:

In D98884#2653361, @madhur13490 wrote:

In D98884#2653291, @rampitec wrote:

This should work I guess. It needs tests for non-callee use and non call base uses. I would also prefer to GlobalOpt change to be extracted into a separate patch.

non call base users - What all such cases possible?We can use invoke but that is also CallBase.

That is more than possible. Take a pointer and store it, compare to another pointer, bitcast and store or compare.

Yes, but we should also consider about hasAddressTaken(). I am now starting to think if hasAddressTaken() needs to be more matured. Technically, you can cast a function pointer to global i8* and do all operations you suggest (e.g. use it as an operand in load) but hasAddressTaken() would return true in such cases because I am considering only CallBase users. If it needs to be more matured then we need to consider *all* instructions which take address as one of the operands.

non callee - ignore-bitcast-call-argument-callee.ll is similar to what to suggest.

Yes, but this is not for global opt.

[rampitec]

In D98884#2653399, @madhur13490 wrote:

In D98884#2653366, @rampitec wrote:

In D98884#2653361, @madhur13490 wrote:

In D98884#2653291, @rampitec wrote:

This should work I guess. It needs tests for non-callee use and non call base uses. I would also prefer to GlobalOpt change to be extracted into a separate patch.

non call base users - What all such cases possible?We can use invoke but that is also CallBase.

That is more than possible. Take a pointer and store it, compare to another pointer, bitcast and store or compare.

Yes, but we should also consider about hasAddressTaken(). I am now starting to think if hasAddressTaken() needs to be more matured. Technically, you can cast a function pointer to global i8* and do all operations you suggest (e.g. use it as an operand in load) but hasAddressTaken() would return true in such cases because I am considering only CallBase users. If it needs to be more matured then we need to consider *all* instructions which take address as one of the operands.

Not necessarily. We ignore some uses in hasAddressTaken() and that can leak to this place.

[madhur13490]

add new test

[madhur13490]

llvm/lib/Transforms/IPO/GlobalOpt.cpp

2154

blockaddr-bitcast-fastcc.ll test address this.

llvm/test/Transforms/GlobalOpt/assumelike-bitcast-fastcc.ll

2

This test case demonstates that fastcc is not set on function bar because its address taken value is true as hasAddressTaken() is used with its default param values.

llvm/test/Transforms/GlobalOpt/bitcast-call-argument-fastcc.ll

2

Test case when bitcast is a call argument.

llvm/test/Transforms/GlobalOpt/blockaddr-bitcast-fastcc.ll

2

This test case demonstrates that fastcc is now set on function "g" unlike trunk which skips this function. This test also covers non call bases users case.

[foad]

llvm/lib/IR/Function.cpp

1643

This does not respect IgnoreCallbackUses, IgnoreAssumeLikeCalls, IgnoreLLVMUsed. Also it does not handle bitcasts-of-bitcasts.

Wouldn't the whole function be better implemented as a worklist algorithm now?

SmallVector<Use &> WorkList(use_begin(), use_end());
while (!WorkList.empty()) {
  U = WorkList.pop_back_val();
  if (isa<BitCast>(U))
    WorkList.push_back(U.use_begin(), U.use_end());
  else if (isa<CallBase>(U))
    ... all the existing logic that handles Ignore* flags ...
}

[rampitec]

It looks fine to me with a style changes in couple places and positive checks in new tests.

llvm/lib/IR/Function.cpp

1643

In general yes, but this patch is growing. This would be a good followup change.
The pass calls hasAddressTaken() without arguments so that is not exerciseable here as is. For the same reason I withdraw my request to split the patch. Also note this may change if this is landed: D96179.

llvm/lib/Transforms/IPO/GlobalOpt.cpp

2137

A readbility change would be nice:

if (!CB && !CB->isCallee(&U))
  return

2167

Same here: if (!CB && !CB->isCallee(&U))

llvm/test/Transforms/GlobalOpt/assumelike-bitcast-fastcc.ll

14

You need to check something positive. You can just check it is 'define internal void @bar() {'.

llvm/test/Transforms/GlobalOpt/bitcast-call-argument-fastcc.ll

10

Same here.

[foad]

llvm/lib/IR/Function.cpp

1643

I disagree. You're changing a core API Function::hasAddressTaken in a way that could change the behaviour for all existing callers, whether they like it or not. So I think this part probably should be a separate patch and the implementation should be as rigorous as possible.

[madhur13490]

llvm/lib/IR/Function.cpp

1643

I disagree. You're changing a core API Function::hasAddressTaken in a way that could change the behaviour for all existing callers, whether they like it or not.

In general sense I do agree agree about the impact and that is why the first diff of this patch had an option to have the new code under an optional flag. Please have a look at the diff and the followup discussion on it.

So I think this part probably should be a separate patch and the implementation should be as rigorous as possible.

I don't understand this part. Worklist based algorithm is more readble and maintenable but that is another way of implementing what this patch does and is not going to do anything different than this does functionality wise. If there is some functionality missing in this patch, then that's a different story. But I don't see that being a reason to block further progress on this patch.

[foad]

llvm/lib/IR/Function.cpp

1643

If there is some functionality missing in this patch, then that's a different story.

The code you added to check callers-via-bitcasts does not respect the Ignore* flags, which could cause problems for existing users of this function. The worklist thing was just a suggestion for how to fix that (and also handle bitcasts-of-bitcasts-... in a natural way, though I'm not sure if they ever occur in practice).

[foad]

llvm/lib/IR/Function.cpp

1643

Sorry, I was confused. I see now that you're just adding yet more cases which don't count as "taking the address" of the function, which shouldn't cause any problems. I still think this function is getting too complicated and could do with refactoring.

[madhur13490]

Addressed Stas' comments

[madhur13490]

llvm/test/Transforms/GlobalOpt/assumelike-bitcast-fastcc.ll

14

Added check for foo as well.

[rampitec]

LGTM with a nit in 2 tests. Please also wait for other responses.

llvm/test/Transforms/GlobalOpt/bitcast-call-argument-fastcc.ll

11

Make a positive check for foo.

22

Also need positive check for test.

[madhur13490]

Ping!

[jdoerfert]

Seems reasonable now, haven't looked into the Attributor test changes but I expect unrelated side effects. I'm not blocking this anymore.

[madhur13490]

make positive checks

[glaubitz]

This change broke the native build on x32 (x86_64 with 32-bit pointers) for me:

