The estimate on target_clobbered_registers based on the call_used arrays is not
correct. This is the worst case
heuristics on the estimate on target_clobbered_registers. This disables many of
the loop Invariant code motion
opportunities in presence of call. Instead of considering the spill cost we
consider only the target_reg_cost
aggressively.
With this change with old logic used in regs_used gave the following gains.
diff --git a/gcc/cfgloopanal.c b/gcc/cfgloopanal.c
--- a/gcc/cfgloopanal.c
+++ b/gcc/cfgloopanal.c
@@ -373,15 +373,23 @@ estimate_reg_pressure_cost (unsigned n_new, unsigned
n_old, bool speed,
/* If there is a call in the loop body, the call-clobbered registers
are not available for loop invariants. */
+
if (call_p)
available_regs = available_regs - target_clobbered_regs;
-
+
/* If we have enough registers, we should use them and not restrict
the transformations unnecessarily. */
if (regs_needed + target_res_regs <= available_regs)
return 0;
- if (regs_needed <= available_regs)
+ /* Estimation of target_clobbered register is based on the call_used
+ arrays which is not the right estimate for the clobbered register
+ used in called function. Instead of considering the spill cost we
+ consider only the reg_cost aggressively. */
+
+ if ((regs_needed <= available_regs)
+ || (call_p && (regs_needed <=
+ (available_regs + target_clobbered_regs))))
/* If we are close to running out of registers, try to preserve
them. */
cost = target_reg_cost [speed] * n_new;
SPEC CPU 2000 INT benchmarks
(Geomean Score without the above change vs Geomean score with the above change
= 3745.193 vs vs 3752.717)
SPEC INT CPU 2000 benchmarks.
(Geomean Score without the above change vs Geomean score with the above change
= 4741.825 vs 4792.085)
Thanks & Regards
Ajit
