Re: c/7871: ICE on legal code, global register variables problems

[Jim Wilson <wilson at tuliptree dot org>]

Let me recap a bit.  The current code calls mark_set_1 (...CLOBBER...)
and mark_used_regs.  This causes CALL_INSNs to get a REG_UNDEAD note for
a global_reg, which confuses combine into deleting the insn that
initializes the global_reg.

If I change this to mark_set_1 (...SET...), then there are no
REG_DEAD/REG_UNUSED notes which is correct.  However, there are
LOG_LINKS from the insn that sets the global_reg to the CALL_INSN, and
from the CALL_INSN to the insn that uses the global_reg.  This causes
combine to merge the 3 instructions together, and the result is that the
first insn that sets the global_reg gets simplified away.  This could
perhaps be fixed by modifying combine to know that calls and global_regs
are special.

If I avoid calling mark_set_1, then there are no REG_DEAD/REG_UNUSED
notes, but there is still a LOG_LINK from the first insn that sets the
global_reg to the CALL_INSN.  When combine merges an unrelated insn into
the CALL_INSN, it sees a LOG_LINK that doesn't appear to belong on the
CALL_INSN, so it moves it to the next insn which uses the global_reg. 
Combine then merges two instructions that set/use the global_reg, and
the first global_reg set before the call disappears again.

Thus it seems that in order to get the right behavior, we need to avoid
adding any REG_NOTES or LOG_LINKS for global_regs.

I had to go back to gcc-2.95.1 to find a compiler that worked for my two
testcases, and it has this behavior.  The change that broke it is:

2000-04-07  Richard Henderson  <rth at cygnus dot com>

	* flow.c (loop_depth): Remove.
	(reg_next_use, cc0_live, mem_set_list): Replace with ...
	...
	(mark_set_reg): Break out of mark_set_1.
	...
I didn't manage to find the mail message for this patch.  This command
	cvs diff -p -r 1.247 -r 1.248 flow.c
will show you the patch if you want to look at it.

This is a large change that was mostly reorganizing code, but another
thing it does is delete one line of code in mark_set_1 that handles
global_regs like the frame pointer and arg pointer register.  If I re-add
this one line of code, then both testcases I have start working again.

I am off on a trip, so it will be a while before I am able to continue
working on this.


2003-03-10  James E Wilson  <wilson at tuliptree dot org>

	* flow.c (mark_set_1): Handle global_regs like the frame pointer.

Index: flow.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/flow.c,v
retrieving revision 1.549
diff -p -r1.549 flow.c
*** flow.c	28 Feb 2003 10:11:47 -0000	1.549
--- flow.c	10 Mar 2003 05:05:08 -0000
*************** mark_set_1 (pbi, code, reg, cond, insn, 
*** 2709,2714 ****
--- 2709,2715 ----
  #if FRAME_POINTER_REGNUM != ARG_POINTER_REGNUM
        && ! (regno_first == ARG_POINTER_REGNUM && fixed_regs[regno_first])
  #endif
+       && ! (regno_first < FIRST_PSEUDO_REGISTER && global_regs[regno_first])
        )
      {
        int some_was_live = 0, some_was_dead = 0;