new-regalloc vs alpha

[Richard Henderson <rth at redhat dot com>]

I've run into two problems so far trying this on alphaev6-linux.

One, you should never try to allocate fixed registers, even if they
are in the proper register class.  Fixed with attachment one.

Two, reload has a sort of phase ordering problem with local-alloc.
I'm not sure how this doesn't show up with current code.  The 
sequence of events is

  * The code around line 1140 in update_equiv_regs decides to re-emit the
    set of a reg near it's (only) use, we lose the memoized INSN_CODE
    that we'd had before.  

  * The allocator chooses a hard reg that is not valid for that insn.

  * reload calls alter_reg for each pseudo, which updates the REGNO
    of each pseudo, which means the bogus hard reg is reflected in
    the insn.

  * eliminate_regs_in_insn calls recog_memoized, which fails.  We abort.

The quickest way around this that I saw is to have update_equiv_regs
copy the insn code while re-emitting the instruction.

I've not committed either of these patches.

You may wish to examine this case further (_fixunsdfsi), as the
allocation is particularly bletcherous, and results in quite a
few bogus fp->int and int->fp copies.



r~
	* new-regalloc.c (assign_regs): Don't consider fixed regs.

Index: new-regalloc.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/Attic/new-regalloc.c,v
retrieving revision 1.1.2.6
diff -c -p -d -r1.1.2.6 new-regalloc.c
*** new-regalloc.c	2001/01/30 04:52:49	1.1.2.6
--- new-regalloc.c	2001/01/30 11:34:41
*************** void assign_regs()
*** 1322,1327 ****
--- 1322,1328 ----
        VARRAY_POP(selectStack);
        COPY_HARD_REG_SET(okColors, reg_class_contents[reg_preferred_class(currReg)]);
        IOR_HARD_REG_SET(okColors, reg_class_contents[reg_alternate_class(currReg)]);
+       AND_COMPL_HARD_REG_SET (okColors, fixed_reg_set);
        
        EXECUTE_IF_SET_IN_SBITMAP(regInfo[currReg]->adjList, 0, entry, 
        {
	* local-alloc.c (update_equiv_regs): Copy INSN_CODE when
	re-emitting an equivalence.

Index: local-alloc.c
===================================================================
RCS file: /cvs/gcc/egcs/gcc/local-alloc.c,v
retrieving revision 1.77
diff -u -p -r1.77 local-alloc.c
--- local-alloc.c	2001/01/01 17:20:08	1.77
+++ local-alloc.c	2001/01/30 12:03:56
@@ -1141,8 +1141,15 @@ update_equiv_regs ()
 		  rtx new_insn;
 
 		  new_insn = emit_insn_before (PATTERN (equiv_insn), insn);
-		  REG_NOTES (PREV_INSN (insn)) = REG_NOTES (equiv_insn);
+		  REG_NOTES (new_insn) = REG_NOTES (equiv_insn);
 		  REG_NOTES (equiv_insn) = 0;
+
+		  /* ??? We may not recognize this insn before 
+		     eliminate_regs_in_insn, which is called after
+		     alter_regs.  And if the allocators have chosen
+		     a bum register class, recognition will at that
+		     point fail, and we will abort.  */
+		  INSN_CODE (new_insn) = INSN_CODE (equiv_insn);
 
 		  PUT_CODE (equiv_insn, NOTE);
 		  NOTE_LINE_NUMBER (equiv_insn) = NOTE_INSN_DELETED;

_fixunsdfsi.i.bz2