--target=esp32s3 builds call0; IDF on the s3 is windowed

Found 2026-09-24 (frankS) while adding the s3 leg of the ESP C conformance run.

Measured

• --target=esp32s3 --emit-obj vs --target=xtensa --xtensa-abi=windowed --platform=esp --emit-obj of test/c_variadic_on_windowed_xtensa.c: different objects (440715 vs 373009 bytes of code); the first is call0 (XtensaABI defaults to XTENSA_ABI_CALL0, compiler.pas, and only --xtensa-abi=windowed changes it).
• All xtensa SoCs in the table (SocIsXtensa: esp32, esp32s2, esp32s3) run ESP-IDF windowed. Bare (--esp-profile=bare) is call0 by requirement, and the refusal of bare+windowed stays.

Why it was not fixed with the finding

The decide ticket (decide-esp-soc-axis-and-capability-table) says a SoC target implies the capability row, and the ABI is in that table's axes, so the implication itself looks decided. Two things stand in the way:

1. Makefile test-xtensa: --target=esp32s3 --emit-obj test/lib_asyncnet6.pas ("THE ESP PARITY THIS WAS ACTUALLY FOR") goes red under windowed, because CoSwitch is call0-only and its negative control asserts windowed refuses. Making the chip imply windowed turns that parity row into a refusal. That is arguably the TRUE answer for IDF on the s3, but it changes a row someone wrote deliberately.
2. c_crtl_s3.o in test-emit-obj builds as call0 today for the same reason.

Condition that would retire this

--target=esp32s3 --emit-obj producing an object whose app_main at .text+0 begins with entry (bytes 36 xx 00), and a booted tools/esp_run.sh --chip esp32s3 of a program built with only the chip name.

Resolution (2026-09-24, frankH)

Both halves were done in one change, since neither is correct alone.

• The default: compiler.pas, after DeriveTargetPlatform. The rule is SocExplicit and SocIsXtensa and not XtensaAbiExplicit and not EspBareBoot and TargetPlatform = PLATFORM_ESP => windowed. XtensaAbiExplicit is new, so --xtensa-abi=call0 still means call0.
• The coroutines: the windowed CoSwitch, recorded in feature-a-a-windowed-abi-coswitch-for-xtensa.

The retire condition, both halves measured:

• --target=esp32s3 --emit-obj gives an app_main that opens with entry a1, 64.
• A program built with ONLY --target=esp32s3 booted and matched its x86-64 oracle:
  • under S3 QEMU;
  • on an ESP32-S3 BOARD, via tools/esp_flash.sh, whose s3 case now spells just the chip name (so does tools/esp_run.sh);
  • the program was test_scheduler_yields_deep_in_a_call_chain, 8 of 8 lines.
• esp32 and esp32s2 chip-name objects are byte-identical to spelled-windowed.
• --target=esp32s3 --esp-profile=bare still builds.

What the s3 asyncnet6 row now asserts: the same command line, and it now builds a WINDOWED object. It still passes, now for the right reason.

Not changed: esp_flash.sh's esp32s2 case still spells the generic xtensa. --target=esp32s2 would also move the SoC from the implied s3 to a real s2, and no s2 was available to verify that.

Log

• 2026-09-24 — resolved; this names the commit that carried the resolve, which is not always the one that carried the change — commit eec095baa.