On Tue, Nov 17, 2026 at 08:21:26AM -0800, Doug Anderson wrote: > Hi David, > > On Mon, Nov 16, 2026 at 8:07 PM David Gibson > wrote: > > > > > > On one stress test (running the fdtfactor tool across 438 arm64 Qualcomm > > > > DTBs) this reduced `fdt_property()` time by ~5.6x. > > > > > > What we really want is memmem(), which would presumably have a similar > > > optimization (and likely more) built in. However, we avoid that to > > > make this easier to environments which lack it. > > > > > > Complexity isn't too great, so this looks reasonable to me > > > > Hmm... actually, thinking about this again, I'm kind of baffled that > > the speedup is so large. In the old version, if p is at a character > > that doesn't match the start of the string we're looking for, memcmp() > > should return almost immediately (after a single byte comparison), so > > I'd expect both versions to be - roughly speaking - a single scan > > through each byte of the strings block. At least with gcc builtin and > > inlined mem*() implementations, I'd expect the performance difference > > to be pretty modest between that and the new version. > > > > What was the environment (host CPU, libc, compiler options) you used > > for the benchmark? > > Environment details: > * CPU: AMD Ryzen Threadripper PRO 5995WX (x86_64, AVX2) > * libc: Debian glibc 2.43 > * Compiler: gcc (Debian 15.2.0-14) 15.2.0 > * Flags: default `meson setup build` (`buildtype=debug`, so `-O0 -g > -fPIC`), though testing at `-Os`, `-O2`, and `-O3` gives the same > ~16-17x speedup on the string-table scan itself. > > There are two things going on even at `-O2` / `-O3`: > > 1. **GCC can't inline `memcmp(p, s, slen)` when `slen` isn't a > compile-time constant**: Because `slen` is a runtime variable (and > GCC also can't synthesize a `*p == *s` pre-check on its own since > `slen` could be `0`), GCC at `-O2` / `-O3` emits a full PLT call to > `memcmp()` on every single byte `p` of `strtab`: > > .L10: > movq %rbx, %rdx > movq %rbp, %rsi > movq %r15, %rdi > call memcmp@PLT > ... > addq $1, %r15 > cmpq %r15, %r12 > jnb .L10 > > Even though glibc's `__memcmp_avx2_movbe` returns quickly on a > first-byte mismatch, paying a PLT call/return and function prologue > on every byte of a ~14 KB `/__symbols__` string table (~5.8M > `memcmp()` calls per 800-symbol DTB) dominates runtime. Just adding a > scalar `*p == *s` guard ahead of `memcmp()` already speeds up the > test below by ~8x at `-O2`. > > 2. **SIMD scanning inside `memchr()` (~2x faster than a scalar byte loop > at `-O2`)**: Instead of stepping one byte per scalar loop iteration, > glibc's `memchr()` (`__memchr_avx2` on x86_64) broadcasts `s[0]` > across a 32-byte YMM register (`vpbroadcastb` + `vpcmpeqb`) and scans > 32 bytes per iteration inside a single call, only returning on the > ~4-6% of offsets where `p[0] == s[0]`. Ah, right, thanks for the analysis. > (In the commit message, overall `fdt_property()` time dropped by ~5.6x > rather than ~16x because ordinary repeated property names like > `compatible`, `reg`, and `phandle` hit near the beginning of `strtab` > right away, and `fdt_property()` also includes copying the property > value and header bookkeeping.) > > Here is a small standalone C program (`bench_find_string.c`) that models > building an 800-symbol string table 200 times (~4.2s -> ~0.25s at `-O2` > on my machine): -- David Gibson (he or they) | I'll have my music baroque, and my code david AT gibson.dropbear.id.au | minimalist, thank you, not the other way | around. http://www.ozlabs.org/~dgibson