[46/3990] Building C object projects/compiler-rt/lib/builtins/CMakeFiles/clang_rt.builtins-x86_64.dir/cpu_model.c.o
FAILED: projects/compiler-rt/lib/builtins/CMakeFiles/clang_rt.builtins-x86_64.dir/cpu_model.c.o
/glaubitz/llvm-project/stage1.install/bin/clang -DVISIBILITY_HIDDEN -D_DEBUG -D_FILE_OFFSET_BITS=64 -D_GNU_SOURCE -D_LARGEFILE_SOURCE -D__STDC_CONSTANT_MACROS -D__STDC_FORMAT_MACROS -D__STDC_LIMIT_MACROS -Iprojects/compiler-rt/lib/builtins -I/glaubitz/llvm-project/compiler-rt/lib/builtins -Iinclude -I/glaubitz/llvm-project/llvm/include -fPIC -Werror=date-time -Werror=unguarded-availability-new -Wall -Wextra -Wno-unused-parameter -Wwrite-strings -Wmissing-field-initializers -pedantic -Wno-long-long -Wc++98-compat-extra-semi -Wimplicit-fallthrough -Wcovered-switch-default -Wdelete-non-virtual-dtor -Wstring-conversion -fdiagnostics-color -ffunction-sections -fdata-sections -Wall -Wno-unused-parameter -O3   -m64 -fno-lto -std=c11 -UNDEBUG -MD -MT projects/compiler-rt/lib/builtins/CMakeFiles/clang_rt.builtins-x86_64.dir/cpu_model.c.o -MF projects/compiler-rt/lib/builtins/CMakeFiles/clang_rt.builtins-x86_64.dir/cpu_model.c.o.d -o projects/compiler-rt/lib/builtins/CMakeFiles/clang_rt.builtins-x86_64.dir/cpu_model.c.o -c /glaubitz/llvm-project/compiler-rt/lib/builtins/cpu_model.c
clang: /glaubitz/llvm-project/llvm/include/llvm/Support/Casting.h:269: typename llvm::cast_retty<X, Y*>::ret_type llvm::cast(Y*) [with X = llvm::CallBase; Y = llvm::User; typename llvm::cast_retty<X, Y*>::ret_type = llvm::CallBase*]: Assertion `isa<X>(Val) && "cast<Ty>() argument of incompatible type!"' failed.
PLEASE submit a bug report to https://bugs.llvm.org/ and include the crash backtrace, preprocessed source, and associated run script.
Stack dump:
0.      Program arguments: /glaubitz/llvm-project/stage1.install/bin/clang -DVISIBILITY_HIDDEN -D_DEBUG -D_FILE_OFFSET_BITS=64 -D_GNU_SOURCE -D_LARGEFILE_SOURCE -D__STDC_CONSTANT_MACROS -D__STDC_FORMAT_MACROS -D__STDC_LIMIT_MACROS -Iprojects/compiler-rt/lib/builtins -I/glaubitz/llvm-project/compiler-rt/lib/builtins -Iinclude -I/glaubitz/llvm-project/llvm/include -fPIC -Werror=date-time -Werror=unguarded-availability-new -Wall -Wextra -Wno-unused-parameter -Wwrite-strings -Wmissing-field-initializers -pedantic -Wno-long-long -Wc++98-compat-extra-semi -Wimplicit-fallthrough -Wcovered-switch-default -Wdelete-non-virtual-dtor -Wstring-conversion -fdiagnostics-color -ffunction-sections -fdata-sections -Wall -Wno-unused-parameter -O3 -m64 -fno-lto -std=c11 -UNDEBUG -MD -MT projects/compiler-rt/lib/builtins/CMakeFiles/clang_rt.builtins-x86_64.dir/cpu_model.c.o -MF projects/compiler-rt/lib/builtins/CMakeFiles/clang_rt.builtins-x86_64.dir/cpu_model.c.o.d -o projects/compiler-rt/lib/builtins/CMakeFiles/clang_rt.builtins-x86_64.dir/cpu_model.c.o -c /glaubitz/llvm-project/compiler-rt/lib/builtins/cpu_model.c
1.      <eof> parser at end of file
2.      Optimizer
 #0 0x01b23df2 llvm::sys::PrintStackTrace(llvm::raw_ostream&, int) (/glaubitz/llvm-project/stage1.install/bin/clang+0x1b23df2)
 #1 0x01b21e10 llvm::sys::RunSignalHandlers() (/glaubitz/llvm-project/stage1.install/bin/clang+0x1b21e10)
 #2 0x01a8ea17 CrashRecoverySignalHandler(int) CrashRecoveryContext.cpp:0:0
 #3 0xfffffffff7f021e0 __restore_rt (/lib/x86_64-linux-gnux32/libpthread.so.0+0x131e0)
 #4 0xfffffffff7a2886d raise (/lib/x86_64-linux-gnux32/libc.so.6+0x3286d)
 #5 0xfffffffff7a12515 abort (/lib/x86_64-linux-gnux32/libc.so.6+0x1c515)
 #6 0xfffffffff7a123fe (/lib/x86_64-linux-gnux32/libc.so.6+0x1c3fe)
 #7 0xfffffffff7a211ff (/lib/x86_64-linux-gnux32/libc.so.6+0x2b1ff)
 #8 0x019f65c0 llvm::runIPSCCP(llvm::Module&, llvm::DataLayout const&, std::function<llvm::TargetLibraryInfo const& (llvm::Function&)>, llvm::function_ref<llvm::AnalysisResultsForFn (llvm::Function&)>) (/glaubitz/llvm-project/stage1.install/bin/clang+0x19f65c0)
 #9 0x015d3b7d llvm::IPSCCPPass::run(llvm::Module&, llvm::AnalysisManager<llvm::Module>&) (/glaubitz/llvm-project/stage1.install/bin/clang+0x15d3b7d)
#10 0x02ae78dd llvm::detail::PassModel<llvm::Module, llvm::IPSCCPPass, llvm::PreservedAnalyses, llvm::AnalysisManager<llvm::Module> >::run(llvm::Module&, llvm::AnalysisManager<llvm::Module>&) (/glaubitz/llvm-project/stage1.install/bin/clang+0x2ae78dd)
#11 0x0146f513 llvm::PassManager<llvm::Module, llvm::AnalysisManager<llvm::Module> >::run(llvm::Module&, llvm::AnalysisManager<llvm::Module>&) (/glaubitz/llvm-project/stage1.install/bin/clang+0x146f513)
#12 0x01de06a1 (anonymous namespace)::EmitAssemblyHelper::EmitAssemblyWithNewPassManager(clang::BackendAction, std::unique_ptr<llvm::raw_pwrite_stream, std::default_delete<llvm::raw_pwrite_stream> >) (.constprop.0) BackendUtil.cpp:0:0
#13 0x01de3ae2 clang::EmitBackendOutput(clang::DiagnosticsEngine&, clang::HeaderSearchOptions const&, clang::CodeGenOptions const&, clang::TargetOptions const&, clang::LangOptions const&, llvm::DataLayout const&, llvm::Module*, clang::BackendAction, std::unique_ptr<llvm::raw_pwrite_stream, std::default_delete<llvm::raw_pwrite_stream> >) (/glaubitz/llvm-project/stage1.install/bin/clang+0x1de3ae2)
#14 0x02a844dd clang::BackendConsumer::HandleTranslationUnit(clang::ASTContext&) (/glaubitz/llvm-project/stage1.install/bin/clang+0x2a844dd)
#15 0x0368be90 clang::ParseAST(clang::Sema&, bool, bool) (/glaubitz/llvm-project/stage1.install/bin/clang+0x368be90)
#16 0x02424b3c clang::FrontendAction::Execute() (/glaubitz/llvm-project/stage1.install/bin/clang+0x2424b3c)
#17 0x023bb424 clang::CompilerInstance::ExecuteAction(clang::FrontendAction&) (/glaubitz/llvm-project/stage1.install/bin/clang+0x23bb424)
#18 0x024e0fd3 clang::ExecuteCompilerInvocation(clang::CompilerInstance*) (/glaubitz/llvm-project/stage1.install/bin/clang+0x24e0fd3)
#19 0x008dd673 cc1_main(llvm::ArrayRef<char const*>, char const*, void*) (/glaubitz/llvm-project/stage1.install/bin/clang+0x8dd673)
#20 0x008d921d ExecuteCC1Tool(llvm::SmallVectorImpl<char const*>&) driver.cpp:0:0
#21 0x02262c99 void llvm::function_ref<void ()>::callback_fn<clang::driver::CC1Command::Execute(llvm::ArrayRef<llvm::Optional<llvm::StringRef> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >*, bool*) const::'lambda'()>(int) Job.cpp:0:0
#22 0x01a8eb8d llvm::CrashRecoveryContext::RunSafely(llvm::function_ref<void ()>) (/glaubitz/llvm-project/stage1.install/bin/clang+0x1a8eb8d)
#23 0x02263ecb clang::driver::CC1Command::Execute(llvm::ArrayRef<llvm::Optional<llvm::StringRef> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >*, bool*) const (.part.0) Job.cpp:0:0
#24 0x0223911c clang::driver::Compilation::ExecuteCommand(clang::driver::Command const&, clang::driver::Command const*&) const (/glaubitz/llvm-project/stage1.install/bin/clang+0x223911c)
#25 0x02239b37 clang::driver::Compilation::ExecuteJobs(clang::driver::JobList const&, llvm::SmallVectorImpl<std::pair<int, clang::driver::Command const*> >&) const (/glaubitz/llvm-project/stage1.install/bin/clang+0x2239b37)
#26 0x022452f4 clang::driver::Driver::ExecuteCompilation(clang::driver::Compilation&, llvm::SmallVectorImpl<std::pair<int, clang::driver::Command const*> >&) (/glaubitz/llvm-project/stage1.install/bin/clang+0x22452f4)
#27 0x0085ecea main (/glaubitz/llvm-project/stage1.install/bin/clang+0x85ecea)
#28 0xfffffffff7a13da2 __libc_start_main (/lib/x86_64-linux-gnux32/libc.so.6+0x1dda2)
#29 0x008d8bbb _start (/glaubitz/llvm-project/stage1.install/bin/clang+0x8d8bbb)
clang-13: error: clang frontend command failed with exit code 134 (use -v to see invocation)
clang version 13.0.0
Target: x86_64-unknown-linux-gnu
Thread model: posix
InstalledDir: /glaubitz/llvm-project/stage1.install/bin
clang-13: note: diagnostic msg:
********************

PLEASE ATTACH THE FOLLOWING FILES TO THE BUG REPORT:
Preprocessed source(s) and associated run script(s) are located at:
clang-13: note: diagnostic msg: /tmp/cpu_model-3ac5e8.c
clang-13: note: diagnostic msg: /tmp/cpu_model-3ac5e8.sh
clang-13: note: diagnostic msg:

********************

I'll file a bug report.

[RKSimon]

@madhur13490 This is causing some build failures: http://lab.llvm.org:8011/#/builders/121/builds/6287

[glaubitz]

Filed x32-related bug as: https://bugs.llvm.org/show_bug.cgi?id=49861

[RKSimon]

@glaubitz Now that there's a repro - please could you revert this so @madhur13490 can investigate it offline?

[glaubitz]

In D98884#2671195, @RKSimon wrote:

@glaubitz Now that there's a repro - please could you revert this so @madhur13490 can investigate it offline?

I don't have commit access, unfortunately.

[RKSimon]

Thanks @jansvoboda11

[davezarzycki]

I have an auto-bisecting cron job that has identified this change as a second-stage regression on Fedora 34 (x86-64). Basically llvm-tblgen crashes reliably:

FAILED: lib/Target/X86/X86GenDAGISel.inc
cd /tmp/_update_lc/t && /tmp/_update_lc/t/bin/llvm-tblgen -gen-dag-isel -I /home/dave/ro_s/lp/llvm/lib/Target/X86 -I/tmp/_update_lc/t/include -I/home/dave/ro_s/lp/llvm/include -I /home/dave/ro_s/lp/llvm/lib/Target -omit-comments /home/dave/ro_s/lp/llvm/lib/Target/X86/X86.td --write-if-changed -o lib/Target/X86/X86GenDAGISel.inc -d lib/Target/X86/X86GenDAGISel.inc.d
PLEASE submit a bug report to https://bugs.llvm.org/ and include the crash backtrace.
Stack dump:
0.  Program arguments: /tmp/_update_lc/t/bin/llvm-tblgen -gen-dag-isel -I /home/dave/ro_s/lp/llvm/lib/Target/X86 -I/tmp/_update_lc/t/include -I/home/dave/ro_s/lp/llvm/include -I /home/dave/ro_s/lp/llvm/lib/Target -omit-comments /home/dave/ro_s/lp/llvm/lib/Target/X86/X86.td --write-if-changed -o lib/Target/X86/X86GenDAGISel.inc -d lib/Target/X86/X86GenDAGISel.inc.d
#0 0x000000000050b153 llvm::sys::PrintStackTrace(llvm::raw_ostream&, int) (/tmp/_update_lc/t/bin/llvm-tblgen+0x50b153)
#1 0x00000000005094b2 llvm::sys::RunSignalHandlers() (/tmp/_update_lc/t/bin/llvm-tblgen+0x5094b2)
#2 0x000000000050b4fa SignalHandler(int) Signals.cpp:0:0
#3 0x00007ffff7fa6a00 __restore_rt sigaction.c:0:0
#4 0x0000000000316a7d (/tmp/_update_lc/t/bin/llvm-tblgen+0x316a7d)

Is this issue known?

[DavidSpickett]

We've also seen failures on all of our 2 stage bots e.g. http://lab.llvm.org:8011/#/builders/7/builds/2348. Which are fixed now this is reverted.

[494/5842] Building Attributes.inc...
FAILED: include/llvm/IR/Attributes.inc 
cd /home/tcwg-buildslave/worker/clang-cmake-aarch64-full/stage2 && /home/tcwg-buildslave/worker/clang-cmake-aarch64-full/stage2/bin/llvm-tblgen -gen-attrs -I /home/tcwg-buildslave/worker/clang-cmake-aarch64-full/llvm/llvm/include/llvm/IR -I/home/tcwg-buildslave/worker/clang-cmake-aarch64-full/stage2/include -I/home/tcwg-buildslave/worker/clang-cmake-aarch64-full/llvm/llvm/include /home/tcwg-buildslave/worker/clang-cmake-aarch64-full/llvm/llvm/include/llvm/IR/Attributes.td --write-if-changed -o include/llvm/IR/Attributes.inc -d include/llvm/IR/Attributes.inc.d
pure virtual method called
terminate called without an active exception

Bisecting this I have seen clang-tblgen crash with a few different asserts. Bad alloc, one in erase() in StringMap, illegal instruction. Our bot doesn't have sanitizers on so it could be the same finicky memory corruption.

This PPC one does though https://lab.llvm.org/buildbot/#/builders/18/builds/1140 and saw similar failures. No immediate explanation but some configs worth checking locally before the reland.

[jansvoboda11]

Back to green after the revert: https://lab.llvm.org/buildbot/#/builders/121/builds/6307

[foad]

I think it was also causing this failure in my Release+Asserts build:

/home/jayfoad2/git/llvm-project/llvm/test/Transforms/WholeProgramDevirt/branch-funnel.ll:150:16: error: CHECK-LABEL: expected string not found in input
; CHECK-LABEL: define internal void @branch_funnel(i8*
               ^
<stdin>:81:16: note: scanning from here
define i32 @fn3(i8* nocapture readonly %obj) local_unnamed_addr #2 {
               ^
<stdin>:95:18: note: possible intended match here
define hidden void @__typeid_typeid1_0_branch_funnel(i8* nest %0, ...) local_unnamed_addr #6 {
                 ^
...
Failed Tests (1):
  LLVM :: Transforms/WholeProgramDevirt/branch-funnel.ll

[madhur13490]

These failures were not caught by LLVM bots which run on a patch. Can I do a trial run of the bots and run all tests that would run when the patch lands?

[davezarzycki]

In D98884#2672170, @madhur13490 wrote:

These failures were not caught by LLVM bots which run on a patch. Can I do a trial run of the bots and run all tests that would run when the patch lands?

I don’t think LLVM is set up for that. Can you try two stage testing first?

[madhur13490]

In D98884#2671293, @DavidSpickett wrote:

We've also seen failures on all of our 2 stage bots e.g. http://lab.llvm.org:8011/#/builders/7/builds/2348. Which are fixed now this is reverted.

[494/5842] Building Attributes.inc...
FAILED: include/llvm/IR/Attributes.inc 
cd /home/tcwg-buildslave/worker/clang-cmake-aarch64-full/stage2 && /home/tcwg-buildslave/worker/clang-cmake-aarch64-full/stage2/bin/llvm-tblgen -gen-attrs -I /home/tcwg-buildslave/worker/clang-cmake-aarch64-full/llvm/llvm/include/llvm/IR -I/home/tcwg-buildslave/worker/clang-cmake-aarch64-full/stage2/include -I/home/tcwg-buildslave/worker/clang-cmake-aarch64-full/llvm/llvm/include /home/tcwg-buildslave/worker/clang-cmake-aarch64-full/llvm/llvm/include/llvm/IR/Attributes.td --write-if-changed -o include/llvm/IR/Attributes.inc -d include/llvm/IR/Attributes.inc.d
pure virtual method called
terminate called without an active exception

Bisecting this I have seen clang-tblgen crash with a few different asserts. Bad alloc, one in erase() in StringMap, illegal instruction. Our bot doesn't have sanitizers on so it could be the same finicky memory corruption.

This PPC one does though https://lab.llvm.org/buildbot/#/builders/18/builds/1140 and saw similar failures. No immediate explanation but some configs worth checking locally before the reland.

Apparently my change exposes the memory corruption bug on X86 too. As the bots show, I can't generate .inc for almost all ARM .td files. Not sure how to fix this reliably. Surprisingly, the builds pass in Debug build type.

[madhur13490]

This change has wider impact than was expected as it affects several components and backends. Bootstrap builds fail with this change and as the buildbots show, this change generates a buggy Tablegen binary in release mode.

Debugging:

1. Disabling the newly added code in hasAddressTaken(), we get sane Tablegen and builds pass. To note, debug builds pass with the change. This clearly shows that codegen is functionally different in release and debug mode for Tablegen.
2. Doing “ninja check-all” on stage1 build does not help although we have ~90000 tests.
3. Running valgrind also did not reveal any new things.
4. Doing ASAN build of the compiler led to several errors and drawing any conclusion from it is close to impossible for this patch.

However, if I disable IPSCCP pass on top of this patch then everything turns green. I tried to fix trivial issues in the pass but it seems the pass needs deeper and major fixes to make this patch land.

