This is an archive of the discontinued LLVM Phabricator instance.

Use the MachineBasicBlock symbol for a callbr target
ClosedPublic

Authored by void on Jul 17 2019, 2:46 PM.

Details

Summary

Inline asm doesn't use labels when compiled as an object file. Therefore, we
shouldn't create one for the (potential) callbr destination. Instead, use the
symbol for the MachineBasicBlock.

Diff Detail

Repository
rL LLVM

Event Timeline

void created this revision.Jul 17 2019, 2:46 PM
Herald added a project: Restricted Project. · View Herald TranscriptJul 17 2019, 2:46 PM
void added a comment.Jul 17 2019, 2:59 PM

Test?

On its way. :-)

void updated this revision to Diff 210432.Jul 17 2019, 3:06 PM

Add testcase.

void updated this revision to Diff 210433.Jul 17 2019, 3:08 PM

Fix bad update.

xbolva00 added inline comments.Jul 17 2019, 3:10 PM
test/CodeGen/X86/callbr-asm-obj-file.ll
3 ↗(On Diff #210433)

; RUN:

I'm curious too, what exactly this solves, maybe a test will elucidate.

lib/CodeGen/AsmPrinter/AsmPrinterInlineAsm.cpp
435 ↗(On Diff #210430)

range based for?

for (const MachineBasicBlock& MBB : MF)
...

might even be able to replace MF with the expression above and still fit on 80 lines.

lib/CodeGen/AsmPrinter/AsmPrinterInlineAsm.cpp
433 ↗(On Diff #210433)

getBasicBlock() would be a more explicit method to call, preferable to getOperand(1).

void updated this revision to Diff 210435.Jul 17 2019, 3:20 PM
void marked 4 inline comments as done.

Fix syntax errors and use better accessor methods.

lib/CodeGen/AsmPrinter/AsmPrinterInlineAsm.cpp
433 ↗(On Diff #210433)

Good call! :-)

435 ↗(On Diff #210430)

I tried that, but never was able to get the compiler to believe that MF had a "begin()" and "end()". Eventually just gave up... I'll try again though.

lib/CodeGen/AsmPrinter/AsmPrinterInlineAsm.cpp
435 ↗(On Diff #210430)

maybe *MF?

void marked an inline comment as done.Jul 17 2019, 3:26 PM
void added inline comments.
lib/CodeGen/AsmPrinter/AsmPrinterInlineAsm.cpp
435 ↗(On Diff #210430)

That gives me this error *sigh*:

error: use of deleted function ‘llvm::MachineBasicBlock::MachineBasicBlock(const llvm::MachineBasicBlock&)’
lib/CodeGen/AsmPrinter/AsmPrinterInlineAsm.cpp
435 ↗(On Diff #210430)

That sounds like you forgot to make the iteration variable a reference, so a deleted copy constructor is being invoked.

diff --git a/llvm/lib/CodeGen/AsmPrinter/AsmPrinterInlineAsm.cpp b/llvm/lib/CodeGen/AsmPrinter/AsmPrinterInlineAsm.cpp
index 7721e996aca5..692877bdeb84 100644
--- a/llvm/lib/CodeGen/AsmPrinter/AsmPrinterInlineAsm.cpp
+++ b/llvm/lib/CodeGen/AsmPrinter/AsmPrinterInlineAsm.cpp
@@ -429,6 +429,10 @@ static void EmitGCCInlineAsmStr(const char *AsmStr, const MachineInstr *MI,
           // PrintAsmOperand?
           if (Modifier[0] == 'l') { // Labels are target independent.
             if (MI->getOperand(OpNo).isBlockAddress()) {
+              const MachineFunction *MF = MI->getParent()->getParent();
+              for (const MachineBasicBlock& MBB : *MF) {
+                errs() << "XXX\n";
+              }

compiles just fine for me.

void updated this revision to Diff 210453.Jul 17 2019, 4:23 PM

Use for-each loop.

void marked an inline comment as done.Jul 17 2019, 4:23 PM
test/CodeGen/X86/callbr-asm-obj-file.ll
7 ↗(On Diff #210453)

Can you add

; CHECK-NOT:

cases so I get a better idea of what LLVM used to print before this change?
(So we have an affirmative test case, and one negative test case)

void marked an inline comment as done.Jul 17 2019, 4:37 PM
void added inline comments.
test/CodeGen/X86/callbr-asm-obj-file.ll
7 ↗(On Diff #210453)

It used to emit an error with no output.

test/CodeGen/X86/callbr-asm-obj-file.ll
1 ↗(On Diff #210453)

is -pc-linux actually recognized? I would have expected x86_64-linux-gnu?

void updated this revision to Diff 210460.Jul 17 2019, 4:55 PM

Update triple.

void marked an inline comment as done.Jul 17 2019, 4:56 PM
void updated this revision to Diff 210466.Jul 17 2019, 5:13 PM

Update test with correct dest labels

lib/CodeGen/AsmPrinter/AsmPrinterInlineAsm.cpp
436 ↗(On Diff #210466)

I'm not sure that either side of this if are being tested/covered by your new test in callbr-asm-obj-file.ll.

In particular, the case that worries me is passing the address of a label as an input to asm goto. Consider:

asm goto(".quad %l0\n\t.quad %l1" :: "i"(&&baz) :: bar);
        bar:;
        baz:;

https://godbolt.org/z/lKa_HD

where 2 blockaddresses are parameters to a callbr, one is the address of a label, one is the jump target list.

I'd be more comfortable signing off on code review if that case was also tested here (and considered in further changes related to callbr)

void marked an inline comment as done.Jul 18 2019, 4:59 PM
void added inline comments.
lib/CodeGen/AsmPrinter/AsmPrinterInlineAsm.cpp
436 ↗(On Diff #210466)

We need to get clang to accept your code snippet first. :-)

I'm not sure I understand what you're saying here. What are you worried will be the outcome?

Despite the looping and whatnot, what this change does is fairly straightforward: instead of getting the IR basic block and creating a label for it, it uses the machine basic block's label directly. Similar to the code on line 436-437.

void updated this revision to Diff 210714.Jul 18 2019, 5:31 PM

Git-ifying my change.

nickdesaulniers accepted this revision.Jul 18 2019, 5:53 PM
nickdesaulniers added inline comments.
lib/CodeGen/AsmPrinter/AsmPrinterInlineAsm.cpp
436 ↗(On Diff #210466)

Sorry, maybe a better example:

int foo(void) {
    asm goto(".quad %l0\n\t.quad %l1" :: "i"(&&baz) :: bar);
    baz: return 1;
    bar: return -1;
}

before your patch clang -O2 x.c -c -S produces:

foo:                                    # @foo
# %bb.0:
	movl	$1, %eax
	#APP
	.quad .Ltmp0
	.quad .Ltmp1
	#NO_APP
	jmp	.LBB0_2
.Ltmp1:                                 # Block address taken
.LBB0_1:
	movl	$-1, %eax
.Ltmp0:                                 # Block address taken
.LBB0_2:
	retq

after:

foo:                                    # @foo
# %bb.0:
	movl	$1, %eax
	#APP
	.quad .LBB0_2
	.quad .LBB0_1
	#NO_APP
	jmp	.LBB0_2
.Ltmp0:                                 # Block address taken
.LBB0_1:
	movl	$-1, %eax
.Ltmp1:                                 # Block address taken
.LBB0_2:
	retq

So that looks like an improvement (and there's more we can do in a follow up to be better).

Before your patch clang -O2 x.c -c fails outright, with it, it succeeds and the disassembly looks good. Ok, sorry for the noise, LGTM.

I think it would be good for us further to not emit the .LtmpX labels, since they both point to the same location.

This revision is now accepted and ready to land.Jul 18 2019, 5:53 PM
void added a comment.Jul 18 2019, 6:01 PM

Thanks! Though your example pointed out a separate problem: not accepting the label as a valid operand. I'll do some work on that.

