arnavsharma990 opened a new pull request, #21214:
URL: https://github.com/apache/nuttx/pull/21214
<html><head></head><body><h2><span><span>Summary</span></span></h2><p><span><span>The
constant 64-bit division optimization in
</span></span><code><span><span>include/nuttx/lib/math32.h</span></span></code><span><span>
is effectively disabled because
</span></span><code><span><span>div64_const</span></span></code><span><span> is
guarded by
</span></span><code><span><span>CONFIG_HAVE_EXPRESSION_STATEMENT</span></span></code><span><span>,
a symbol that is never defined in NuttX or
nuttx-apps.</span></span></p><p><span><span>The guard has remained ineffective
since the optimization was introduced in commit
</span></span><code><span><span>7fc03e0084</span></span></code><span><span>.
Commit </span></span><code><span><span>c32b683</span></span></code><span><span>
later removed the
</span></span><code><span><span>CONFIG_HAVE_LONG_LONG</span></span></code><span><span>
condition, leaving the guard dependent entirely on the undefined
symbol.</span></span></p><p><span><span>As a result, affe
cted 64-bit division call sites fall back to ordinary
</span></span><code><span><span>/</span></span></code><span><span> operations.
On 32-bit targets, this can introduce expensive software division helpers
instead of using the intended Granlund–Montgomery reciprocal
multiplication.</span></span></p><p><span><span>This PR replaces the dead guard
with a compiler capability check:</span></span></p><pre><code><span><span>#if
!defined(__ghs__) && (defined(__GNUC__) ||
defined(__clang__))</span></span></code></pre><p><span><span>This follows the
existing compiler conventions in
</span></span><code><span><span>include/nuttx/compiler.h</span></span></code><span><span>
and enables the optimization for GCC and Clang while excluding Green Hills,
which does not support statement expressions in all required
contexts.</span></span></p><p><strong><span><span>Affected call
sites:</span></span></strong></p><ul><li><p><code><span><span>include/nuttx/clock.h</span></span></code><span><span>
— clock conversion functions and
macros.</span></span></p></li><li><p><code><span><span>sched/wdog/wd_start.c:369</span></span></code><span><span>
— watchdog
timing.</span></span></p></li><li><p><code><span><span>include/nuttx/timers/oneshot.h:366</span></span></code><span><span>
— oneshot timer
calculations.</span></span></p></li><li><p><code><span><span>include/nuttx/timers/clkcnt.h:183,224,317,321</span></span></code><span><span>
— clock counter conversions.</span></span></p></li></ul><p><span><span>The
separate
</span></span><code><span><span>lib_ultoa_invert.c</span></span></code><span><span>
optimization is not included in this change. It requires additional
modifications to
</span></span><code><span><span>div64_const32</span></span></code><span><span>
and will be considered separately.</span></span></p><p><span><span>Related
issue:
#21213</span></span></p><h2><span><span>Impact</span></span></h2><ul><li><p><strong><span><span>Binary
size:</span></span></strong><span>
<span> Reduced by 104 bytes in the tested
</span></span><code><span><span>stm32f4discovery</span></span></code><span><span>
minimal build (28,212 → 28,108
bytes).</span></span></p></li><li><p><strong><span><span>Build
system:</span></span></strong><span><span> No Kconfig or build-system
changes.</span></span></p></li><li><p><strong><span><span>Hardware:</span></span></strong><span><span>
No hardware-specific
changes.</span></span></p></li><li><p><strong><span><span>Documentation:</span></span></strong><span><span>
No changes
required.</span></span></p></li><li><p><strong><span><span>Security:</span></span></strong><span><span>
No security
implications.</span></span></p></li><li><p><strong><span><span>Compatibility:</span></span></strong><span><span>
The optimized path preserves the original integer division semantics.
