arch/arm: stop boards silently discarding --fixed-r10

[casaroli]

Summary

arch/arm/src/common/Toolchain.defs adds --fixed-r10 to CFLAGS when
CONFIG_PIC is enabled. Nearly every board Make.defs includes that file and
then assigns

CFLAGS := $(ARCHCFLAGS) $(ARCHOPTIMIZATION) $(ARCHCPUFLAGS) ...

with :=, which discards it. 346 of the tree's board files contain that exact
CPICFLAGS = $(ARCHPICFLAGS) $(CFLAGS) idiom, and every one of them assigns
CFLAGS this way first.

That flag is what stops the base firmware from allocating r10 — the register a
PIC module reaches its own data through. When it is lost, a base firmware
routine that calls back into module code (qsort() with a module comparison
function is the standard case) arrives with r10 holding whatever the firmware
last put there, and the module reads its data through a bad base pointer.

The failure is silent and remote from its cause: the build succeeds, every
module loads, and only the callback misbehaves.

mps3-an547:picostest is the only defconfig in the tree that sets
CONFIG_PIC=y, so it is the only board changed here. It re-applies the flag
after the := and filters it back out of CPICFLAGS and CELFFLAGS, because
GCC rejects --fixed-r10 alongside the -mpic-register=r10 those carry once
-fpic is in effect:

cc1: error: unable to use 'r10' for PIC register

That filtering is also why the flag cannot simply move to ARCHCPUFLAGS to
survive the := — the module flags derive from CFLAGS, so it would land in
exactly that rejected combination. Toolchain.defs now records this, since the
next board to enable CONFIG_PIC will otherwise hit the same trap.

Why the second commit touches the spellcheck config

tools: exempt the "*.siz" gsize comment from codespell is unrelated to the
register, and is here only because CI will not go green without it. An existing
comment in arch/arm/src/common/Toolchain.defs reads:

# instruct the gsize to generate a "*.siz" file that contains the detailed section size

codespell reports siz ==> size, six. It is neither — .siz is the filename
extension gsize produces. The line is on master and this PR does not touch it;
it adds no line containing siz at all.

It fails anyway because tools/checkpatch.sh:224 invokes

codespell $codespell_args ${@: -1}

passing the filename, so codespell scans the whole file rather than the
changed hunks. checkpatch builds its file list from the commit range, so
editing Toolchain.defs for any reason pulls all ~600 of its lines into the
check. Confirmed by the other two PRs in this series: they do not touch that
file and report ✔️ All checks pass.

I put the line in .codespell-ignore-lines, which .codespellrc documents as
the place for exotic-but-correct occurrences, rather than in
ignore-words-list, which that file reserves for frequently used words. It is a
separate commit so it can be dropped, taken separately, or replaced with a
different remedy without disturbing the fix itself.

Impact

• Users: affects CONFIG_PIC configurations only. mps3-an547:picostest
  now genuinely reserves r10 in the base firmware.
• Build process: no change for the other 345 boards — they get a comment in
  Toolchain.defs and nothing else. No new warnings.
• Hardware: none directly; changes generated code for CONFIG_PIC firmware
  only (one fewer allocatable register, which is the point).
• Documentation / security / compatibility: unaffected.
• Not fixed here: any board that enables CONFIG_PIC in future still needs
  the same three lines. A tree-wide mechanical change to all 346 files would
  fix it once and for all; I did not attempt that, and would rather a
  maintainer say whether it is wanted before 346 files are touched.
• Spellcheck config: the second commit adds one line to
  .codespell-ignore-lines (see above). It changes no build behaviour and
  unblocks CI for every future PR touching Toolchain.defs, not just this one.

Testing

Board: mps3-an547:picostest — the only defconfig with CONFIG_PIC=y.

