Re: O0 cleanup_cfg vs debuggability

[James E Wilson <wilson at specifixinc dot com>]

Olivier Hainque wrote:
> This optimization may affect debuggability even at O0, which is a bit
> annoying. For instance, on x86-linux a breakpoint on the "break" statement in
> the code below:

In gcc-3.4, I think you can solve this by redefining what a forwarder 
block is, or how we handle them.  See forwarder_block_p in cfganal.c. 
Normally, a forwarder block is something that contains no active insns 
other than a branch, and we optimize them away whenever possible to 
simplify the cfg.  However, if we want full debugging info, then a block 
that contains a line number note should not be considered a forwarder 
block, or else, we should not be optimizing away forwarder blocks.  The 
rest of what the jump pass is doing is probably not relevant to your 
problem.

In gcc-3.5 however, we are already screwed before the first rtl pass. 
As a side effect of the gimple to rtl transformation, we don't bother to 
emit the empty block that is only needed to hold the line number note. 
So we would have to fix the gimple to rtl transformation first to 
preserve this info, then fix the cleanup_cfg code to not optimize away 
blocks that contain line number notes.  Try compiling with -da 
-fdump-tree-all and look at the last tree dump and the first rtl dump.
--
Jim Wilson, GNU Tools Support, http://www.SpecifixInc.com