assembler output customization

[Trass3r]

Bugzilla Link	17465
Version	trunk
OS	All
CC	@gnzlbg,@jryans,@nico

Extended Description

It would be nice to have if clang supported gcc's -masm=dialect switch so I can get readable asm without a special case for clang to use -mllvm --x86-asm-syntax=intel.

But more importantly I'd love to have a way to generate assembler output with properly embedded source code, especially on Linux. But that would even allow to support the /FAs switch in clang-cl and enable me to seamlessly use my existing VS addin for quick asm inspection with the newly available clang toolkit.

I already tried several different approaches.

• -fverbose-asm is pretty useless.
• Then there's the -Wa trick (http://www.fclose.com/240/generate-a-mixed-source-and-assembly-listing-using-gcc) which only works with gcc.
• objdump works with both: clang++ -c -g -O3 test.cpp objdump -dSC -M intel test.o

But neither of these approaches reaches the readability of /FAs (in combination with AsmHighlighter):

?set@Point@@QEAAXNNN@Z PROC ; Point::set, COMDAT ; 47 : void set(const double X, const double Y, const double Z) { _xyz[0] = X; _xyz[1] = Y; _xyz[2] = Z; }

movsdx    QWORD PTR [rcx], xmm1
movsdx    QWORD PTR [rcx+8], xmm2
movsdx    QWORD PTR [rcx+16], xmm3
ret       0

?set@Point@@QEAAXNNN@Z ENDP ; Point::set

In this case source code is properly injected as assembler comments and there are no line numbers or machine code confusing the syntax highlighter. But the /FA switch also allows you to add these if you need.

[nico]

Cross-ref "[cfe-dev] RFC: Add a way to interleave source code in assembler output" and https://reviews.llvm.org/D30897 / https://reviews.llvm.org/D30898 linked from there (apparently stalled, but maybe a good start).

See also bug 39360.

[Trass3r]

MCAsmStreamer::EmitDwarfLocDirective is a POI for this.

[nico]

fwiw, -masm=intel has been working since 208683 (clang 3.5).

[54aefcd4-c07d-4252-8441-723563c8826f]

+1 I would also like to generate source code annotated with assembly in a particular dialect. This is kind of a must to verify the assembly generated for a particular snippet of code within a large project.

[nico]

Here's a patch for -masm: http://lists.cs.uiuc.edu/pipermail/cfe-commits/Week-of-Mon-20140505/105299.html

For the other thing, -fno-integrated-as -Xassembler -adhln might do the trick.

[nickdesaulniers]

-fverbose-asm support in clang has come up at least twice so far on LKML (this being most recent).

@efriedma-quic mentions in https://reviews.llvm.org/D30897#inline-268745 that clang should have the source buffer around, but I thought clang was pretty aggressive about discarding the AST from memory ASAP? I think @aeubanks had some changes recently related to -clear-ast-before-backend? https://reviews.llvm.org/D111270 https://lists.llvm.org/pipermail/cfe-dev/2021-September/068930.html

[aeubanks]

I'm not sure if the AST is needed in this case, perhaps just the source buffer is enough? But if you do need the AST, you could skip passing -clear-ast-before-backend if -fverbose-asm.

[efriedma-quic]

You shouldn't need the AST, just the source code. We don't discard the source code; it's usually just mmap'ed anyway.