https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114729
Vineet Gupta <vineetg at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|NEW |ASSIGNED
--- Comment #17 from Vineet Gupta <vineetg at gcc dot gnu.org> ---
There's one more place where ECC pressure reduction is being ignored to favor
more spills.
With the following non production change (ref to fix4)
@ -2401,12 +2401,18 @@ model_excess_group_cost (struct model_pressure_group
*group,
int pressure, cl;
cl = ira_pressure_classes[pci];
- if (delta < 0 && point >= group->limits[pci].point)
+
+ if (delta < 0)
+ {
+ if (point >= group->limits[pci].point)
{
pressure = MAX (group->limits[pci].orig_pressure,
curr_reg_pressure[cl] + delta);
return -model_spill_cost (cl, pressure, curr_reg_pressure[cl]);
}
+ else
+ return delta;
+ }
We see further spill reduction:
$ -std=c++03 -Ofast -g -S -static -fno-lto -ftree-vectorize ML_BSSN_RHS.cpp
-march=rv64gcv_zvl256b_zba_zbb_zbs_zicond_zfa -mrvv-vector-bits=zvl
--verbose-asm
$ grep sfp ML_BSSN_RHS.s | wc -l
--- tc-up-240920-sched1-baseline : 48,523
--- tc-up-240920-sched1-fix1-ECC : 35,403
--- tc-up-240920-sched1-fix1-fix4-ECC-fixes : 28,991
And following are the qemu dynamic icounts for cactu for corresponding builds
2,544,112,250,412 # baseline
2,185,048,432,166 # fix1 ECC
2,048,796,612,056 # fix1 + fix4 ECC