aboutsummaryrefslogtreecommitdiffstats
path: root/docs/design/validation
diff options
context:
space:
mode:
authoreunhhu <[email protected]>2026-09-27 13:43:49 +0100
committereunhhu <[email protected]>2026-09-27 13:43:49 +0100
commit628e72621693af9be8bf94ef42f3164c45763a7b (patch)
tree838534da7561476d625b31afc2341322d85d33a4 /docs/design/validation
parent37d3a1e152707d2da587440c4cea8601b9ef0fd2 (diff)
downloadpure-628e72621693af9be8bf94ef42f3164c45763a7b.tar.gz
pure-628e72621693af9be8bf94ef42f3164c45763a7b.zip
Show the new layout costs nothing, and keep code where it was
The benchmark runs an earlier pure-serve build in turn with the current one (--baseline). Against the build before the move, linking through libpure.a costs nothing: a one-command build with the library's object order has the same size and the same CPU per request. What moved was code placement, so the runtime links last again, as it did; the two orders now differ by less than rounds of one build do (total -0.4%, page +1.2%, mixed -0.1%). The binary is 5,008 B smaller: unused flow.c functions stay in the library, and the linker's A53 erratum padding is gone.
Diffstat (limited to 'docs/design/validation')
0 files changed, 0 insertions, 0 deletions