[Committed] Fix 54524: Wrong code with some 64bit addition with registers as 32bits

Andrew Pinski andrew.pinski@caviumnetworks.com
Fri Nov 2 23:32:00 GMT 2012


Hi,
  For some reason cse produced:
(insn 27 43 28 2 (set (reg:SI 231 [+4 ])
        (plus:SI (reg:SI 229 [+4 ])
            (const_int 0 [0]))) t.c:24 10 {*addsi3}
     (nil))

(insn 28 27 29 2 (set (reg:SI 212)
        (ltu:SI (reg:SI 231 [+4 ])
            (reg:SI 229 [+4 ]))) t.c:24 521 {*sltu_sisi}
     (nil))

And then forwprop goes and does the following:
In insn 28, replacing
 (ltu:SI (reg:SI 231 [+4 ])
        (reg:SI 229 [+4 ]))
 with (const_int 1 [0x1])

This is due to simplify-rtx.c (simplify_relational_operation_1) Having
the following optimization:
  /* (LTU/GEU (PLUS a C) C), where C is constant, can be simplified to
     (GEU/LTU a -C).  Likewise for (LTU/GEU (PLUS a C) a).  */

But this is wrong when C is 0.  Anyways I committed this patch as
obvious to fix the issue in simplify-rtx.c but I have not looked into
why CSE does simplify the plus.

Thanks,
Andrew Pinski

ChangeLog:

        PR rtl-opt/54524
        * simplify-rtx.c (simplify_relational_operation_1): Don't simplify
        (LTU/GEU (PLUS a 0) 0) into (GEU/LTU a 0) since they are not equivalent.
-------------- next part --------------
Index: simplify-rtx.c
===================================================================
--- simplify-rtx.c	(revision 193110)
+++ simplify-rtx.c	(working copy)
@@ -4546,7 +4546,9 @@ simplify_relational_operation_1 (enum rt
       && GET_CODE (op0) == PLUS
       && CONST_INT_P (XEXP (op0, 1))
       && (rtx_equal_p (op1, XEXP (op0, 0))
-	  || rtx_equal_p (op1, XEXP (op0, 1))))
+	  || rtx_equal_p (op1, XEXP (op0, 1)))
+      /* (LTU/GEU (PLUS a 0) 0) is not the same as (GEU/LTU a 0). */
+      && XEXP (op0, 1) != const0_rtx)
     {
       rtx new_cmp
 	= simplify_gen_unary (NEG, cmp_mode, XEXP (op0, 1), cmp_mode);


More information about the Gcc-patches mailing list