On 03/01/13 22:39, Andrew Pinski wrote:
Hi,
For aarch64, we don't CSE some cmp away. This patch fixes the case
where we are CSE across some basic-blocks like:
int f(int a, int b)
{
if(a<b)
return 1;
if(a>b)
return -1;
return 0;
}
--- CUT ---
To fix this, I implemented the target hook
TARGET_FIXED_CONDITION_CODE_REGS as there was already code in CSE
which uses this target hook to find the extra setting of the CC
registers.
OK? Build and tested on aarch64-thunder-elf (using Cavium's internal
simulator). Build a cross to aarch64-thunder-linux-gnu and a Canadian
cross with that one for the native toolchain.
Thanks,
Andrew Pinski
* config/aarch64/aarch64.c (aarch64_fixed_condition_code_regs):
New function.
(TARGET_FIXED_CONDITION_CODE_REGS): Define.
* gcc.target/aarch64/cmp-1.c: New testcase.
fixed_condition_code.diff.txt
* config/aarch64/aarch64.c (aarch64_fixed_condition_code_regs):
New function.
(TARGET_FIXED_CONDITION_CODE_REGS): Define.
* gcc.target/aarch64/cmp-1.c: New testcase.
Index: config/aarch64/aarch64.c
===================================================================
--- config/aarch64/aarch64.c (revision 194872)
+++ config/aarch64/aarch64.c (working copy)
@@ -3041,6 +3041,16 @@ aarch64_const_double_zero_rtx_p (rtx x)
return REAL_VALUES_EQUAL (r, dconst0);
}
+/* Return the fixed registers used for condition codes. */
+
+static bool
+aarch64_fixed_condition_code_regs (unsigned int *p1, unsigned int *p2)
+{
+ *p1 = CC_REGNUM;
+ *p2 = -1;
Please use INVALID_REGNUM here (as the documentation states).
Otherwise, OK.
A backport to the AArch64-4.7 branch would be appreciated.
R.
