https://gcc.gnu.org/bugzilla/show_bug.cgi?id=127035
--- Comment #1 from kugan at gcc dot gnu.org --- I can reproduce this in AArch64 with PGO. However, The profile-scaling change correctly redistributes GetVirtualPixelsFromNexus’s original count among its two specialized clones and the remaining function. This leaves the profitable SetPixelCacheNexusPixels specialization with a saved-time count below the whole-program hot-count threshold. Although its IPA-CP profitability score is far above the required threshold, the hard hotness check rejects it. Consequently, GCC no longer creates and inlines the aggregate-specialized SetPixelCacheNexusPixels.constprop versions in the critical pixel-processing path. IMO, We need to relax the cost model. Looking into this.