Till we fix issues in IPSCCP, this patch stands reverted. The whole intention of this patch was to handle a subset of cases for AMDGPU backend, specifically, needed for https://reviews.llvm.org/D99347.

[arsenm]

In D98884#2678593, @madhur13490 wrote:

This change has wider impact than was expected as it affects several components and backends. Bootstrap builds fail with this change and as the buildbots show, this change generates a buggy Tablegen binary in release mode.

Debugging:

1. Disabling the newly added code in hasAddressTaken(), we get sane Tablegen and builds pass. To note, debug builds pass with the change. This clearly shows that codegen is functionally different in release and debug mode for Tablegen.
2. Doing “ninja check-all” on stage1 build does not help although we have ~90000 tests.
3. Running valgrind also did not reveal any new things.
4. Doing ASAN build of the compiler led to several errors and drawing any conclusion from it is close to impossible for this patch.

However, if I disable IPSCCP pass on top of this patch then everything turns green. I tried to fix trivial issues in the pass but it seems the pass needs deeper and major fixes to make this patch land.

Till we fix issues in IPSCCP, this patch stands reverted. The whole intention of this patch was to handle a subset of cases for AMDGPU backend, specifically, needed for https://reviews.llvm.org/D99347.

This doesn't necessarily mean there's a problem with IPSCCP. What is the failing IR?

[madhur13490]

In D98884#2680577, @arsenm wrote:

In D98884#2678593, @madhur13490 wrote:

This change has wider impact than was expected as it affects several components and backends. Bootstrap builds fail with this change and as the buildbots show, this change generates a buggy Tablegen binary in release mode.

Debugging:

1. Disabling the newly added code in hasAddressTaken(), we get sane Tablegen and builds pass. To note, debug builds pass with the change. This clearly shows that codegen is functionally different in release and debug mode for Tablegen.
2. Doing “ninja check-all” on stage1 build does not help although we have ~90000 tests.
3. Running valgrind also did not reveal any new things.
4. Doing ASAN build of the compiler led to several errors and drawing any conclusion from it is close to impossible for this patch.

However, if I disable IPSCCP pass on top of this patch then everything turns green. I tried to fix trivial issues in the pass but it seems the pass needs deeper and major fixes to make this patch land.

Till we fix issues in IPSCCP, this patch stands reverted. The whole intention of this patch was to handle a subset of cases for AMDGPU backend, specifically, needed for https://reviews.llvm.org/D99347.

This doesn't necessarily mean there's a problem with IPSCCP. What is the failing IR?

There is no failing IR but the failing binary. Tablegen constitutes around ~200 C/C++ files and all I know is that the binary is misompiling. It is pretty difficult to say at this point if one single file is miscompiling or a bunch of them. All I know at this point is

1. A lot of new functions are now added to IPSCCP's radar as hasAddressTaken() for them is 'false' now as it meets the condition my patch adds i.e all users of bitcast operator are CB. See llvm::canTrackArgumentsInterprocedurally(Function *F) which is called from IPSCCP.
2. IPSCCP may not be the culprit but it is definitely a starting point of my investigation. There are obvious places in it which needs fix e.g. properly removing argmemonly and inaccessiblemem_or_argmemonly attributes here.

[jdoerfert]

Are we sure this is not a single bad cast that evolves into something different if assertions are disabled. Is the cast problem located, that should be easy after all given the minimal reproducer in PR49861?

[madhur13490]

In D98884#2681986, @jdoerfert wrote:

Are we sure this is not a single bad cast that evolves into something different if assertions are disabled. Is the cast problem located, that should be easy after all given the minimal reproducer in PR49861?

The two line reproducer points to one part of the problem and it does fix cpu_model.c assert (mentioned in PR49861) but it does not help the miscompilation of TableGen. All my experiments are done after I fix the problem pointed by assert in cpu_model.c.

[jdoerfert]

In D98884#2681227, @madhur13490 wrote:

1. A lot of new functions are now added to IPSCCP's radar as hasAddressTaken() for them is 'false' now as it meets the condition my patch adds i.e all users of bitcast operator are CB. See llvm::canTrackArgumentsInterprocedurally(Function *F) which is called from IPSCCP.

You can use a Debug Counter and bisect it to the first function for which the hasAddressTaken change will cause the issue.
With the pre- and post-IPSCCP module we should be able to figure this out.

llvm/lib/IR/Function.cpp

 
 /// hasAddressTaken - returns true if there are any uses of this function
 /// other than direct calls or invokes to it. Optionally ignores callback
 /// uses, assume like pointer annotation calls, and references in llvm.used
 /// and llvm.compiler.used variables.
 bool Function::hasAddressTaken(const User **PutOffender,
                                bool IgnoreCallbackUses,
                                bool IgnoreAssumeLikeCalls,
                                bool IgnoreLLVMUsed) const {
   for (const Use &U : uses()) {
     const User *FU = U.getUser();
     if (isa<BlockAddress>(FU))
       continue;
 
     if (IgnoreCallbackUses) {
       AbstractCallSite ACS(&U);
       if (ACS && ACS.isCallbackCall())
         continue;
     }
 
+    if (isa<BitCastOperator>(FU) && isa<ConstantExpr>(FU) &&
+        llvm::all_of(FU->uses(), [](const Use &U) {
+          if (const CallBase *CB = dyn_cast<CallBase>(U.getUser()))
+            return CB->isCallee(&U);
+          return false;
+        }))
+      continue;
+
     const auto *Call = dyn_cast<CallBase>(FU);
     if (!Call) {
       if (IgnoreAssumeLikeCalls) {
         if (const auto *FI = dyn_cast<Instruction>(FU)) {
           if (FI->isCast() && !FI->user_empty() &&
               llvm::all_of(FU->users(), [](const User *U) {
                 if (const auto *I = dyn_cast<IntrinsicInst>(U))
                   return I->isAssumeLikeIntrinsic();

llvm/lib/Transforms/IPO/GlobalOpt.cpp

if (!GVar)
     return Changed;
 
   if (GVar->isConstant() || !GVar->hasInitializer())
     return Changed;
 
   return processInternalGlobal(GVar, GS, GetTLI, LookupDomTree) || Changed;
 }
 
+static void SetCallBaseFastCallingConv(Use &U) {
+  CallBase *CB = dyn_cast<CallBase>(U.getUser());
+  if (!CB || !CB->isCallee(&U))
+    return;
+  CB->setCallingConv(CallingConv::Fast);
+}
+
 /// Walk all of the direct calls of the specified function, changing them to
 /// FastCC.
 static void ChangeCalleesToFastCall(Function *F) {
-  for (User *U : F->users()) {
-    if (isa<BlockAddress>(U))
+  for (Use &U : F->uses()) {
+    User *FU = U.getUser();
+    if (isa<BlockAddress>(FU))
+      continue;
+    if (isa<BitCastOperator>(FU) && isa<ConstantExpr>(FU)) {
+      for (Use &BU : FU->uses())
+        SetCallBaseFastCallingConv(BU);
       continue;
-    cast<CallBase>(U)->setCallingConv(CallingConv::Fast);
+    }
+    SetCallBaseFastCallingConv(U);
   }
 }
 
 static AttributeList StripAttr(LLVMContext &C, AttributeList Attrs,
                                Attribute::AttrKind A) {
   unsigned AttrIndex;
   if (Attrs.hasAttrSomewhere(A, &AttrIndex))
     return Attrs.removeAttribute(C, AttrIndex, A);
   return Attrs;
 }
 
+static void SetCallBaseAttributes(Use &U, Function *F, Attribute::AttrKind A) {
+  CallBase *CB = dyn_cast<CallBase>(U.getUser());
+  if (!CB || !CB->isCallee(&U))
+    return;
+  CB->setAttributes(StripAttr(F->getContext(), CB->getAttributes(), A));
+}
+
 static void RemoveAttribute(Function *F, Attribute::AttrKind A) {
   F->setAttributes(StripAttr(F->getContext(), F->getAttributes(), A));
-  for (User *U : F->users()) {
-    if (isa<BlockAddress>(U))
+  for (Use &U : F->uses()) {
+    User *FU = U.getUser();
+    if (isa<BlockAddress>(FU))
       continue;
-    CallBase *CB = cast<CallBase>(U);
-    CB->setAttributes(StripAttr(F->getContext(), CB->getAttributes(), A));
+    if (isa<BitCastOperator>(FU) && isa<ConstantExpr>(FU)) {
+      for (Use &BU : FU->uses())
+        SetCallBaseAttributes(BU, F, A);
+      continue;
+    }
+    SetCallBaseAttributes(U, F, A);
   }
 }
 
