> Heh.  Well then - the original patch is ok.

Thanks.  However, I simplified the original testcase far too much so the fix 
isn't general enough, sorry about that.  The new testcase (attached) involves 
nested GIMPLE_TRY_FINALLYs and GIMPLE_BINDs.

So, in the end, we really need to track things.  I think it's better to think 
in terms of fallthruness since there is already a lot of code in gimple-low.c 
that does this.  The patch fixes a typo in the process, whereby GIMPLE_SWITCH 
is deemed may_fallthru by the code of gimple_stmt_may_fallthru whereas the 
comment just above explains why it is not.  It also attempts to remove only 
unreachable return statements, that's sufficient and probably safer.

Tested on x86_64-suse-linux, OK for mainline?


2009-10-18  Eric Botcazou  <ebotcazou@adacore.com>

        * gimple-low.c (struct lower_data): Add cannot_fallthru field.
	(lower_stmt) <GIMPLE_BIND>: Add comment.
	<GIMPLE_COND, GIMPLE_GOTO, GIMPLE_SWITCH>: Set cannot_fallthru to true
	and return.
	<GIMPLE_RETURN>: Remove the statement if cannot_fallthru is set.
	Otherwise lower it and set cannot_fallthru to true.
	<GIMPLE_TRY>: Update cannot_fallthru for GIMPLE_TRY_FINALLY.  Set it
	to false otherwise.  Return.
	<GIMPLE_CATCH, GIMPLE_EH_FILTER>; Set cannot_fallthru to false.
	<GIMPLE_CALL>: Set cannot_fallthru to false for BUILT_IN_SETJMP and
	to true for a noreturn call.  Do not remove statements.
	<GIMPLE_OMP_PARALLEL, GIMPLE_OMP_TASK>: Set cannot_fallthru to false.
	Set cannot_fallthru to false on function exit.
	(gimple_stmt_may_fallthru) <GIMPLE_SWITCH>: Really return false.
	<GIMPLE_ASSIGN>: Remove.


2009-10-18  Eric Botcazou  <ebotcazou@adacore.com>

        * gnat.dg/noreturn2.ad[sb]: New test.


-- 
Eric Botcazou