This is the mail archive of the mailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Committed: fix PR37939, CRIS regression in code quality (insn not used)

See PR37939.  The cost of the MULT made the cost of the combined
shift-and-add higher than the individual insns.  The benefit of
this fix is highly code-dependent, with results in the range
0.2-7.0% improvement but with no regressions; it depends on how
much focus the code has on indexing 2 or 4-byte arrays. :)
(BIAP is the name of the prefix insn implementing the mode
pre-CRIS v32; better not have "addi" in the file or its name
other than in the scan-assembler directive.)

Right, the patterns (and combine!) should probably use ASHIFT,
not MULT; see comment #1, but fixing that would be unsuitable at
this stage.  Note also that there'd then be assummetry for the
side-effect addressing modes (a parallel assignment of the
address to a fourth register):
  [(set reg0 (mem (plus (mult reg1 2) reg2)))
   (set reg3 (plus (ashift reg1 1) reg2))])

The fix was regtested on trunk and gcc-4_3-branch for
cris-axis-elf and crisv32-axis-linux-gnu.  Committed, trunk and

	PR target/37939
	* New test.

	PR target/37939
	* config/cris/cris.c (cris_rtx_costs) <MULT>: Return 0 for an ADDI

--- /dev/null	2008-02-22 17:23:55.052000250 +0100
+++	2008-10-28 17:23:14.000000000 +0100
@@ -0,0 +1,11 @@
+/* Make sure ADDI is combined and emitted successfully.
+   See also PR37939.  */
+/* { dg-do compile } */
+/* { dg-options "-O2" } */
+/* { dg-final { scan-assembler "addi" } } */
+/* { dg-final { scan-assembler-not "lsl" } } */
+int xyzzy (int r10, int r11)
+  return r11 * 4 + r10;

Index: config/cris/cris.c
--- config/cris/cris.c	(revision 141344)
+++ config/cris/cris.c	(working copy)
@@ -1795,6 +1795,21 @@
       return true;
     case MULT:
+      /* If we have one arm of an ADDI, make sure it gets the cost of
+	 one insn, i.e. zero cost for this operand, and just the cost
+	 of the PLUS, as the insn is created by combine from a PLUS
+	 and an ASHIFT, and the MULT cost below would make the
+	 combined value be larger than the separate insns.  The insn
+	 validity is checked elsewhere by combine.
+	 FIXME: this case is a stop-gap for 4.3 and 4.4, this whole
+	 function should be rewritten.  */
+      if (outer_code == PLUS && BIAP_INDEX_P (x))
+	{
+	  *total = 0;
+	  return true;
+	}
       /* Identify values that are no powers of two.  Powers of 2 are
          taken care of already and those values should not be changed.  */
       if (!CONST_INT_P (XEXP (x, 1))

brgds, H-P

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]