c/8028: GCC 3.x __builtin_apply() passes wrong arguments

[achurch at achurch dot org]

>Number:         8028
>Category:       c
>Synopsis:       GCC 3.x __builtin_apply() passes wrong arguments
>Confidential:   no
>Severity:       serious
>Priority:       medium
>Responsible:    unassigned
>State:          open
>Class:          wrong-code
>Submitter-Id:   net
>Arrival-Date:   Tue Sep 24 19:36:01 PDT 2002
>Closed-Date:
>Last-Modified:
>Originator:     Andrew Church
>Release:        3.2
>Organization:
>Environment:
System: Linux crystal 2.4.18 #9 SMP Mon Sep 2 16:03:12 JST 2002 i686 unknown
Architecture: i686

	
host: i686-pc-linux-gnu
build: i686-pc-linux-gnu
target: i686-pc-linux-gnu
configured with: /usr/src/gcc-3.2/configure --prefix=/packages/gcc3 --program-suffix=3 --enable-languages=c,c++
>Description:
Calling a function via __builtin_apply() passes the wrong arguments to the
function if another function which takes arguments has been called
previously.
>How-To-Repeat:
gcc3 -O0 foo.c -o foo
./foo ; echo $?
--------------- BEGIN foo.c ---------------
int foo(int n)
{
    return n;
}

int bar(int n)
{
    return n+1;
}

int quux(int n)
{
    foo(0);
    __builtin_return(__builtin_apply((void *)bar, __builtin_apply_args(), 64));
}

int main(int argc, char **argv)
{
    return quux(argc);
}
---------------- END foo.c ----------------
The program should exit with a value of argc+1, but instead exits with a
stack-dependent return value (I get 99).  The following part of the
assembly listing for quux() looks suspicious:
	...
	movl	$0, (%esp)
	call	foo		 <-- call foo(0)
	movl	(%ebx), %edx
	movl	%esp, -172(%ebp)
	subl	$64, %esp	 <-- space for __builtin_apply() arguments?
	leal	4(%esp), %eax	 <-- copy arguments to ESP+4 (BUG???)
	...
The offset of "4" in the last line above seems to reflect the amount of
argument space required for foo()--if I pass foo() two parameters, it
changes to 8, if three parameters then 12, etc.
>Fix:
Move everything but the __builtin_apply() call into a separate function
that takes no parameters, and call that function before calling
__builtin_apply().  (It seems to be safe to declare the new function
"static inline".)
>Release-Note:
>Audit-Trail:
>Unformatted: