samparker added a comment.
I just ran some numbers for when -Oz == HighCost and it's interesting... For Arm it's not good:
master minsize-high-cost-expand diff
count 310.000000 310.000000 310.000000
mean 19710.077419 19662.567742 0.000666
std 48882.246111 48773.870505 0.010779
min 292.000000 292.000000 -0.042781
25% 1090.000000 1084.000000 -0.002887
50% 2776.000000 2784.000000 0.000000
75% 10185.000000 10138.000000 0.002424
max 324060.000000 323508.000000 0.058974
Geomean difference 0.1%
For AArch64 it's okay, but there are still plenty of large regressions:
master minsize-high-cost-expand diff
count 310.000000 310.000000 310.000000
mean 26712.283871 26650.709677 -0.000930
std 66147.455977 66033.469245 0.010996
min 476.000000 476.000000 -0.050374
25% 1294.000000 1304.000000 -0.003676
50% 3168.000000 3168.000000 0.000000
75% 11252.000000 11216.000000 0.001597
max 471648.000000 470888.000000 0.051485
Geomean difference -0.1%
But for X86, it's great:
master minsize-high-cost-expand diff
count 313.000000 313.000000 313.000000
mean 27472.361022 27354.431310 -0.005084
std 68350.278069 68101.639463 0.013028
min 389.000000 389.000000 -0.065654
25% 1157.000000 1157.000000 -0.007393
50% 2933.000000 2949.000000 0.000000
75% 11093.000000 11077.000000 0.000000
max 444676.000000 442660.000000 0.039900
Geomean difference -0.5%
CHANGES SINCE LAST ACTION
https://reviews.llvm.org/D76434/new/
https://reviews.llvm.org/D76434