Because the failure is a lost compiler flag, the test is what the build system
computes for CFLAGS. Reproduce with:

./tools/configure.sh mps3-an547:picostest
cat > /tmp/show.mk <<'EOF'
include $(TOPDIR)/Make.defs
show:
	@echo "CONFIG_PIC = $(CONFIG_PIC)"
	@echo "CFLAGS    has --fixed-r10 : $(if $(filter --fixed-r10,$(CFLAGS)),YES,NO)"
	@echo "CPICFLAGS has --fixed-r10 : $(if $(filter --fixed-r10,$(CPICFLAGS)),YES,NO)"
	@echo "CELFFLAGS has --fixed-r10 : $(if $(filter --fixed-r10,$(CELFFLAGS)),YES,NO)"
EOF
make -f /tmp/show.mk TOPDIR=$PWD show

Before (master 7df7c6ee) — the flag is gone:

CONFIG_PIC = y
CFLAGS    has --fixed-r10 : NO
CPICFLAGS has --fixed-r10 : NO
CELFFLAGS has --fixed-r10 : NO

After this patch — firmware reserves r10, module flags do not:

CONFIG_PIC = y
CFLAGS    has --fixed-r10 : YES
CPICFLAGS has --fixed-r10 : NO
CELFFLAGS has --fixed-r10 : NO

The conflict that makes the filtering necessary, shown directly:

$ arm-none-eabi-gcc -mcpu=cortex-m55 -mthumb -fpic -msingle-pic-base \
      --fixed-r10 -mpic-register=r10 -c t.c -o t.o
cc1: error: unable to use 'r10' for PIC register

$ arm-none-eabi-gcc -mcpu=cortex-m55 -mthumb -fpic -msingle-pic-base \
      -mpic-register=r10 -c t.c -o t.o          # accepted

Limitation, stated plainly: I could not produce a full link of
mps3-an547:picostest, and it is not for want of a toolchain — I repeated this
with the Arm GNU Toolchain 15.2.Rel1 (full newlib). The Makefile build of this
board does not complete on unmodified master either, for reasons unrelated
to this patch:

$ ./tools/configure.sh mps3-an547:picostest && make
arm-none-eabi-gcc: error: missing argument to '-Wstack-usage='

CONFIG_STACK_USAGE_WARNING is unset for this config, and Toolchain.defs
only omits -Wstack-usage= when the value is exactly 0, so an empty value is
passed through. Setting it to 0 gets further, and then:

make[1]: *** No rule to make target `arm_exception.S', needed by `.depend'.  Stop.
make: *** [pass2dep] Error 2

Both failures reproduce identically on master with and without this patch, so
they are pre-existing and out of scope here — but they do mean the one board in
the tree that enables CONFIG_PIC currently has no working Makefile build,
which is probably why this flag went missing unnoticed. The flag-level evidence
above is therefore the strongest available; a maintainer with a working
CONFIG_PIC build is very welcome to add a link log.

The same mechanism is exercised end-to-end in a downstream tree of mine on
RP2350 hardware, where the equivalent register (r9, for FDPIC modules) being
lost to this same := caused a firmware qsort() callback into module code to
read garbage, and re-applying the flag fixed it.

Tree checks, both commits applied:

$ ./tools/checkpatch.sh -c -u -m -g master..HEAD
Used config files:
    1: .codespellrc
✔️ All checks pass.

Without the second commit this reports
arch/arm/src/common/Toolchain.defs:485: siz ==> size, six.

[github-actions]

MemBrowse Memory Report

No memory changes detected for:

• arduino-mega2560
• esp32-devkitc
• hifive1-revb
• mirtoo
• qemu-armv8a
• qemu-intel64
• rx65n-rsk2mb
• s698pm-dkit
• stm32-nucleo-f103rb

[xiaoxiang781216]

let's move to ARCHCFLAGS instead. it's better to fix the root cause directly.

[casaroli]

I agree that your suggestion is the correct fix, however we need to change all 500 arm files, so that ARCHPICFLAGS and CPICFLAGS all move to Toolchain.defs

This is a huge change, so I have prepared the fix in a separate PR so you can analyze and compare, and we can check the CI status:

#19510