On 6/30/2026 12:19 PM, Radosav Krunic wrote:
It tunes ifcvt parameters so that we get if-conversion in more cases.
gcc/
* config/mips/mips.cc (mips_rtx_costs): Reduce cost of
if_then_else pattern.
(mips_max_noce_ifcvt_seq_cost): New function. Decrease
maximum permissible cost for the unconditional sequence which
should be generated during if-conversion (for all non-r6
targets). This disables if-conversion for non-r6 targets in
branch-cost-1.c test.
(mips_noce_conversion_profitable_p): New function.
(TARGET_MAX_NOCE_IFCVT_SEQ_COST): Define hook.
(TARGET_NOCE_CONVERSION_PROFITABLE_P): Define hook.
gcc/testsuite/
* gcc.target/mips/branch-cost-1.c: Disable for -Os.
Note this is MIPS specific, so in an ideal world someone more familiar
with MIPS targets would step in with an ACK or list of things to fix.
In the absence of that feedback...
gcc/config/mips/mips.cc | 56 +++++++++++++++++++
gcc/testsuite/gcc.target/mips/branch-cost-1.c | 2 +-
2 files changed, 57 insertions(+), 1 deletion(-)
diff --git a/gcc/config/mips/mips.cc b/gcc/config/mips/mips.cc
index 277ec419826..81a41b8ee3a 100644
--- a/gcc/config/mips/mips.cc
+++ b/gcc/config/mips/mips.cc
@@ -67,6 +67,7 @@ along with GCC; see the file COPYING3. If not see
#include "rtl-iter.h"
#include "flags.h"
#include "opts.h"
+#include "ifcvt.h"
/* This file should be included last. */
#include "target-def.h"
@@ -4795,6 +4796,12 @@ mips_rtx_costs (rtx x, machine_mode mode, int outer_code,
break;
}
return false;
+
+ case IF_THEN_ELSE:
+ if (reg_or_0_operand (XEXP (x, 1), VOIDmode)
+ || reg_or_0_operand (XEXP (x, 2), VOIDmode))
+ *total = 0;
+ return false;
Formatting looks wrong here. Make sure the case IF_THEN_ELSE lines up
with other cases in that switch. The IF should be indented two spaces
relative to the case label and so-on following GNU style guidelines.
+
+/* Return true if SEQ is a good candidate as a replacement for the
+ if-convertible sequence described in IF_INFO. */
+
+static bool
+mips_noce_conversion_profitable_p (rtx_insn *seq, struct noce_if_info *if_info)
+{
+ struct noce_if_info mips_if_info = *if_info;
+
+ for (rtx_insn *insn = seq; insn; insn = NEXT_INSN (insn))
+ {
+ /* Compensate target-specific costs by incrementing original_cost
+ (mips_if_info.original_cost) and max_seq_costif
(mips_if_info.ax_seq_costif) */
+ break;
+ }
+
+ return default_noce_conversion_profitable_p(seq, &mips_if_info);
So is this hook definition really doing anything useful? It doesn't
really do anything with the insns in the sequence and just seems to hand
back off to the generic variant. Am I missing something? Formatting
nit. if we're keeping this code there's always supposed to be a space
between the function name and the open paren for its arguments.
Basically it seems sensible from an overall structure standpoint. I
can't really comment on the cost model as I haven't done meaningful work
on mips in probably 20 years at this point. THe biggest issue I see is
the conversion_profitable_hook seems pointless and should probably just
be removed unless there's a followup which is going to improve it further.
jeff
