Where does LLVM mangle characters from llvm-ir names while generating native code?

So I was taking a look at Microsoft C++ ABI support while on vacation,
and ran into a major issue. Given the following llvm-ir:

$ clang++ -S -emit-llvm -O3 mangling.cpp -o - -Xclang -cxx-abi -Xclang microsoft
; ModuleID = 'mangling.cpp'
target datalayout =
"e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-f80:128:128-v64:64:64-
v128:128:128-a0:0:64-f80:32:32-n8:16:32-S32"
target triple = "i686-pc-win32"

define i32 @"?heyimacxxfunction@@YAHXZ"() nounwind readnone {
entry:
  ret i32 42
}

Note the ?heyimacxxfunction@@YAHXZ. Now if I generate assembly (using
clang or llc) I get:

$ clang++ -S -O3 mangling.cpp -o - -Xclang -cxx-abi -Xclang microsoft
        .def __3F_heyimacxxfunction@@YAHXZ;
        .scl 2;
        .type 32;
        .endef
        .text
        .globl __3F_heyimacxxfunction@@YAHXZ
        .align 16, 0x90
__3F_heyimacxxfunction@@YAHXZ: # @"?heyimacxxfunction@@YAHXZ"
# BB#0: # %entry
        movl $42, %eax
        ret

It turned the ? into _3F_, and prepended _ someplace. I get the same
symbol if I use integrated-as to generate an object file.

I've been unable thus far to find out exactly where this occurs. And
once I do find it, I'm not sure of the correct fix. I assume that LLVM
mangles the '?' because it means something special in the gas syntax.
There also has to be a way to communicate to llvm from clang that this
symbol should not receive a prepended _.

Thanks,

- Michael Spencer

Looks to me like it converted the ? into the ascii hexadecimal representation 3F. I don’t think another underscore was pre-pended.

This is probably thanks to lib/Target/Mangler.cpp. You’ll want to let ? be treated as an acceptable character.

static bool isAcceptableChar(char C, bool AllowPeriod) {
if ((C < ‘a’ || C > ‘z’) &&
(C < ‘A’ || C > ‘Z’) &&
(C < ‘0’ || C > ‘9’) &&
C != ‘_’ && C != ‘$’ && C != ‘@’ &&
!(AllowPeriod && C == ‘.’))
return false;
return true;
}

But you might want to add doesAllowQuestionMarksInName() to MCAsmInfo, it’s possible that other name manglers don’t accept question marks.

Cheers,

Joe Abbey
Software Architect
Arxan Technologies, Inc.
1305 Cumberland Ave, Ste 215
West Lafayette, IN 47906
jabbey@arxan.com
www.arxan.com

it’s possible that other name manglers don’t accept question marks.

I mean assemblers/linkers.

Dur,

Joe

Joe Abbey
Software Architect
Arxan Technologies, Inc.
1305 Cumberland Ave, Ste 215
West Lafayette, IN 47906
jabbey@arxan.com
www.arxan.com

I've been unable thus far to find out exactly where this occurs. And
once I do find it, I'm not sure of the correct fix. I assume that LLVM
mangles the '?' because it means something special in the gas syntax.
There also has to be a way to communicate to llvm from clang that this
symbol should not receive a prepended _.

You can prepend a \01. That is what is done with inline assembly.

I would love to remove all mangling from the LLVM IL, but last I
mentioned sabre wanted to keep at the lest the prepending of '_'. Not
sure about the rest (like expanding ?).

Thanks,

- Michael Spencer

Cheers,
Rafael

So I was taking a look at Microsoft C++ ABI support while on vacation,
and ran into a major issue. Given the following llvm-ir:

$ clang++ -S -emit-llvm -O3 mangling.cpp -o - -Xclang -cxx-abi -Xclang microsoft
; ModuleID = 'mangling.cpp'
target datalayout =
"e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-f80:128:128-v64:64:64-
v128:128:128-a0:0:64-f80:32:32-n8:16:32-S32"
target triple = "i686-pc-win32"

define i32 @"?heyimacxxfunction@@YAHXZ"() nounwind readnone {
entry:
ret i32 42
}

