libs/libbuiltin/compiler-rt: skip unsupported Arm VFP builtins asm

[uditjainstjis]

Summary

libs/libbuiltin/compiler-rt globs every arm/*.S source into the builtins
library ($(wildcard .../arm/*.S) in Make.defs, file(GLOB ...) in
CMakeLists.txt). Several of those hand-written VFP assembly files require FPU
features the selected target may not have. Upstream compiler-rt selects them
conditionally in its own CMake logic; NuttX did not, so enabling
BUILTIN_COMPILER_RT on a single-precision-FPU Arm target (e.g. Cortex-M33 with
-mfpu=fpv5-sp-d16, as on the Raspberry Pi Pico 2 / RP2350) breaks the build
with assembler errors such as:

compiler-rt/lib/builtins/arm/adddf3vfp.S:24: Error: selected FPU does not support instruction -- `vadd.f64 d6,d6,d7'
compiler-rt/lib/builtins/arm/chkstk.S:29:   Error: instruction not supported in Thumb16 mode -- `subs r5,r5,#4096'

This filters the Arm .S source list to match the configured FPU, in both the
Makefile and CMake builds:

• chkstk.S / chkstk2.S are Windows/MinGW-only stack probes — always dropped;
• with no hardware FPU (!CONFIG_ARCH_FPU), every arm/*vfp.S is dropped;
• with a single-precision FPU (!CONFIG_ARCH_DPFPU), the double-precision
  *df*vfp.S routines are dropped.

The excluded set mirrors upstream compiler-rt's own conditional selection.

Fixes #17386.

Impact

• Fixes the build for CONFIG_BUILTIN_COMPILER_RT=y on single-precision-FPU and
  soft-float Arm targets (previously failed to link/assemble).
• No behavioural change for targets with a double-precision FPU or for
  non-Arm architectures: those targets keep the same source set (the double-
  precision *df*vfp.S files are only removed when CONFIG_ARCH_DPFPU is unset,
  and the whole block is guarded by CONFIG_ARCH_ARM).
• Build system only; no runtime code, API, or documentation changes.

Testing

Host: Ubuntu 24.04 (x86_64), arm-none-eabi-gcc 14.2.rel1 (matching the
toolchain in the original report), compiler-rt 17.0.1 (NuttX default
LIB_COMPILER_RT_VERSION).

Reproduced the failure by assembling every arm/*.S builtin with the Cortex-M33
single-precision flags from the issue
(-march=armv8-m.main+dsp -mtune=cortex-m33 -mfpu=fpv5-sp-d16 -mfloat-abi=softfp -mthumb -mcmse):

[RESULT] total=86  FAIL=18  PASS=68
FAILING: adddf3vfp.S chkstk.S divdf3vfp.S eqdf2vfp.S extendsfdf2vfp.S
         fixdfsivfp.S fixunsdfsivfp.S floatsidfvfp.S floatunssidfvfp.S
         gedf2vfp.S gtdf2vfp.S ledf2vfp.S ltdf2vfp.S muldf3vfp.S nedf2vfp.S
         subdf3vfp.S truncdfsf2vfp.S unorddf2vfp.S

Applying this patch's exclusion set (17 double-precision *df*vfp.S files +
chkstk.S) and re-assembling the remainder:

[POST-FIX] assembled=68  remaining_failures=0

negdf2vfp.S and the single-precision *sf*vfp.S routines are correctly
retained (they assemble under a single-precision FPU). Verified the Make filter
logic independently: for a single-precision target it removes exactly the DP-VFP
and chkstk files while keeping negdf2vfp.S, addsf3vfp.S, and .c sources.

./tools/checkpatch.sh -f passes on both changed files (including
cmake-format).

[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