 /// Return true if this is a calling convention that we'd like to change.  The
 /// idea here is that we don't want to mess with the convention if the user
 /// explicitly requested something with performance implications like coldcc,
 /// GHC, or anyregcc.
 static bool hasChangeableCC(Function *F) {
   CallingConv::ID CC = F->getCallingConv();
 
   // FIXME: Is it worth transforming x86_stdcallcc and x86_fastcallcc?
   if (CC != CallingConv::C && CC != CallingConv::X86_ThisCall)

llvm/test/Analysis/CallGraph/ignore-bitcast-call-argument-callee.ll

+; RUN:  opt < %s -print-callgraph -disable-output 2>&1 | FileCheck %s
+
+; CHECK: Call graph node <<null function>><<{{.*}}>>  #uses=0
+; CHECK-NEXT:   CS<None> calls function 'foo'
+; CHECK-EMPTY:
+; CHECK-NEXT: Call graph node for function: 'bar'<<{{.*}}>>  #uses=1
+; CHECK-EMPTY:
+; CHECK-NEXT: Call graph node for function: 'foo'<<{{.*}}>>  #uses=1
+; CHECK-EMPTY:
+; CHECK-NEXT: Call graph node for function: 'test'<<{{.*}}>>  #uses=0
+; CHECK-NEXT:  CS<{{.*}}> calls function 'bar'
+; CHECK-NEXT:  CS<{{.*}}> calls external node
+
+define internal i32 @foo() {
+entry:
+    ret i32 5
+}
+
+define internal i32 @bar(float()* %arg) {
+    ret i32 5
+}
+
+define internal i32 @test() {
+  %v1 = call i32 @bar(float()* bitcast (i32()* @foo to float()*))
+  %v2 = call float bitcast (i32()* @foo to float()*)()
+  %v3 = fptoui float %v2 to i32
+  %v4 = add i32 %v1, %v3
+  ret i32 %v4
+}
+

llvm/test/Analysis/CallGraph/ignore-bitcast-call-argument.ll

+; RUN:  opt < %s -print-callgraph -disable-output 2>&1 | FileCheck %s
+
+; CHECK: Call graph node <<null function>><<{{.*}}>>  #uses=0
+; CHECK-NEXT:   CS<None> calls function 'foo'
+; CHECK-EMPTY:
+; CHECK-NEXT: Call graph node for function: 'bar'<<{{.*}}>>  #uses=1
+; CHECK-EMPTY:
+; CHECK-NEXT: Call graph node for function: 'foo'<<{{.*}}>>  #uses=1
+; CHECK-EMPTY:
+; CHECK-NEXT: Call graph node for function: 'test'<<{{.*}}>>  #uses=0
+; CHECK-NEXT:   CS<{{.*}}> calls function 'bar'
+
+define internal i32 @foo() {
+entry:
+    ret i32 5
+}
+
+define internal i32 @bar(float()* %arg) {
+    ret i32 5
+}
+
+define internal i32 @test() {
+  %v1 = call i32 @bar(float()* bitcast (i32()* @foo to float()*))
+  %v2 = add i32 %v1, 6
+  ret i32 %v2
+}
+

llvm/test/Analysis/CallGraph/ignore-bitcast-callees.ll

+; RUN:  opt < %s -print-callgraph -disable-output 2>&1 | FileCheck %s
+; CHECK: Call graph node <<null function>><<{{.*}}>>  #uses=0
+; CHECK-EMPTY:
+; CHECK-NEXT: Call graph node for function: 'foo'<<{{.*}}>>  #uses=0
+; CHECK-EMPTY:
+; CHECK-NEXT: Call graph node for function: 'test_bitcast_callees'<<{{.*}}>>  #uses=0
+; CHECK-NEXT:   CS<{{.*}}> calls external node
+; CHECK-NEXT:   CS<{{.*}}> calls external node
+
+define internal i32 @foo() {
+entry:
+    ret i32 5
+}
+
+define internal float @test_bitcast_callees() {
+  %v1 = call float bitcast (i32()* @foo to float()*)()
+  %v2 = fadd float %v1, 1.0
+  %v3 = call i8 bitcast (i32()* @foo to i8()*)()
+  %v4 = uitofp i8 %v3 to float
+  %v5 = fadd float %v2, %v4
+  ret float %v5
+}
+

llvm/test/Analysis/CallGraph/ignore-bitcast-callees2.ll

+; RUN:  opt < %s -print-callgraph -disable-output 2>&1 | FileCheck %s
+; CHECK: Call graph node <<null function>><<{{.*}}>>  #uses=0
+; CHECK-EMPTY:
+; CHECK-NEXT: Call graph node for function: 'foo'<<{{.*}}>>  #uses=0
+; CHECK-EMPTY:
+; CHECK-NEXT: Call graph node for function: 'test_bitcast_callees2'<<{{.*}}>>  #uses=0
+; CHECK-NEXT:   CS<{{.*}}> calls external node
+; CHECK-NEXT:   CS<{{.*}}> calls external node
+; CHECK-NEXT:   CS<{{.*}}> calls external node
+
+define internal i32 @foo() {
+entry:
+    ret i32 5
+}
+
+define internal float @test_bitcast_callees2() {
+  %v1 = call float bitcast (i32()* @foo to float()*)()
+  %v2 = fadd float %v1, 1.0
+  %v3 = call float bitcast (i32()* @foo to float()*)()
+  %v4 = fadd float %v3, %v2
+  %v5 = call i8 bitcast (i32()* @foo to i8()*)()
+  %v6 = uitofp i8 %v5 to float
+  %v7 = fadd float %v4, %v6
+  ret float %v7
+}
+

llvm/test/Transforms/Attributor/IPConstantProp/arg-count-mismatch.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --function-signature --check-attributes --check-globals
 ; RUN: opt -attributor -enable-new-pm=0 -attributor-manifest-internal  -attributor-max-iterations-verify -attributor-annotate-decl-cs -attributor-max-iterations=1 -S < %s | FileCheck %s --check-prefixes=CHECK,NOT_CGSCC_NPM,NOT_CGSCC_OPM,NOT_TUNIT_NPM,IS__TUNIT____,IS________OPM,IS__TUNIT_OPM
 ; RUN: opt -aa-pipeline=basic-aa -passes=attributor -attributor-manifest-internal  -attributor-max-iterations-verify -attributor-annotate-decl-cs -attributor-max-iterations=1 -S < %s | FileCheck %s --check-prefixes=CHECK,NOT_CGSCC_OPM,NOT_CGSCC_NPM,NOT_TUNIT_OPM,IS__TUNIT____,IS________NPM,IS__TUNIT_NPM
 ; RUN: opt -attributor-cgscc -enable-new-pm=0 -attributor-manifest-internal  -attributor-annotate-decl-cs -S < %s | FileCheck %s --check-prefixes=CHECK,NOT_TUNIT_NPM,NOT_TUNIT_OPM,NOT_CGSCC_NPM,IS__CGSCC____,IS________OPM,IS__CGSCC_OPM
 ; RUN: opt -aa-pipeline=basic-aa -passes=attributor-cgscc -attributor-manifest-internal  -attributor-annotate-decl-cs -S < %s | FileCheck %s --check-prefixes=CHECK,NOT_TUNIT_NPM,NOT_TUNIT_OPM,NOT_CGSCC_OPM,IS__CGSCC____,IS________NPM,IS__CGSCC_NPM
 
 ; The original C source looked like this:
 ;
 ;   long long a101, b101, e101;
 ;   volatile long c101;
 ;   int d101;
 ;
 ;   static inline int bar(p1, p2)
 ;   {

 ; Note that -lint will report this as UB, but it passes -verify.
 
 ; This test is just to verify that we do not crash/assert due to mismatch in
 ; argument count between the caller and callee.
 
