cause of all the mainline Sparc regressions

["David S. Miller" <davem at redhat dot com>]

It is this change:

+Wed Mar  6 10:59:39 CET 2002  Jan Hubicka  <jh@suse.cz>
+
+       * cfgcleanup.c (try_optimize_cfg): Kill blocks.
+       (try_optimize_cfg): Do not update liveness.
+       (cleanup-cfg): Loop until try_optimize_cfg and dead code
+       removal stabilizes; use delete_trivially_dead_insns.

Specifically the bit that makes delete_trivially_dead_insns
get invoked from cleanup_cfg.

It avoids this when CLEANUP_PRE_SIBCALL is set, but there is
another case where we either have to avoid delete_trivially_dead_insns
or make it smarter.  That is when we are being called from
finish_eh_generation().

The reason is that the dead insn removal does not understand
REG_SAVE_AREA notes, they are like uses of the register
mentioned in such notes.  Specifically the calls to count_reg_usage
does not look for these notes.

When this bug triggers, alloca()'s occuring in code using EH regions
subtract from the stack pointer using uninitialized registers because:

1) cleanup_cfg from finish_eh_generation blew away all references
   to the register in the REG_SAVE_AREA note

2) optimize_save_area_alloca transformed the instructions to
   use the REG_SAVE_AREA register, but all the code to initialize
   that register has been deleted by #1

I would rather not add yet another CLEANUP_* flag to check for
this, because auditing for all of the cleanup_cfg calls that occur
before optimize_save_area_alloca is nontrivial.  I'd rather use
some flag set by toplev when it gets past the
optimize_save_area_alloca call.

Jan, any preferences for a fix?

Maybe the best thing to do is to make delete_trivially_dead_insns look
for the REG_SAVE_AREA notes.  I wonder if there is other stuff that is
getting confused up by how these REG_SAVE_AREA notes work.