Propagate eh_frame strip policy to ARM Cortex-M / Teensy / RISC-V GCC targets (follow-up to #243) #244

Follow-up to #243 (ESP32: auto-strip .eh_frame when exception handling and panic backtrace are both off).

#243 lands the policy + decision function + flag injection in the ESP32 compiler path only. The same -fno-asynchronous-unwind-tables -fno-unwind-tables pair is a win on every GCC-based embedded target where C++ exceptions and platform-level backtrace are both off — which is most of fbuild's matrix. This ticket tracks propagating that machinery uniformly.

Why this is its own issue

The decision function eh_frame_policy::decide() from #243 is platform-neutral (its inputs are build_flags, build_unflags, profile, env vars, plus an optional platform-specific summary like SdkConfigSummary). The wiring into each per-platform compiler is platform-specific. Splitting them keeps #243's PR small and lets #243 prove out the savings on one platform before fanning out.

Platforms in scope

Plan

1. Wait for ESP32: auto-strip .eh_frame when exception handling and panic backtrace are both off (~225 KB savings on FastLED Blink) #243 to merge and confirm:

   • The decision function (crates/fbuild-build/src/esp32/eh_frame_policy.rs) is a pure function over plain inputs.
   • Tests cover the matrix (Preserve / Strip per trigger).
   • Integration test asserts ≥ 150 KB savings on ESP32-S3 NEOPIXEL Blink.

2. Hoist the decision function up out of esp32/ into a shared module (likely crates/fbuild-build/src/eh_frame_policy.rs or a small fbuild-build-policy sub-module). The inputs stay generic; the per-platform extras (sdkconfig for ESP32, possibly Teensy usb_type/board config for Teensy) plug in as an optional trait or as a &dyn PlatformPolicyHints.

3. Per platform, wire the policy into the existing common_flags() (or equivalent) method:

   • generic_arm/arm_compiler.rs
   • teensy/teensy_compiler.rs
   • esp8266/* (if applicable)

4. Per platform, add an integration test mirroring the ESP32 one: build a tiny canonical sketch, assert binary-size delta. Use the same object-crate .eh_frame section-size diagnostic from ESP32: auto-strip .eh_frame when exception handling and panic backtrace are both off (~225 KB savings on FastLED Blink) #243 so failure messages stay informative.

5. Document in the same fbuild docs page ESP32: auto-strip .eh_frame when exception handling and panic backtrace are both off (~225 KB savings on FastLED Blink) #243 created (docs/.../size-optimization.md or similar). One table of supported platforms, defaults, and overrides.

Decision policy — same as #243

Preserve eh_frame if ANY of:
  - build_flags contains `-fexceptions` or `-funwind-tables`
    or `-fasynchronous-unwind-tables`
  - debug profile (-Og / -O0 / !NDEBUG)
  - FBUILD_KEEP_EH_FRAME=1
  - (platform-specific) backtrace / fault-decoder enabled in platform config
Strip otherwise (default for release builds with no override).

The platform-specific hint is what differs:

• ESP32: CONFIG_ESP_SYSTEM_PANIC_* / CONFIG_ESP_DEBUG_OCDAWARE (already ESP32: auto-strip .eh_frame when exception handling and panic backtrace are both off (~225 KB savings on FastLED Blink) #243).
• Generic ARM: no equivalent today — STM32/RP2040/NRF52 fault handlers don't walk .eh_frame. Default is Strip-unless-overridden.
• Teensy: Teensy's fault handler doesn't use eh_frame either. Default is Strip-unless-overridden.

Estimated impact

The ESP32 measurement (#243) shows ~225 KB savings on FastLED Blink with full propagation. Per-platform numbers will be smaller because the framework cores are smaller, but the ratio is roughly the same (~30 % of release-build size is .eh_frame for an exception-handling-disabled C++ library).

Rough projections, to be validated against the integration test:

Acceptance

• Shared eh_frame_policy::decide() lives outside the esp32/ module and is consumed by every GCC-based platform compiler.
• Per-platform integration test: build canonical sketch with FBUILD_KEEP_EH_FRAME=1 and FBUILD_STRIP_EH_FRAME=1, assert positive delta.
• No regression in existing per-platform tests (crates/fbuild-build/tests/teensy_build.rs, stm32_acceptance.rs, etc.).
• Docs page from ESP32: auto-strip .eh_frame when exception handling and panic backtrace are both off (~225 KB savings on FastLED Blink) #243 expanded to list every supported platform's default + override.
• Optional bonus: the Teensy LC / 3.0 binary size drops enough to give the RAM-overflow work in Teensy 3.0 / LC link-time RAM overflow (.dmabuffers / .bss) across all examples FastLED#2406 some headroom (not a hard requirement; mostly a happy side effect).

Out of scope

• AVR. Audit first; almost certainly nothing to strip.
• WASM. Different unwind model.
• The decision policy itself. Owned by ESP32: auto-strip .eh_frame when exception handling and panic backtrace are both off (~225 KB savings on FastLED Blink) #243; this ticket is wiring only.
• Per-platform fault-decoder semantics (e.g., if a Teensy CrashReport-style feature wants eh_frame, that goes in a separate decision-policy ticket).

Related

• ESP32: auto-strip .eh_frame when exception handling and panic backtrace are both off (~225 KB savings on FastLED Blink) #243 — ESP32-side implementation that lands the decision function.
• ESP32-S3 binary size: 657 KB on master vs 344 KB on 3.10.3 — remaining +313 KB delta breakdown FastLED#2473 — symbol-level audit that motivated the work.
• Teensy 3.0 / LC link-time RAM overflow (.dmabuffers / .bss) across all examples FastLED#2406 — Teensy 3.0 / LC RAM overflow; any code-size cut on those boards helps incidentally.
• feat(channels): compile-time Bus binding via templated ChannelConfig (eliminates Channel API binary bloat) FastLED#2428 — driver registry default-off (already merged); the other big size win.