Note the ?heyimacxxfunction@@YAHXZ. Now if I generate assembly (using
clang or llc) I get:

$ clang++ -S -O3 mangling.cpp -o - -Xclang -cxx-abi -Xclang microsoft
       .def __3F_heyimacxxfunction@@YAHXZ;
       .scl 2;
       .type 32;
       .endef
       .text
       .globl __3F_heyimacxxfunction@@YAHXZ
       .align 16, 0x90
__3F_heyimacxxfunction@@YAHXZ: # @"?heyimacxxfunction@@YAHXZ"
# BB#0: # %entry
       movl $42, %eax
       ret

It turned the ? into _3F_, and prepended _ someplace. I get the same
symbol if I use integrated-as to generate an object file.

I've been unable thus far to find out exactly where this occurs. And
once I do find it, I'm not sure of the correct fix. I assume that LLVM
mangles the '?' because it means something special in the gas syntax.
There also has to be a way to communicate to llvm from clang that this
symbol should not receive a prepended _.

I could swear that when I wrote that code I stuck a '\1' character in front of the name to prevent LLVM's mangler from doing anything to it (see lib/AST/MicrosoftMangle.cpp:162). (I knew the prepending of '_' characters was a problem, but I didn't know LLVM magically transformed chars it doesn't like into _xx_.) That is a magic marker that says "this is the literal name of this symbol, don't mangle it in the usual way."

It looks like for some reason that the '\1' character isn't getting emitted. I wish I had an answer for you, but I can't debug this because I don't run Windows. Works fine for me here on Mac. Running your command to generate LLVM IR, I get:

; ModuleID = 'mangling.cpp'
target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f80:128:128-n8:16:32-S128"
target triple = "i386-apple-macosx10.6.8"

define i32 @"\01?heyimacxxfunction@@YAHXZ"() nounwind readnone ssp {
  ret i32 42
}

Note the '\01'.

Chip

In the case I posted I had removed that line, however, you still get the __3F_ in the generated assembly with it.

In the case I posted I had removed that line, however, you still get the __3F_ in the generated assembly with it.

Huh. It only seems to happen with a Windows triple or a Linux triple. Doesn't happen with a Mac triple, though--probably because the Darwin assembler supports quoted symbols (i.e. you can enclose an identifier in quotes). Maybe Joe is right, and you should change the LLVM mangler to not mangle '?' chars (if the assembler can handle it). Or maybe we should see if GAS supports quoted symbols (MCAsmInfo::doesAllowQuotesInName()).

Chip

In the case I posted I had removed that line, however, you still get the __3F_ in the generated assembly with it.

Huh. It only seems to happen with a Windows triple or a Linux triple. Doesn't happen with a Mac triple, though--probably because the Darwin assembler supports quoted symbols (i.e. you can enclose an identifier in quotes). Maybe Joe is right, and you should change the LLVM mangler to not mangle '?' chars (if the assembler can handle it). Or maybe we should see if GAS supports quoted symbols (MCAsmInfo::doesAllowQuotesInName()).

Chip

So setting AllowQuotesInName=true gets the correct names in the
output. However, it breaks MinGW. The issue is of course that
MCAsmInfo mixes both assembler and object file format capabilities.
Allowing quoted names has nothing to do with COFF ;/. So it seems we
need a new AsmInfo class for "real" Windows. Any other ideas on how to
implement this?

- Michael Spencer

So setting AllowQuotesInName=true gets the correct names in the
output. However, it breaks MinGW. The issue is of course that
MCAsmInfo mixes both assembler and object file format capabilities.
Allowing quoted names has nothing to do with COFF ;/. So it seems we
need a new AsmInfo class for "real" Windows. Any other ideas on how to
implement this?

Something like what we do for cfi support in the assembly is probably
reasonable. I would suggest

* add a flag like -fno-asm-quotes, default it to false since MC is
fairly common these days.
* Remove AllowQuotesInName
* Instead of mangling names when given -fno-asm-quotes, just error if
the string has unprintable characters.

- Michael Spencer

Cheers,
Rafael