Unsupported compilers retain the existing
fallback.</span></span></p></li><li><p><strong><span><span>Compiler
support:</span></span></strong><span><span> Enabled for
GCC and Clang; Green Hills is explicitly
excluded.</span></span></p></li></ul><p><span><span>The remaining
</span></span><code><span><span>_udivmoddi4</span></span></code><span><span>
helper (816 bytes in the tested image) comes from
</span></span><code><span><span>lib_ultoa_invert</span></span></code><span><span>
and is outside the scope of this
PR.</span></span></p><h2><span><span>Testing</span></span></h2><p><strong><span><span>Environment:</span></span></strong><span><span>
macOS, Apple Clang 21.0.0 and
</span></span><code><span>arm-none-eabi-gcc</span></code><span><span>
16.2.0.</span></span></p><p><strong><span><span>Target:</span></span></strong><span><span>
</span></span><code><span><span>stm32f4discovery</span></span></code><span><span>
(ARM Cortex-M4,
32-bit).</span></span></p><p><strong><span><span>Configuration:</span></span></strong><span><span>
Minimal
</span></span><code><span><span>stm32f4discovery:nsh</span></span></code><span><span>
build with </span></span><code>
<span><span>CONFIG_INIT_ENTRYPOINT="hello_main"</span></span></code><span><span>
and unnecessary features disabled.</span></span></p><h3><span>1. Baseline
build</span></h3><pre><code><span><span>$ arm-none-eabi-size nuttx
</span>text data bss dec hex
<span> 28212 </span>800 2136 31148 79ac<span>
$ grep -o "udivmoddi4\|aeabi_ldivmod\|aeabi_uldivmod" nuttx.map | sort |
uniq -c
1 aeabi_ldivmod
20 aeabi_uldivmod
20 udivmoddi4</span></span></code></pre><p><span><span>Disassembly of
</span></span><code><span><span>clock_systime_timespec</span></span></code><span><span>
confirmed the software division
call:</span></span></p><pre><code><span><span>8001c56: f003 fd65 bl 8005724
<__aeabi_ldivmod></span></span></code></pre><h3><span>2. Modified
build</span></h3><p><span>After replacing the guard with the compiler
capability check:</span></p><pre><code><span><span>$ arm-none-eabi-size nuttx
</span>text data bss dec hex
<span> 28108 </span>800 2136 31044 7944<span>
$ grep -o "udivmoddi4\|aeabi_ldivmod\|aeabi_uldivmod" nuttx.map | sort |
uniq -c
20 aeabi_uldivmod
20 udivmoddi4</span></span></code></pre><p><span>Disassembly showed the
division-helper call replaced by <span>hardware
multiply-accumulate:</span></span></p><pre><code><span><span>8001c58: fbe0 2305
umlal r2, r3, r0, r5</span></span></code></pre><h3><span><span>3. Build
result</span></span></h3><pre><code><span>EXIT=0</span></code></pre><p><span>The
build completed successfully without errors or warnings related to
</span><code><span>math32.h</span></code><span>.</span></p><h3><span>4. Clang
verification</span></h3><p><span>A minimal reproduction <span>of the
statement-expression macro </span>compiled and ran successfully with Apple
Clang 21.0.0:</span></p><pre><code><span>clang -std=gnu11 -Wall -Wextra
test_math32.c -o test_math32
./test_math32</span></code></pre><p><span>Output:</span></p><pre><code><span>12345678</span></code></pre><p><span>A
full Clang cross-target build was not performed<span> because
</span></span><code><span><span>arm-none-eabi-clang</span></span></code><span><span>
was unavailable.</span></span></p><h3><span><span>5. Binary size
comparison</span></span></h3>
| Text | Data | BSS
-- | -- | -- | --
Baseline | 28,212 | 800 | 2,136
Modified | 28,108 | 800 | 2,136
Difference | −104 | 0 | 0
<p><span><span>The
</span></span><code><span><span>_udivmoddi4</span></span></code><span><span>
and
</span></span><code><span><span>__aeabi_uldivmod</span></span></code><span><span>
references remain in the image and are outside the scope of this
change.</span></span></p></body></html>
--
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]
