On 06.02.2026 17:15, Alejandro Vallejo wrote: > Hi, > > This series is a big revamp of vendor-checking to enable it to perform DCE. > It improves all configurations useful in practice at minimal cost in the full > build, and at a massive advantage for the single-vendor case. Many ifdefs can > go away as a side effect of the series. > > This series depends on cross-vendor removal: > > https://lore.kernel.org/xen-devel/[email protected]/T/#m4c3d318f37e4f24d0f8c62b104221aa5d428cebc > > Patch 1 in this series matches that of cross-vendor removal. It's logically > required, but that's the single requirement. > > High level description > ====================== > > When compared to the RFC this makes a different approach The series introduces > cpu_vendor() which maps to a constant in the single vendor case and to > (boot_cpu_data.vendor & X86_ENABLED_VENDORS), where X86_ENABLED_VENDORS is a > mask of the compile-time chosen vendors. This enables the compiler to detect > dead-code at the uses and remove all unreachable branches, including in > switches. > > When compared to the x86_vendor_is() macro introduced in the RFC, this is > simpler. It achieves MOST of what the older macro did without touching the > switches, with a few caveats: > > 1. Compiled-out vendors cause a panic, they don't fallback onto the unknown > vendor case. In retrospect, this is a much saner thing to do.
I'm unconvinced here. Especially our Centaur and Shanghai support is at best rudimentary. Treating those worse when configured-out than when configured-in feels questionable. > 2. equalities and inequalities have been replaced by equivalent > (cpu_vendor() & ...) > forms. This isn't stylistic preference. This form allows the compiler > to merge the compared-against constant with X86_ENABLED_VENDORS, yielding > much better codegen throughout the tree. > > The effect of (2) triples the delta in the full build below. > > Some differences might be attributable to the change from policy vendor checks > to boot_cpu_data. In the case of the emulator it caused a 400 bytes increase > due to the way it checks using LOTS of macro invocations, so I left that one > piece using the policy vendor except for the single vendor case. For the emulator I'd like to point out this question that I raised in the AVX10 series: "Since it'll be reducing code size, we may want to further convert host_and_vcpu_must_have() to just vcpu_must_have() where appropriate (should be [almost?] everywhere)." Sadly there was no feedback an that (or really on almost all of that work) at all so far. Jan
