Re: [Fortran] RFC: Patch to support STDCALL etc. (PR 34112) * PING *

[FX <fxcoudert at gmail dot com>]

> The problem is that the symbols are not name-decorated on the  
> Fortran side:


I've investigated a bit, using my cross-compiler, and found the  
following:

  * i386_pe_mangle_decl_assembler_name() (in config/i386/winnt.c) is  
where the decoration happens; it's never called when compiling  
Fortran, but called when compiling equivalent C code

  * in C, this is done from lhd_set_decl_assembler_name() (in  
langhooks.c), which calls targetm.mangle_decl_assembler_name (which,  
on i386 PE targets, is i386_pe_mangle_decl_assembler_name)

  * in Fortran, we simply use SET_DECL_ASSEMBLER_NAME directly (in  
trans-common.c and trans-decl.c), without allowing further mangling/ 
decoration


To fix this, I propose to add a function gfc_set_decl_assembler_name,  
which calls targetm.mangle_decl_assembler_name first and then uses  
SET_DECL_ASSEMBLER_NAME. All curent uses of SET_DECL_ASSEMBLER_NAME  
should probably be switched to gfc_set_decl_assembler_name().

The patch attached does what I describe above. I have been able to  
test that it does build for my cross-compiler (i386-darwin to i586- 
mingw32), gives object files with properly decorated function names  
and not warning at link time. The resulting executable (from Dennis'  
test) runs fine in a VirtualBox, and does not hit any abort. I am not  
able to bootstrap (nor regtest) a native compiler on a non-win32  
platform, which should be tested before applying the patch. I don't  
expect any tricky issue though, as the only targets using the  
mangle_decl_assembler_name target hook are i386-winnt (cygwin and  
mingw) and i386-netware (don't know what that is).

I hope this helps, sorry I can't do more.

FX


PS: one misc thing I noted: add_attributes_to_decl() has a misleading  
name: it doesn't add anything, it simply builds a list of the  
appropriate attributes.

Attachment: stdcall.diff
Description: Binary data