This revision was automatically updated to reflect the committed changes.
xbolva00 added a subscriber: hans.Jul 19 2019, 2:31 AM

It would be nice to merge this fix into 9.0 branch after some time in trunk. What do you think?

@hans @void

It would be nice to merge this fix into 9.0 branch after some time in trunk. What do you think?

@hans @void

So two regressions were immediately reported:

  1. https://github.com/ClangBuiltLinux/linux/issues/614
  2. https://github.com/ClangBuiltLinux/linux/issues/615

so honestly I'd rather revert it until we understand the exact issue better.

@void has also authored https://reviews.llvm.org/D64982.

For the life of me I swear there's a bug that @craig.topper commented on that explains an issue with symbols not being added to the symbol table at the MC layer, but I cannot find it (and it's killing me). I think that should be pursued (why were we generating a .tmp symbol anyways).

For the life of me I swear there's a bug that @craig.topper commented on that explains an issue with symbols not being added to the symbol table at the MC layer, but I cannot find it (and it's killing me). I think that should be pursued (why were we generating a .tmp symbol anyways).

https://bugs.llvm.org/show_bug.cgi?id=42147

llvm/trunk/lib/CodeGen/AsmPrinter/AsmPrinterInlineAsm.cpp
434

So another issue I though of with this approach:

It makes sense to me as implemented; we need some mapping from BlockAddress of the BasicBlock to the corresponding MachineBasicBlock (we have the reverse mapping; given a MachineBasicBlock we can get the BasicBlock it comes from, but we don't have the reverse mapping; GetBlockAddressSymbol kind of does this but has other issues). This approach works for the test case provided, and probably all C/C++ that could ever hit this path BUT it is not safe to assume that the MachineBasicBlock corresponding to the BlockAddress is within this MachineFunction.

Stated another way, I could have blockaddress(@foo, %bar) used in MachineFunction @baz. The code in this patch would iterate the MachineBasicBlocks of the MachineFunction corresponding to @baz since the MachineInstruction using the blockaddress constant was in @baz. It would never find %bar since it needs to iterate MachineBasicBlocks of @foo, not @baz.

I think this would be fixed by initializing MF to BB->getParent().

Reverted in https://reviews.llvm.org/rL366600. While GetBlockAddressSymbol has some complicated machinery related to managing symbols of temporary addresses, I get the feeling a complete fix will lie in modifications there.

hans added a comment.Jul 22 2019, 1:37 PM

It would be nice to merge this fix into 9.0 branch after some time in trunk. What do you think?

@hans

Sounds like this was reverted on trunk. Please let me know if there are follow-ups that might be candidates for merging.

void added a comment.Jul 22 2019, 9:30 PM

It would be nice to merge this fix into 9.0 branch after some time in trunk. What do you think?

@hans

Sounds like this was reverted on trunk. Please let me know if there are follow-ups that might be candidates for merging.

I have a patch in my local branch that should fix some of the issues. However, I can't seem to support the situation Nick pointed out. If you have something like:

define void @foo() {
entry:
  callbr void asm sideeffect ".word ${0:l}", "X"(i8* blockaddress(@bar, %next))
          to label %if.end10 [label %l_yes]

if.end10:
  br label %l_yes

l_yes:
  ret void
}

define i32 @bar() {
entry:
  br label %next

next:
  ret i32 42
}

This will fail, because @bar hasn't been created yet, so it can't get the address of next. Or, rather, it *can*, but the assembly parser won't realize that it's a label and will create a new label for it. So it goes from this:

foo:
    .word .Ltmp0
  ...
bar:
.Ltmp0:

to this:

foo:
    .word .Ltmp00  ; <- Notice the '0' suffix, which is applied to a label to uniquify it.
  ...
bar:
.Ltmp0:

when parsing the inline assembly. I *think* we need to convince MCContext that .Ltmp0 is a label that it doesn't need to uniquify—i.e., register it as a symbol in MCContext instead of treating it as a temporary label (there's a difference, and I'm not sure why).

The question is whether this is a case we need to care about at this point in time.

Thoughts?