Temporary paired trials showed:
LTO alone: about 2.3% faster on both phosphatidylcholine and clarithromycin.
Corpus-trained PGO+LTO: about 7.7% on phosphatidylcholine, 5.1% on clarithromycin, and 15% on scaling-suite clock totals.
PGO regressed one small path case, so validate every suite and tune training weights before making it the default. Keep x86-64-v3 for distributed binaries; -march=native added only about 0.7% on a representative long case.
Temporary paired trials showed:
LTO alone: about 2.3% faster on both phosphatidylcholine and clarithromycin.
Corpus-trained PGO+LTO: about 7.7% on phosphatidylcholine, 5.1% on clarithromycin, and 15% on scaling-suite clock totals.
PGO regressed one small path case, so validate every suite and tune training weights before making it the default. Keep x86-64-v3 for distributed binaries; -march=native added only about 0.7% on a representative long case.