Another czero related adjustment. This time in costing of conditional
move sequences. Essentially a copy from a promoted subreg can and
should be ignored from a costing standpoint. We had some code to do
this, but its conditions were too strict.
No real surprises evaluating spec. This should be a minor, but probably
not measurable improvement in x264 and xz. It is if-converting more in
some particular harm to hot routines, but not necessarily in the hot
parts of those routines.
It's been tested on riscv32-elf and riscv64-elf. Versions of this have
bootstrapped and regression tested as well, though perhaps not this
exact version.
Waiting on pre-commit testing.
Jeff
gcc/
* config/riscv/riscv_noce_conversion_profitable_p): Relax condition
for adjustments due to copies from promoted SUBREGs.
diff --git a/gcc/config/riscv/riscv.cc b/gcc/config/riscv/riscv.cc
index aa8cd97b310..ec858625d24 100644
--- a/gcc/config/riscv/riscv.cc
+++ b/gcc/config/riscv/riscv.cc
@@ -4603,16 +4603,14 @@ riscv_noce_conversion_profitable_p (rtx_insn *seq,
rtx dest = SET_DEST (x);
- /* Do something similar for the moves that are likely to
+ /* Do something similar for the moves that are likely to
turn into NOP moves by the time the register allocator is
- done. These are also side effects of how our sCC expanders
- work. We'll want to check and update LAST_DEST here too. */
- if (last_dest
- && REG_P (dest)
+ done. We don't require src to be something set in this
+ sequence, just a promoted SUBREG. */
+ if (REG_P (dest)
&& GET_MODE (dest) == SImode
&& SUBREG_P (src)
- && SUBREG_PROMOTED_VAR_P (src)
- && REGNO (SUBREG_REG (src)) == REGNO (last_dest))
+ && SUBREG_PROMOTED_VAR_P (src))
{
riscv_if_info.original_cost += COSTS_N_INSNS (1);
riscv_if_info.max_seq_cost += COSTS_N_INSNS (1);
