Patch: 3334280

[Dale Johannesen]

This patch prevents combine from generating a compare to an invalid 
subreg, as
discussed in my earlier mail on this bug (at bottom).  Running 
testsuite.

Index: combine.c
===================================================================
RCS file: /cvs/repository/CoreTools/gcc3/gcc/combine.c,v
retrieving revision 1.22.2.3
diff -u -d -b -w -r1.22.2.3 combine.c
--- combine.c   2003/04/23 00:14:16     1.22.2.3
+++ combine.c   2003/07/25 21:38:20
@@ -11360,6 +11360,7 @@
        if (have_insn_for (COMPARE, tmode))
         {
           int zero_extended;
+         rtx new_op0, new_op1;

           /* If the only nonzero bits in OP0 and OP1 are those in the
              narrower mode and this is an equality or unsigned 
comparison,
@@ -11392,11 +11393,20 @@
                                   gen_lowpart_for_combine (tmode,
                                                            XEXP (op0, 
1)));

-             op0 = gen_lowpart_for_combine (tmode, op0);
+             /* APPLE LOCAL ban undefined SUBREG */
+             /* We can't accept a (necessarily paradoxical) SUBREG
+                here, as the high bits of it are undefined. */
+             new_op0 = gen_lowpart_for_combine (tmode, op0);
+             new_op1 = gen_lowpart_for_combine (tmode, op1);
+             if (GET_CODE (new_op0) != SUBREG
+                 && (op1 == new_op1 || GET_CODE (new_op1) != SUBREG))
+               {
+                 op0 = new_op0;
+                 op1 = new_op1;
               if (zero_extended && GET_CODE (op1) == CONST_INT)
                 op1 = GEN_INT (INTVAL (op1) & GET_MODE_MASK (mode));
-             op1 = gen_lowpart_for_combine (tmode, op1);
               break;
+               }
             }

           /* If this is a test for negative, we can make an explicit

> This is a bad-codegen regression from 3.1, still broken in FSF 3.3 and 
> 3.4.
> The RTL starts to differ as far back as expand, and I could track that 
> down,
> but the new RTL looks correct for a while.  The actual problem seems 
> to be
> introduced by combine, which merges
>
> (insn 28 27 30 0 0x0 (set (reg:SI 129)
>         (sign_extend:SI (reg:QI 127))) 21 {extendqisi2_ppc} (insn_list 
> 27 (nil))
>     (expr_list:REG_DEAD (reg:QI 127)
>         (nil)))
>
> (insn 30 28 31 0 0x0 (set (reg:CC 131)
>         (compare:CC (reg:SI 129)
>             (const_int -1 [0xffffffff]))) 426 {*cmpsi_internal1} 
> (insn_list 28 (nil))
>     (expr_list:REG_DEAD (reg:SI 129)
>         (nil)))
>
> into
>
> (insn 30 28 31 0 0x0 (set (reg:CC 131)
>         (compare:CC (subreg:SI (reg:QI 127) 0)
>             (const_int 255 [0xff]))) 426 {*cmpsi_internal1} (insn_list 
> 27 (nil))
>     (expr_list:REG_DEAD (reg:QI 127)
>         (nil)))
>
> This seems wrong; it is doing a 32-bit comparison, but in the 
> subreg:SI(QI)
> expression, the high 24 bits are undefined as I read the docs.  But 
> I'm not on
> firm ground with subreg's, so I want to get confirmation before 
> proceeding...