c/7574: Code generation problem involving regparm attribute

>Number:         7574
>Category:       c
>Synopsis:       Code generation problem involving regparm attribute
>Confidential:   no
>Severity:       serious
>Priority:       medium
>Responsible:    unassigned
>State:          open
>Class:          wrong-code
>Submitter-Id:   net
>Arrival-Date:   Sun Aug 11 17:46:01 PDT 2002
>Closed-Date:
>Last-Modified:
>Originator:     Eric R. Krause
>Release:        gcc 3.2 20020809 (prerelease)
>Organization:
>Environment:
Cygwin 1.3.12, Windows 2000 Professional SP2 (5.0.2195)
AMD Athlon Processor
>Description:
gcc configuration:
Reading specs from /usr/local/lib/gcc-lib/i686-pc-cygwin/3.2/specs
Configured with: /usr/src/gcc-3.2-20020809/configure --with-gcc --with-gnu-as 
--with-gnu-ld --enable-shared --program-suffix=-3.2 --enable-threads=win32 
--disable-win32-registry --disable-nls --enable-languages=c,c++ 
--enable-version-specific-runtime-libs
Thread model: win32
gcc version 3.2 20020809 (prerelease)

Details:
gcc 3.2 produces incorrect code for situations where a function defined with 
the regparm (1) attribute is called via a pointer, but only if that pointer is 
stored in an array AND the array subscript is used as the pointer.

What I mean by using the array subscript as a pointer, is that the function 
would be called with an expression like:
  fnPtrArray[x](...);

Note that retrieving the pointer from the array, assigning it to a function 
pointer variable, and then dispatching the call using the pointer variable DOES 
work as expected with gcc 3.2:
  fnPtr = fnPtrArray[x];
  fnPtr(...);

Bug:
gcc 3.2 seems not to recognize that the function being called has the regparm 
attribute: as a matter of fact, it PUSHES the intended argument value on the 
stack, loads the pointer into %eax (where the argument value SHOULD be) and 
then does a 'call *%eax'.

gcc 2.95.x would use a different register to set up the call-through-pointer; 
%eax contained the correct argument value.

I suspect this problem is also present in gcc 3.1 and gcc 3.1.1.
>How-To-Repeat:
In the attached tarball, there are two files: x86codebug.c and x86codebug.i.  
The latter is the pre-processed output generated under Cygwin.

Compile x86codebug.c _with gcc 3.2_ and run it.  The first line of output 
should read "Should be 100: 100", but it won't when you run it.
>Fix:
Workaround: Assign the pointer array element to a temporary, then dispatch the 
call using the temporary.
>Release-Note:
>Audit-Trail:
>Unformatted: