Re: [PATCH/RFC] PR other/22313: Hot/cold sections vs. dwarf2 (take 2)

[David Daney <ddaney at avtrex dot com>]

Roger Sayle wrote:

> 2006-08-31  Roger Sayle  <roger@eyesopen.com>
>
> 	PR other/22313
> 	* dwarf2out.c (add_fde_cfi): Use a set_loc if the current label is
> 	NULL, otherwise use an advance_loc4 to adjust relative to the
> 	current label.
> 	(output_cfi) <DW_CFA_set_loc>: Update the current label.
> 	(dwarf2out_switch_text_section): Reset the current label to avoid
> 	using advance_loc4 over section boundaries.

This breaks exception handling on mipsel-linux-gnu.  I did a regression 
hunt and r116604 (this patch) fails, where as r116603 works.

The problem is that with this patch applied the first instruction in a 
FDE is typically DW_CFA_set_loc, before the patch it was DW_CFA_advance_loc.

The mips linker changes the CIE augmentation to be PC relative when it 
generates a shared object (and perhaps at other times as well, but I 
have not investigated exactly where it happens).  This transformation 
does not effect the DW_CFA_advance_loc instrunctions as they already use 
relative addressing.  But the transformation fails with DW_CFA_set_loc 
as it uses an absolute address and is not easily transformed.

It is possible that this problem needs to be addressed in the linker, 
here is a message I posted to the binutils list further discussing the 
problem:

http://sourceware.org/ml/binutils/2006-09/msg00131.html

In any event there is a regression for mipsel-linux.  libgcj is totally 
broken, C++ exceptions are broken.


David Daney