 ; FIXME we should recognize this as UB and make it an unreachable.
 
 define dso_local i16 @foo(i16 %a) {
-; NOT_CGSCC_NPM-LABEL: define {{[^@]+}}@foo
-; NOT_CGSCC_NPM-SAME: (i16 [[A:%.*]]) {
-; NOT_CGSCC_NPM-NEXT:    [[CALL:%.*]] = call i16 @bar()
-; NOT_CGSCC_NPM-NEXT:    ret i16 [[CALL]]
+; IS__TUNIT____-LABEL: define {{[^@]+}}@foo
+; IS__TUNIT____-SAME: (i16 [[A:%.*]]) {
+; IS__TUNIT____-NEXT:    [[CALL:%.*]] = call i16 @bar()
+; IS__TUNIT____-NEXT:    ret i16 [[CALL]]
+;
+; IS__CGSCC_OPM-LABEL: define {{[^@]+}}@foo
+; IS__CGSCC_OPM-SAME: (i16 [[A:%.*]]) {
+; IS__CGSCC_OPM-NEXT:    [[CALL:%.*]] = call i16 bitcast (i16 (i16, i16)* @bar to i16 (i16)*)(i16 [[A]])
+; IS__CGSCC_OPM-NEXT:    ret i16 [[CALL]]
 ;
 ; IS__CGSCC_NPM: Function Attrs: nofree norecurse nosync nounwind readnone willreturn
 ; IS__CGSCC_NPM-LABEL: define {{[^@]+}}@foo
 ; IS__CGSCC_NPM-SAME: (i16 [[A:%.*]]) #[[ATTR0:[0-9]+]] {
 ; IS__CGSCC_NPM-NEXT:    ret i16 0
 ;
   %call = call i16 bitcast (i16 (i16, i16) * @bar to i16 (i16) *)(i16 %a)
   ret i16 %call
 }
 
 define internal i16 @bar(i16 %p1, i16 %p2) {
 ; IS__TUNIT____: Function Attrs: nofree nosync nounwind readnone willreturn
 ; IS__TUNIT____-LABEL: define {{[^@]+}}@bar
 ; IS__TUNIT____-SAME: () #[[ATTR0:[0-9]+]] {
 ; IS__TUNIT____-NEXT:    ret i16 0
 ;
-; IS__CGSCC_OPM: Function Attrs: nofree norecurse nosync nounwind readnone willreturn
 ; IS__CGSCC_OPM-LABEL: define {{[^@]+}}@bar
-; IS__CGSCC_OPM-SAME: () #[[ATTR0:[0-9]+]] {
+; IS__CGSCC_OPM-SAME: (i16 [[P1:%.*]], i16 [[P2:%.*]]) {
 ; IS__CGSCC_OPM-NEXT:    ret i16 0
 ;
 ; IS__CGSCC_NPM: Function Attrs: nofree norecurse nosync nounwind readnone willreturn
 ; IS__CGSCC_NPM-LABEL: define {{[^@]+}}@bar
 ; IS__CGSCC_NPM-SAME: () #[[ATTR0]] {
 ; IS__CGSCC_NPM-NEXT:    ret i16 undef
 ;
   ret i16 0

 
 define internal i16 @bar2(i16 %p1, i16 %p2) {
 ; IS__TUNIT____: Function Attrs: nofree nosync nounwind readnone willreturn
 ; IS__TUNIT____-LABEL: define {{[^@]+}}@bar2
 ; IS__TUNIT____-SAME: (i16 [[P1:%.*]], i16 [[P2:%.*]]) #[[ATTR0]] {
 ; IS__TUNIT____-NEXT:    [[A:%.*]] = add i16 [[P1]], [[P2]]
 ; IS__TUNIT____-NEXT:    ret i16 [[A]]
 ;
-; IS__CGSCC____: Function Attrs: nofree norecurse nosync nounwind readnone willreturn
-; IS__CGSCC____-LABEL: define {{[^@]+}}@bar2
-; IS__CGSCC____-SAME: (i16 [[P1:%.*]], i16 [[P2:%.*]]) #[[ATTR0:[0-9]+]] {
-; IS__CGSCC____-NEXT:    [[A:%.*]] = add i16 [[P1]], [[P2]]
-; IS__CGSCC____-NEXT:    ret i16 [[A]]
+; IS__CGSCC_OPM-LABEL: define {{[^@]+}}@bar2
+; IS__CGSCC_OPM-SAME: (i16 [[P1:%.*]], i16 [[P2:%.*]]) {
+; IS__CGSCC_OPM-NEXT:    [[A:%.*]] = add i16 [[P1]], [[P2]]
+; IS__CGSCC_OPM-NEXT:    ret i16 [[A]]
+;
+; IS__CGSCC_NPM: Function Attrs: nofree norecurse nosync nounwind readnone willreturn
+; IS__CGSCC_NPM-LABEL: define {{[^@]+}}@bar2
+; IS__CGSCC_NPM-SAME: (i16 [[P1:%.*]], i16 [[P2:%.*]]) #[[ATTR0]] {
+; IS__CGSCC_NPM-NEXT:    [[A:%.*]] = add i16 [[P1]], [[P2]]
+; IS__CGSCC_NPM-NEXT:    ret i16 [[A]]
 ;
   %a = add i16 %p1, %p2
   ret i16 %a
 }
 
 
 ;-------------------------------------------------------------------------------
 ; Additional tests to verify that we still optimize when having a mismatch
 ; in argument count due to varargs (as long as all non-variadic arguments have
 ; been provided),
 
 define dso_local i16 @vararg_tests(i16 %a) {
-; CHECK-LABEL: define {{[^@]+}}@vararg_tests
-; CHECK-SAME: (i16 [[A:%.*]]) {
-; CHECK-NEXT:    [[CALL2:%.*]] = call i16 bitcast (i16 (i16, i16, ...)* @vararg_no_prop to i16 (i16)*)(i16 noundef 7)
-; CHECK-NEXT:    [[ADD:%.*]] = add i16 7, [[CALL2]]
-; CHECK-NEXT:    ret i16 [[ADD]]
+; NOT_CGSCC_OPM-LABEL: define {{[^@]+}}@vararg_tests
+; NOT_CGSCC_OPM-SAME: (i16 [[A:%.*]]) {
+; NOT_CGSCC_OPM-NEXT:    [[CALL2:%.*]] = call i16 bitcast (i16 (i16, i16, ...)* @vararg_no_prop to i16 (i16)*)(i16 noundef 7)
+; NOT_CGSCC_OPM-NEXT:    [[ADD:%.*]] = add i16 7, [[CALL2]]
+; NOT_CGSCC_OPM-NEXT:    ret i16 [[ADD]]
+;
+; IS__CGSCC_OPM-LABEL: define {{[^@]+}}@vararg_tests
+; IS__CGSCC_OPM-SAME: (i16 [[A:%.*]]) {
+; IS__CGSCC_OPM-NEXT:    [[CALL2:%.*]] = call i16 bitcast (i16 (i16, i16, ...)* @vararg_no_prop to i16 (i16)*)(i16 7)
+; IS__CGSCC_OPM-NEXT:    [[ADD:%.*]] = add i16 7, [[CALL2]]
+; IS__CGSCC_OPM-NEXT:    ret i16 [[ADD]]
 ;
   %call1 = call i16 (i16, ...) @vararg_prop(i16 7, i16 8, i16 %a)
   %call2 = call i16 bitcast (i16 (i16, i16, ...) * @vararg_no_prop to i16 (i16) *) (i16 7)
   %add = add i16 %call1, %call2
   ret i16 %add
 }
 
 define internal i16 @vararg_prop(i16 %p1, ...) {
 ; IS__CGSCC____: Function Attrs: nofree norecurse nosync nounwind readnone willreturn
 ; IS__CGSCC____-LABEL: define {{[^@]+}}@vararg_prop
-; IS__CGSCC____-SAME: (i16 returned [[P1:%.*]], ...) #[[ATTR0]] {
+; IS__CGSCC____-SAME: (i16 returned [[P1:%.*]], ...) #[[ATTR0:[0-9]+]] {
 ; IS__CGSCC____-NEXT:    ret i16 7
 ;
   ret i16 %p1
 }
 
 define internal i16 @vararg_no_prop(i16 %p1, i16 %p2, ...) {
 ; IS__TUNIT____: Function Attrs: nofree nosync nounwind readnone willreturn
 ; IS__TUNIT____-LABEL: define {{[^@]+}}@vararg_no_prop
 ; IS__TUNIT____-SAME: (i16 returned [[P1:%.*]], i16 [[P2:%.*]], ...) #[[ATTR0]] {
 ; IS__TUNIT____-NEXT:    ret i16 7
 ;
-; IS__CGSCC____: Function Attrs: nofree norecurse nosync nounwind readnone willreturn
-; IS__CGSCC____-LABEL: define {{[^@]+}}@vararg_no_prop
-; IS__CGSCC____-SAME: (i16 returned [[P1:%.*]], i16 [[P2:%.*]], ...) #[[ATTR0]] {
-; IS__CGSCC____-NEXT:    ret i16 7
+; IS__CGSCC_OPM-LABEL: define {{[^@]+}}@vararg_no_prop
+; IS__CGSCC_OPM-SAME: (i16 [[P1:%.*]], i16 [[P2:%.*]], ...) {
+; IS__CGSCC_OPM-NEXT:    ret i16 [[P1]]
+;
+; IS__CGSCC_NPM: Function Attrs: nofree norecurse nosync nounwind readnone willreturn
+; IS__CGSCC_NPM-LABEL: define {{[^@]+}}@vararg_no_prop
+; IS__CGSCC_NPM-SAME: (i16 returned [[P1:%.*]], i16 [[P2:%.*]], ...) #[[ATTR0]] {
+; IS__CGSCC_NPM-NEXT:    ret i16 7
 ;
   ret i16 %p1
 }
 
 ;.
 ; IS__TUNIT____: attributes #[[ATTR0]] = { nofree nosync nounwind readnone willreturn }
 ;.
 ; IS__CGSCC____: attributes #[[ATTR0]] = { nofree norecurse nosync nounwind readnone willreturn }
 ;.

llvm/test/Transforms/Attributor/IPConstantProp/arg-type-mismatch.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --function-signature --check-attributes --check-globals
 ; RUN: opt -attributor -enable-new-pm=0 -attributor-manifest-internal  -attributor-max-iterations-verify -attributor-annotate-decl-cs -attributor-max-iterations=1 -S < %s | FileCheck %s --check-prefixes=CHECK,NOT_CGSCC_NPM,NOT_CGSCC_OPM,NOT_TUNIT_NPM,IS__TUNIT____,IS________OPM,IS__TUNIT_OPM
 ; RUN: opt -aa-pipeline=basic-aa -passes=attributor -attributor-manifest-internal  -attributor-max-iterations-verify -attributor-annotate-decl-cs -attributor-max-iterations=1 -S < %s | FileCheck %s --check-prefixes=CHECK,NOT_CGSCC_OPM,NOT_CGSCC_NPM,NOT_TUNIT_OPM,IS__TUNIT____,IS________NPM,IS__TUNIT_NPM
 ; RUN: opt -attributor-cgscc -enable-new-pm=0 -attributor-manifest-internal  -attributor-annotate-decl-cs -S < %s | FileCheck %s --check-prefixes=CHECK,NOT_TUNIT_NPM,NOT_TUNIT_OPM,NOT_CGSCC_NPM,IS__CGSCC____,IS________OPM,IS__CGSCC_OPM
 ; RUN: opt -aa-pipeline=basic-aa -passes=attributor-cgscc -attributor-manifest-internal  -attributor-annotate-decl-cs -S < %s | FileCheck %s --check-prefixes=CHECK,NOT_TUNIT_NPM,NOT_TUNIT_OPM,NOT_CGSCC_OPM,IS__CGSCC____,IS________NPM,IS__CGSCC_NPM
 
 ; This test is just to verify that we do not crash/assert due to mismatch in
 ; argument type between the caller and callee.
 
 define dso_local i16 @foo(i16 %a) {
 ; CHECK-LABEL: define {{[^@]+}}@foo
 ; CHECK-SAME: (i16 [[A:%.*]]) {
 ; CHECK-NEXT:    [[CALL:%.*]] = call i16 bitcast (i16 (i16, i16)* @bar to i16 (i16, i32)*)(i16 [[A]], i32 7)
 ; CHECK-NEXT:    ret i16 [[CALL]]
 ;
   %call = call i16 bitcast (i16 (i16, i16) * @bar to i16 (i16, i32) *)(i16 %a, i32 7)
   ret i16 %call
 }
 
 define internal i16 @bar(i16 %p1, i16 %p2) {
 ; IS__TUNIT____: Function Attrs: nofree nosync nounwind readnone willreturn
 ; IS__TUNIT____-LABEL: define {{[^@]+}}@bar
 ; IS__TUNIT____-SAME: (i16 [[P1:%.*]], i16 returned [[P2:%.*]]) #[[ATTR0:[0-9]+]] {
 ; IS__TUNIT____-NEXT:    ret i16 [[P2]]
 ;
-; IS__CGSCC____: Function Attrs: nofree norecurse nosync nounwind readnone willreturn
-; IS__CGSCC____-LABEL: define {{[^@]+}}@bar
-; IS__CGSCC____-SAME: (i16 [[P1:%.*]], i16 returned [[P2:%.*]]) #[[ATTR0:[0-9]+]] {
-; IS__CGSCC____-NEXT:    ret i16 [[P2]]
+; IS__CGSCC_OPM-LABEL: define {{[^@]+}}@bar
+; IS__CGSCC_OPM-SAME: (i16 [[P1:%.*]], i16 [[P2:%.*]]) {
+; IS__CGSCC_OPM-NEXT:    ret i16 [[P2]]
+;
+; IS__CGSCC_NPM: Function Attrs: nofree norecurse nosync nounwind readnone willreturn
+; IS__CGSCC_NPM-LABEL: define {{[^@]+}}@bar
+; IS__CGSCC_NPM-SAME: (i16 [[P1:%.*]], i16 returned [[P2:%.*]]) #[[ATTR0:[0-9]+]] {
+; IS__CGSCC_NPM-NEXT:    ret i16 [[P2]]
 ;
   ret i16 %p2
 }
 
 
 ;.
 ; IS__TUNIT____: attributes #[[ATTR0]] = { nofree nosync nounwind readnone willreturn }
 ;.
-; IS__CGSCC____: attributes #[[ATTR0]] = { nofree norecurse nosync nounwind readnone willreturn }
+; IS__CGSCC_NPM: attributes #[[ATTR0]] = { nofree norecurse nosync nounwind readnone willreturn }
 ;.

llvm/test/Transforms/GlobalOpt/assumelike-bitcast-fastcc.ll

+; RUN: opt < %s -globalopt -S | FileCheck %s
+
+; Check that fastccc is not set on
+; function "bar" because its address is taken
+; when no optional parameters are passed to
+; hasAddressTaken() as it returns "true"
+; If the optional parameter corresponding to
+; ignore llvmassumelike is passed to
+; hasAddressTaken() then this test would fail.
+
+declare void @llvm.lifetime.start.p0i8(i64, i8* nocapture)
+
+; CHECK-NOT: define internal fastcc void @bar() {
+; CHECK: define internal void @bar() {
+define internal void @bar() {
+entry:
+  ret void
+}
+
+; CHECK: define void @foo() local_unnamed_addr {
+define void @foo() {
+entry:
+  %c = bitcast void()* @bar to i8*
+  call void @llvm.lifetime.start.p0i8(i64 4, i8* %c)
+  ret void
+}
+

llvm/test/Transforms/GlobalOpt/bitcast-call-argument-fastcc.ll

+; RUN:  opt < %s -globalopt -S | FileCheck %s
+
+; Check that globalopt doesn't set fastcc on
+; functions and their call sites if its hasAddressTaken()
+; is true. "foo" function's hasAddressTaken() is
+; true so globalopt should not set fastcc on it
+; and its call sites. However, for "bar" fastcc must be set.
+
+; CHECK-NOT: define internal fastcc i32 @foo() {
+; CHECK:     define internal i32 @foo() {
+define internal i32 @foo() {
+entry:
+    ret i32 5
+}
+
+; CHECK: define internal fastcc i32 @bar(float ()* %arg) unnamed_addr {
+define internal i32 @bar(float()* %arg) {
+    ret i32 5
+}
+
+; CHECK-NOT: define fastcc i32 @test()
+; CHECK:     define i32 @test()
+define i32 @test() {
+  ; CHECK: call fastcc i32 @bar(float ()* bitcast (i32 ()* @foo to float ()*))
+  %v1 = call i32 @bar(float ()* bitcast (i32 ()* @foo to float ()*))
+  %v2 = add i32 %v1, 6
+  ret i32 %v2
+}
+

