res=__builtin_setjmp(); -> invalid assembler

[llvmbot]

Bugzilla Link	18974
Version	trunk
OS	Linux
Attachments	minimal example
Reporter	LLVM Bugzilla Contributor
CC	@hfinkel,@bogner

Extended Description

__builtin_setjmp() only works when we immediately switch on its result and branch into two different code paths. If instead we write this kind of code:

int res = __builtin_setjmp();
do_this_in_all_cases();

then, at least on x86-64, it produces this kind of assembler for the first line:

movq    %rbp, -80(%rbp)
movq    %rsp, -64(%rbp)
movq    $".LBB4_-1", -72(%rbp)

where ".LBB4_-1" is a bogus label not defined anywhere in the .s file. It fails to assemble, of course.

[bogner]

Hmm, looks like that fails too though. We're doing something wrong here.

[bogner]

This is undefined behaviour. From C99 (n1256) 7.13:

4. An invocation of the setjmp macro shall appear only in one of the following contexts:

   — the entire controlling expression of a selection or iteration statement;

   — one operand of a relational or equality operator with the other operand an integer constant expression, with the resulting expression being the entire controlling expression of a selection or iteration statement;

   — the operand of a unary ! operator with the resulting expression being the entire controlling expression of a selection or iteration statement; or

   • the entire expression of an expression statement (possibly cast to void).

5. If the invocation appears in any other context, the behavior is undefined.

If you want to ignore __builtin_setjmp()'s result, you should write:

(void)__builtin_setjmp(...);

[llvmbot]

Note that the C standard says specifically this:

s = setjmp(env);  /* WRONG! */

It's of little relevance here. If you consider it is, please view this bug report as complaining about the following kind of code, which is explicitly allowed by the C standard:

__builtin_setjmp(..);   /* ignore the return value */
do_this_in_all_cases();

Or:

if (__builtin_setjmp(..) == 1)
     ;   /* nothing */
do_this_in_all_cases();

[llvmbot]

A workaround that should always work: replace a bare __builtin_setjmp(buf) line with:

while (__builtin_setjmp(buf)) { }

Here, LLVM cannot optimize the two paths back into one, unlike e.g. the same statement with "if" instead of "while".

It means that after a __builtin_longjmp() we will repeat the __builtin_setjmp() once, which is at worst a minor performance problem, and which can be considered a good thing to do if you're (rarely but) possibly doing several longjmp() to go back to this point, like in some search algorithm.