This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
RFA: tune arm_rtx_costs function for xscale multiply
- From: James Lemke <jim at wasabisystems dot com>
- To: gcc-patches at gcc dot gnu dot org
- Date: 21 Nov 2003 13:53:41 -0500
- Subject: RFA: tune arm_rtx_costs function for xscale multiply
- Organization:
For expediency, I have split my earlier patch.
This portion omits the changes to arm.md.
2003-07-07 James Lemke <jim@wasabisystems.com>
* config/arm/arm.c (arm_rtx_costs): Improve for xscale multiply.
* testsuite/gcc.dg/arm-g2.c: New file.
Index: gcc/config/arm/arm.c
===================================================================
RCS file: /cvsroot/gcc/gcc/gcc/config/arm/arm.c,v
retrieving revision 1.294
diff -u -r1.294 arm.c
--- gcc/config/arm/arm.c 7 Oct 2003 08:49:35 -0000 1.294
+++ gcc/config/arm/arm.c 9 Oct 2003 13:40:15 -0000
@@ -3315,19 +3315,41 @@
{
unsigned HOST_WIDE_INT i = (INTVAL (XEXP (x, 1))
& (unsigned HOST_WIDE_INT) 0xffffffff);
- int add_cost = const_ok_for_arm (i) ? 4 : 8;
- int j;
+ int cost, const_ok = const_ok_for_arm (i);
+ int j, booth_unit_size;
+
+ if (arm_tune_xscale)
+ {
+ unsigned HOST_WIDE_INT masked_const;
+
+ /* The cost will be related to two insns.
+ First a load of the constant (MOV or LDR), then a multiply. */
+ cost = 2;
+ if (! const_ok)
+ cost += 1; /* LDR is probably more expensive because
+ of longer result latency. */
+ masked_const = i & 0xffff8000;
+ if (masked_const != 0 && masked_const != 0xffff8000)
+ {
+ masked_const = i & 0xf8000000;
+ if (masked_const == 0 || masked_const == 0xf8000000)
+ cost += 1;
+ else
+ cost += 2;
+ }
+ return cost;
+ }
/* Tune as appropriate. */
- int booth_unit_size = ((tune_flags & FL_FAST_MULT) ? 8 : 2);
-
+ cost = const_ok ? 4 : 8;
+ booth_unit_size = ((tune_flags & FL_FAST_MULT) ? 8 : 2);
for (j = 0; i && j < 32; j += booth_unit_size)
{
i >>= booth_unit_size;
- add_cost += 2;
+ cost += 2;
}
- return add_cost;
+ return cost;
}
return (((tune_flags & FL_FAST_MULT) ? 8 : 30)
And the new file: gcc/testsuite/gcc.dg/arm-g2.c
===================================================================
/* Verify that hardware multiply is preferred on XScale. */
/* { dg-do compile { target xscale*-*-* } } */
/* { dg-options "-mcpu=xscale -O" } */
/* Brett Gaines' test case. */
unsigned BCPL(unsigned) __attribute__ ((naked));
unsigned BCPL(unsigned seed)
{
/* Best code would be:
ldr r1, =2147001325
ldr r2, =715136305
mla r0, r1, r0, r2
mov pc, lr */
return seed * 2147001325U + 715136305U;
}
/* We want to suppress running for -mthumb but not for -mthumb-interwork. */
/* { dg-final { global compiler_flags; if ![string match "*-mthumb *" $compiler_flags] { scan-assembler "mla\[ ].*" } } } */
--
James Lemke jim@wasabisystems.com Orillia, Ontario
http://www.wasabisystems.com