casaroli opened a new pull request, #21845: URL: https://github.com/apache/nuttx/pull/21845
## Summary Each make process of a build parses `tools/Config.mk` and the arch `Toolchain.defs` again. Three `$(shell)` commands in these files run again in each make process, with the same result each time. One configure and build of qemu-armv8a:nsh starts 1115 shells, and these three commands start about 280 of them. This PR runs each of them once and gives the result to the sub-makes in the environment, as `tools/Config.mk` already does for `DEFINE_PREFIX` and `INCDIR_PREFIX`. - tools/Config.mk: Look for bash once per build. `command -v bash` ran in each make process (169 times in qemu-armv8a:nsh). `BASHCMD` is now set only when it is not defined, and it is exported. - build: Compare GCCVER with 12 once per build. `GCCVER` is exported, but `expr "$(GCCVER)" \>= 12` ran in each make process (106 times in qemu-armv8a:nsh), in arch/arm, arch/arm64, arch/risc-v (two places) and libs/libxx/libcxx. The result is now exported in `GCCVER_GE_12`, together with the `GCCVER` that it is for (`GCCVER_GE_12_FOR`). A sub-make with a different `GCCVER` (the ameba boards set their own) compares again. The comparison stays in each file, because 3 efm32 boards include their `Toolchain.defs` before `tools/Config.mk`. - tools/Unix.mk: Update .version once per make invocation. `tools/version.sh` starts several git commands. It ran in each parse of the top-level Makefile, also in its sub-makes: 3 times in one `configure.sh`, because `olddefconfig` runs `clean_context` and `apps_preconfig` as sub-makes. Now the first run exports `NUTTX_VERSION_UPDATED`, and the sub-makes do not run it again. ## Impact Make build only. CMake does not run these commands. The values are the same as before, so `.config`, `include/nuttx/config.h` and all build commands do not change (see Testing). The build environment gets 4 new variables: `BASHCMD`, `GCCVER_GE_12`, `GCCVER_GE_12_FOR` and `NUTTX_VERSION_UPDATED`. A `BASHCMD` in the environment of `make` is now used; before, the makefile replaced it. Windows native builds do not use the changed lines of `tools/Config.mk` and `tools/Unix.mk`, and the `GCCVER` comparison runs the same `$(shell expr ...)` as before, so its result does not change. Documentation: no page describes these variables, so no change. ## Testing Same output: for each configuration, `configure.sh`, a real `make context`, then `make -n V=1` (all build commands), on master and with this PR. `.config`, `include/nuttx/config.h` and the full `make -n` output are identical: - GNU make 3.81 (macOS): qemu-armv8a:nsh, qemu-armv8a:knsh (kernel), stm32f4discovery:nsh, stm32f4discovery:kostest (protected), efm32-g8xx-stk:nsh (Toolchain.defs before Config.mk), rv-virt:nsh64, rv-virt:knsh64 (kernel; both risc-v comparisons), sim:nsh, sabre-6quad:libcxx (libcxx). - GNU make 4.4: qemu-armv8a:nsh, rv-virt:nsh64, stm32f4discovery:kostest. The only difference is the line number in the expected `make -n` error message for the link inputs (tools/Unix.mk has 5 more lines). - pke8721daf:nsh (ameba, `GCCVER` 10): `.config` and `make -n` output identical. Its context needs Linux prebuilt tools, so `config.h` was not made on macOS. The `GCCVER` cache alone, with GNU make 3.81 and 4.4: a top-level make with `GCCVER` 13, then sub-makes with 13, 10, an empty value, 9, and a sub-make of the sub-make with 9. Each result is the same as a plain `expr`; `expr` runs 4 times instead of 6. Shells that make starts (configure, `make context`, `make -n`): | configuration | master | this PR | |---|---|---| | qemu-armv8a:nsh | 1115 | 841 (-25%) | | rv-virt:nsh64 | 1410 | 997 (-29%) | | stm32f4discovery:nsh | 860 | 634 (-26%) | Time on macOS (arm64, no other load), `configure.sh` 10 times per tree, alternating, median: | configuration | master wall / CPU | this PR wall / CPU | |---|---|---| | qemu-armv8a:nsh | 3.86 s / 3.92 s | 3.45 s / 3.33 s | | rv-virt:nsh64 | 3.75 s / 3.90 s | 3.44 s / 3.32 s | | stm32f4discovery:nsh | 3.88 s / 4.00 s | 3.49 s / 3.39 s | Full builds of the same 3 configurations (`make -j8`, warm ccache, 3 runs per tree): no difference larger than the noise (-2% to +2%). In the CI container on a Linux host with other load: no difference larger than the noise. `tools/checkpatch.sh -c -u -m -g` passes. Not tested: Windows native builds, a build of every configuration (CI does that). -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
