backend should generate right name for longjmp on netbsd

[jsonn]

Bugzilla Link	8765
Resolution	FIXED
Resolved on	May 13, 2011 16:17
Version	trunk
OS	All
CC	@lattner

Extended Description

NetBSD is using asm("foo") renames in the system headers to provide multiple versions of certain functions for ABI compatibility. One such instance is the longjmp family.

If clang is used with -fbuiltin, the renames for the longjmp functions are lost in the code generation. This directly results in link errors due to missing functions.

[lattner]

As written, this bug is "behaves correctly". The real fix is a backend fix: if longjmp has a different name on netbsd, then the code generator should always generate the right name, independently of asm renaming.

[jsonn]

The problem seems to be that the renaming is not represented on the IR level for the builtin. In theory a complex program linked with LTO could end up with different longjmp versions, so this isn't a fixed mapping. I haven't checked yet if other builtins are affected.

[jsonn]

This is the only remaining critical issue for NetBSD, so it would be nice to address this before LLVM 2.9.

I can see one short term fix and a long term fix. Short term fix would be to not recognize a function as builtin, if it has an assembler name set. At worst, this should result in missed optimisations, not bad code. Having support to control if a specific function is builtin would be another option and useful for kernel use.

The long term fix would be to either keep the mapping intact, even when using the intrinsic (making the real function to call an attribute of the intrinsic call) or to emit a mapping table of intrinsic name to library implementation. That is going to require a lot more changes, one way or the other though.

[lattner]

Jörg, please attach a .bc file that shows the setjmp and a snippet of the assembly generated, along with what is wrong. I'd like to know precisely what the right and wrong names are. Thanks!

[jsonn]

src/bin/sh/main.c preprocessed
This is the preprocessed main.c file as generated by clang -emit-llvm -O2 -save-temps ...

[jsonn]

[Bytecode version of srchttps://user-images.githubusercontent.com/296817/143742421-52f77396-209f-4cdd-9711-a82d527b3f4e.gz)
This is the generated bytecode by clang -emit-llvm -O2.

Notice that it uses _setjmp etc., when the assembler symbol in the pre-processed input is (main.i:2687) is __setjmp14.

[llvmbot]

The problem seems to be with longjmp(), not __asm handling.
Example C-source code (no #includes):

typedef long jmp_buf[14];
void jmp(jmp_buf, int) __asm("__jmp14");
void longjmp(jmp_buf, int) __asm("__longjmp14");

void test()
{
jmp((void *)0, 0);
longjmp((void *)0, 0);
}

GNU C generates (via cc -S):

    .section        __TEXT,__text,regular,pure_instructions
    .globl  _test

_test:
Leh_func_begin1:
subq $8, %rsp
Ltmp0:
xorl %edi, %edi
xorl %esi, %esi
callq __jmp14
xorl %edi, %edi
xorl %esi, %esi
callq __longjmp14
Leh_func_end1:

Clang generates (via clang -S):

    .section        __TEXT,__text,regular,pure_instructions
    .globl  _test

_test: ## @​test
Leh_func_begin0:

BB#0:

    pushq   %rax

Ltmp0:
xorl %edi, %edi
xorl %esi, %esi
callq __jmp14
xorl %edi, %edi
xorl %esi, %esi
callq _longjmp
Leh_func_end0:

(I am changing Platform to "All", since this is not NetBSD-specific problem. The above have been generated on Mac OS X 10.6.7 with LLVM/Clang built with today's sources.)

[jsonn]

Adam, the problem is the combination of the LIBBUILTION with the __asm renaming. The presence of the former overrides the latter.

[jsonn]

http://lists.cs.uiuc.edu/pipermail/cfe-commits/Week-of-Mon-20110502/041658.html has a possible fix.

[llvmbot]

The patch fixes longjmp, but breaks strerror (an probably more) on Darwin:

char *strerror(int) __asm("_" "strerror");
void test() {
strerror(1);
}

Original Clang gives:

    .section        __TEXT,__text,regular,pure_instructions
    .globl  _test
    .align  4, 0x90

_test: ## @​test
Ltmp0:
.cfi_startproc

BB#0:

    movl    $1, %edi
    jmp     _strerror               ## TAILCALL

Ltmp1:
.cfi_endproc
Leh_func_end0:

Patched Clang:

    .section        __TEXT,__text,regular,pure_instructions
    .globl  _test
    .align  4, 0x90

_test: ## @​test
Ltmp0:
.cfi_startproc

BB#0:

    movl    $1, %edi
    jmp     __strerror              ## TAILCALL

Ltmp1:
.cfi_endproc
Leh_func_end0:

[jsonn]

Updated patch to deal with name mangling issues

[llvmbot]

This is it! Good work! :)

[jsonn]

Updated patch to deal with name mangling issues
Remove dead variable.

[lattner]

This patch looks great to me, please commit it with a testcase, thanks!

[jsonn]

Fixed with r131311.