https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116743
--- Comment #6 from Rama Malladi <rvmallad at amazon dot com> --- I am trying to create a reproducer for this issue. Interim, I wanted to share some stats I got from the MySQL build to highlight this issue w GCC 12.3.0 vs. 11.5.0. Executable Size (B) Baseline AutoFDO GCC 11.5.0 1,577,584,040 1,622,046,224 GCC 12.3.0 1,525,753,912 1,602,504,784 So, I got some inline stats from the compiler opt-report. GCC 11.5.0 Baseline PGO missed_not_inlinable - Total 1,716,533 1,745,229 max-inline-insns-auto limit reached 61,236 63,220 function body not available 1,052,113 1,061,064 can be overwritten at link time 94,640 97,968 function not inlinable 6,808 6,799 large-stack-frame-growth limit reached 18,789 26,347 inline-unit-growth limit reached 327,025 338,244 call is unlikely and code size grow 100,979 103,536 GCC 12.3.0 Baseline PGO missed_not_inlinable - Total 1,838,987 1,728,352 max-inline-insns-auto limit reached 59,480 61,981 function body not available 927,989 944,740 can be overwritten at link time 92,993 93,044 function not inlinable 7,042 7,047 large-stack-frame-growth limit reached 18,629 18,483 inline-unit-growth limit reached 497,948 367,222 call is unlikely and code size grow 180,154 185,387 I tried to reason why "call is unlikely and code size would grow". Here is an example function which got inlined in GCC 11.5.0 w/o and w PGO. But in GCC 12.3.0, it got inlined w/o PGO but not w PGO. GCC 11.5.0: w/o PGO: optimized: Inlined void mem_heap_free(mem_heap_t*)/8218 into bool btr_pcur_t::restore_position(ulint, mtr_t*, ut::Location)/16707 which now has time 186.710665 and size 255, net change of -5. w PGO: optimized: Inlined void mem_heap_free(mem_heap_t*)/8218 into bool btr_pcur_t::restore_position(ulint, mtr_t*, ut::Location)/16707 which now has time 501561.786133 and size 227, net change of -6. GCC 12.3.0: w/o PGO: optimized: Inlined void mem_heap_free(mem_heap_t*)/8120 into bool btr_pcur_t::restore_position(ulint, mtr_t*, ut::Location)/16600 which now has time 186.710665 and size 255, net change of -5. w PGO: missed: not inlinable: bool btr_pcur_t::restore_position(ulint, mtr_t*, ut::Location)/16600 -> void mem_heap_free(mem_heap_t*)/8120, call is unlikely and code size would grow The inline optimizations get tossed by the AutoFDO count propagation algorithm in that the inlining of a callee (w one edge) gets impacted by the perf stats of the caller.