llvm/test/Transforms/GlobalOpt/bitcast-callees-fastcc.ll

+; RUN:  opt < %s -globalopt -S | FileCheck %s
+
+; CHECK-LABEL: define internal fastcc i32 @foo() unnamed_addr
+define internal i32 @foo() {
+   ret i32 8
+}
+
+; CHECK-LABEL: define void @test_bitcast_callees2() local_unnamed_addr {
+define void @test_bitcast_callees2() {
+  ; CHECK: %1 = call fastcc float bitcast (i32 ()* @foo to float ()*)()
+  call float bitcast (i32()* @foo to float()*)()
+  ; CHECK-NEXT: %2 = call fastcc float bitcast (i32 ()* @foo to float ()*)()
+  call float bitcast (i32()* @foo to float()*)()
+  ; CHECK-NEXT: %3 = call fastcc i8 bitcast (i32 ()* @foo to i8 ()*)()
+  call i8 bitcast (i32()* @foo to i8()*)()
+  ret void
+}
+

llvm/test/Transforms/GlobalOpt/blockaddr-bitcast-fastcc.ll

+; RUN: opt < %s -globalopt -S | FileCheck %s
+
+; Check if fastcc is set on function "g"
+; as its address is not taken and other
+; conditions are met to tag "g" as fastcc.
+
+@x = internal global i8* zeroinitializer
+
+; CHECK: define internal fastcc void @g() unnamed_addr {
+define internal void @g() {
+entry:
+  br label %here
+
+here:
+  ret void
+}
+
+define void @f() {
+  store i8* blockaddress(@g, %here), i8** @x, align 8
+; CHECK: call fastcc i32 bitcast (void ()* @g to i32 ()*)()
+  call i32 bitcast (void ()* @g to i32 ()*)()
+  ret void
+}
+

llvm/test/Transforms/OpenMP/parallel_deletion_cg_update.ll

 ; RUN: opt < %s -instcombine -attributor-cgscc -print-callgraph -disable-output -verify 2>&1 -enable-new-pm=0 | FileCheck %s
 
 ; CHECK: Call graph node <<null function>><<{{.*}}>>  #uses=0
 ; CHECK:   CS<None> calls function 'dead_fork_call'
-; CHECK:   CS<None> calls function '.omp_outlined..0'
 ; CHECK:   CS<None> calls function '__kmpc_fork_call'
 ; CHECK:   CS<None> calls function 'live_fork_call'
 ; CHECK:   CS<None> calls function '.omp_outlined..1'
 ; CHECK:   CS<None> calls function 'd'
 ;
-; CHECK: Call graph node for function: '.omp_outlined..0'<<{{.*}}>>  #uses=1
-; 
+; CHECK: Call graph node for function: '.omp_outlined..0'<<{{.*}}>>  #uses=0
+;
 ; CHECK: Call graph node for function: '.omp_outlined..1'<<{{.*}}>>  #uses=3
 ; CHECK:   CS<{{.*}}> calls function 'd'
 ;
 ; CHECK: Call graph node for function: '__kmpc_fork_call'<<{{.*}}>>  #uses=3
 ; CHECK:   CS<None> calls external node
 ;
 ; CHECK: Call graph node for function: 'd'<<{{.*}}>>  #uses=2
 ; CHECK:   CS<None> calls external node

llvm/test/Transforms/SCCP/arg-count-mismatch.ll

-; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --function-signature --check-attributes --check-globals
 ; RUN: opt < %s -ipsccp -S -o - | FileCheck %s
 
 ; The original C source looked like this:
 ;
 ;   long long a101, b101, e101;
 ;   volatile long c101;
 ;   int d101;
 ;

 ; This ll reproducer has been reduced to only include tha call.
 ;
 ; Note that -lint will report this as UB, but it passes -verify.
 
 ; This test is just to verify that we do not crash/assert due to mismatch in
 ; argument count between the caller and callee.
 
 define dso_local void @foo(i16 %a) {
-; CHECK-LABEL: @foo(
-; CHECK-NEXT:    [[CALL:%.*]] = call i16 bitcast (i16 (i16, i16)* @bar to i16 (i16)*)(i16 [[A:%.*]])
+; CHECK-LABEL: define {{[^@]+}}@foo
+; CHECK-SAME: (i16 [[A:%.*]]) {
+; CHECK-NEXT:    [[CALL:%.*]] = call i16 bitcast (i16 (i16, i16)* @bar to i16 (i16)*)(i16 [[A]])
 ; CHECK-NEXT:    ret void
 ;
   %call = call i16 bitcast (i16 (i16, i16) * @bar to i16 (i16) *)(i16 %a)
   ret void
 }
 
 define internal i16 @bar(i16 %p1, i16 %p2) {
-; CHECK-LABEL: @bar(
-; CHECK-NEXT:    ret i16 0
+; CHECK-LABEL: define {{[^@]+}}@bar
+; CHECK-SAME: (i16 [[P1:%.*]], i16 [[P2:%.*]]) {
+; CHECK-NEXT:    unreachable
 ;
   ret i16 0
 }
 
 ;-------------------------------------------------------------------------------
 ; Additional tests to verify that we still optimize when having a mismatch
 ; in argument count due to varargs (as long as all non-variadic arguments have
 ; been provided),
 
 define internal i16 @vararg_prop(i16 %p1, ...) {
-; CHECK-LABEL: @vararg_prop(
+; CHECK-LABEL: define {{[^@]+}}@vararg_prop
+; CHECK-SAME: (i16 [[P1:%.*]], ...) {
 ; CHECK-NEXT:    ret i16 undef
 ;
   ret i16 %p1
 }
 
 define dso_local i16 @vararg_tests(i16 %a) {
-; CHECK-LABEL: @vararg_tests(
-; CHECK-NEXT:    [[CALL1:%.*]] = call i16 (i16, ...) @vararg_prop(i16 7, i16 8, i16 [[A:%.*]])
+; CHECK-LABEL: define {{[^@]+}}@vararg_tests
+; CHECK-SAME: (i16 [[A:%.*]]) {
+; CHECK-NEXT:    [[CALL1:%.*]] = call i16 (i16, ...) @vararg_prop(i16 7, i16 8, i16 [[A]])
 ; CHECK-NEXT:    [[CALL2:%.*]] = call i16 bitcast (i16 (i16, i16, ...)* @vararg_no_prop to i16 (i16)*)(i16 7)
 ; CHECK-NEXT:    [[ADD:%.*]] = add i16 7, [[CALL2]]
 ; CHECK-NEXT:    ret i16 [[ADD]]
 ;
   %call1 = call i16 (i16, ...) @vararg_prop(i16 7, i16 8, i16 %a)
   %call2 = call i16 bitcast (i16 (i16, i16, ...) * @vararg_no_prop to i16 (i16) *) (i16 7)
   %add = add i16 %call1, %call2
   ret i16 %add
 }
 
 define internal i16 @vararg_no_prop(i16 %p1, i16 %p2, ...) {
-; CHECK-LABEL: @vararg_no_prop(
-; CHECK-NEXT:    ret i16 [[P1:%.*]]
+; CHECK-LABEL: define {{[^@]+}}@vararg_no_prop
+; CHECK-SAME: (i16 [[P1:%.*]], i16 [[P2:%.*]], ...) {
+; CHECK-NEXT:    unreachable
 ;
   ret i16 %p1
 }
 

llvm/test/Transforms/SCCP/arg-type-mismatch.ll

-; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --function-signature --check-attributes --check-globals
 ; RUN: opt < %s -ipsccp -S -o - | FileCheck %s
 
 ; This test is just to verify that we do not crash/assert due to mismatch in
 ; argument type between the caller and callee.
 
 define dso_local void @foo(i16 %a) {
-; CHECK-LABEL: @foo(
-; CHECK-NEXT:    [[CALL:%.*]] = call i16 bitcast (i16 (i16, i16)* @bar to i16 (i16, i32)*)(i16 [[A:%.*]], i32 7)
+; CHECK-LABEL: define {{[^@]+}}@foo
+; CHECK-SAME: (i16 [[A:%.*]]) {
+; CHECK-NEXT:    [[CALL:%.*]] = call i16 bitcast (i16 (i16, i16)* @bar to i16 (i16, i32)*)(i16 [[A]], i32 7)
 ; CHECK-NEXT:    ret void
 ;
   %call = call i16 bitcast (i16 (i16, i16) * @bar to i16 (i16, i32) *)(i16 %a, i32 7)
   ret void
 }
 
 define internal i16 @bar(i16 %p1, i16 %p2) {
-; CHECK-LABEL: @bar(
-; CHECK-NEXT:    ret i16 [[P2:%.*]]
+; CHECK-LABEL: define {{[^@]+}}@bar
+; CHECK-SAME: (i16 [[P1:%.*]], i16 [[P2:%.*]]) {
+; CHECK-NEXT:    unreachable
 ;
   ret i16 %p2
 }