[llvm] r237876 - AsmPrinter: Compute absolute label difference directly

[Duncan P. N. Exon Smith]

> On 2015 Jun 4, at 14:24, David Blaikie <dblaikie at gmail.com> wrote:
> 
> 
> 
> On Thu, Jun 4, 2015 at 2:11 PM, Duncan P. N. Exon Smith <dexonsmith at apple.com> wrote:
> 
>> On 2015 Jun 4, at 12:48, David Blaikie <dblaikie at gmail.com> wrote:
>> 
>> 
>> 
>> On Thu, Jun 4, 2015 at 12:22 PM, Duncan P. N. Exon Smith <dexonsmith at apple.com> wrote:
>> 
>>> On 2015-Jun-04, at 11:48, David Blaikie <dblaikie at gmail.com> wrote:
>>> 
>>> 
>>> 
>>> On Wed, May 20, 2015 at 7:41 PM, Duncan P. N. Exon Smith <dexonsmith at apple.com> wrote:
>>> Author: dexonsmith
>>> Date: Wed May 20 21:41:23 2015
>>> New Revision: 237876
>>> 
>>> URL: http://llvm.org/viewvc/llvm-project?rev=237876&view=rev
>>> Log:
>>> AsmPrinter: Compute absolute label difference directly
>>> 
>>> Create a low-overhead path for `EmitLabelDifference()` that emits a
>>> emits an absolute number when (1) the output is an object stream and (2)
>>> the two symbols are in the same data fragment.
>>> 
>>> This drops memory usage on Mach-O from 975 MB down to 919 MB (5.8%).
>>> The only call is when `!doesDwarfUseRelocationsAcrossSections()` --
>>> i.e., on Mach-O -- since otherwise an absolute offset from the start of
>>> the section needs a relocation.  (`EmitLabelDifference()` is cheaper on
>>> ELF anyway, since it creates 1 fewer temp symbol, and it gets called far
>>> less often.  It's not clear to me if this is even a bottleneck there.)
>>> 
>>> (I'm looking at `llc` memory usage on `verify-uselistorder.lto.opt.bc`;
>>> see r236629 for details.)
>>> 
>>> Modified:
>>>    llvm/trunk/include/llvm/MC/MCObjectStreamer.h
>>>    llvm/trunk/include/llvm/MC/MCStreamer.h
>>>    llvm/trunk/lib/CodeGen/AsmPrinter/AsmPrinter.cpp
>>>    llvm/trunk/lib/MC/MCObjectStreamer.cpp
>>> 
>>> Modified: llvm/trunk/include/llvm/MC/MCObjectStreamer.h
>>> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/include/llvm/MC/MCObjectStreamer.h?rev=237876&r1=237875&r2=237876&view=diff
>>> ==============================================================================
>>> --- llvm/trunk/include/llvm/MC/MCObjectStreamer.h (original)
>>> +++ llvm/trunk/include/llvm/MC/MCObjectStreamer.h Wed May 20 21:41:23 2015
>>> @@ -136,6 +136,17 @@ public:
>>>   void EmitZeros(uint64_t NumBytes) override;
>>>   void FinishImpl() override;
>>> 
>>> +  /// Emit the absolute difference between two symbols if possible.
>>> +  ///
>>> +  /// Emit the absolute difference between \c Hi and \c Lo, as long as we can
>>> +  /// compute it.  Currently, that requires that both symbols are in the same
>>> +  /// data fragment.  Otherwise, do nothing and return \c false.
>>> +  ///
>>> +  /// \pre Offset of \c Hi is greater than the offset \c Lo.
>>> +  /// \return true on success.
>>> +  bool emitAbsoluteSymbolDiff(const MCSymbol *Hi, const MCSymbol *Lo,
>>> +                              unsigned Size) override;
>>> +
>>>   bool mayHaveInstructions() const override {
>>>     return getCurrentSectionData()->hasInstructions();
>>>   }
>>> 
>>> Modified: llvm/trunk/include/llvm/MC/MCStreamer.h
>>> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/include/llvm/MC/MCStreamer.h?rev=237876&r1=237875&r2=237876&view=diff
>>> ==============================================================================
>>> --- llvm/trunk/include/llvm/MC/MCStreamer.h (original)
>>> +++ llvm/trunk/include/llvm/MC/MCStreamer.h Wed May 20 21:41:23 2015
>>> @@ -652,6 +652,15 @@ public:
>>>                                      unsigned Isa, unsigned Discriminator,
>>>                                      StringRef FileName);
>>> 
>>> +  /// Emit the absolute difference between two symbols if possible.
>>> +  ///
>>> +  /// \pre Offset of \c Hi is greater than the offset \c Lo.
>>> +  /// \return true on success.
>>> +  virtual bool emitAbsoluteSymbolDiff(const MCSymbol *Hi, const MCSymbol *Lo,
>>> +                                      unsigned Size) {
>>> +    return false;
>>> +  }
>>> +
>>>   virtual MCSymbol *getDwarfLineTableSymbol(unsigned CUID);
>>>   virtual void EmitCFISections(bool EH, bool Debug);
>>>   void EmitCFIStartProc(bool IsSimple);
>>> 
>>> Modified: llvm/trunk/lib/CodeGen/AsmPrinter/AsmPrinter.cpp
>>> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/CodeGen/AsmPrinter/AsmPrinter.cpp?rev=237876&r1=237875&r2=237876&view=diff
>>> ==============================================================================
>>> --- llvm/trunk/lib/CodeGen/AsmPrinter/AsmPrinter.cpp (original)
>>> +++ llvm/trunk/lib/CodeGen/AsmPrinter/AsmPrinter.cpp Wed May 20 21:41:23 2015
>>> @@ -1591,6 +1591,10 @@ void AsmPrinter::EmitInt32(int Value) co
>>> /// .set if it avoids relocations.
>>> void AsmPrinter::EmitLabelDifference(const MCSymbol *Hi, const MCSymbol *Lo,
>>>                                      unsigned Size) const {
>>> 
>>> Is this function only used for DWARF emission? It seems sufficiently generally named that that's not obvious & could be surprising for other users to find this \/ behavior?
>>> 
>> 
>> Yes, it's only used for DWARF emission; at least, I looked, and only
>> found DWARF callers.  I'll look into changing the name.  I agree it's
>> confusing.
>> 
>> Wonder what we use for label difference in inline asm or other things, then... *shrug*
> 
> Inline asm doesn't use this, anyway.
> 
> However, I just did a more thorough look (I was starting to work on
> the patch) and found a single non-DWARF caller:
> `EHStreamer::emitExceptionTable()`.  I'll fix that pronto; looks
> like I should leave `EmitLabelDifference()` around, create a new
> `emitDwarfLabelDifference()` that eventually calls the former, and
> change all the callers but that one.
> 
>> 
>> 
>> Any suggestions?  Maybe, `emitDwarfLabelDifference()`?
>> 
>> Sounds plausible - can we assert that it's in one of the blessed sections that MachO doesn't link into the executable, too?
> 
> Seems like a good idea, but I don't know how to check that.  Do you?
> 
> I don't know what logic MachO uses to bless the sections - whether it's name based or what.

I think it's just a static set, but it might be name-based.

Adrian?  Fred?

> If it's just debug_* sections, you should be able to get the section from the symbol (hmm, only if we've already emitted the symbol which often/usually isn't the case?) and the name from the section, etc.

I *think* what matters is section we're emitting the offset into, not
the section that the symbols are from.  Fred?