protobuf v34: descriptor tables marked __attribute__((section("protodesc_cold")))

[ltowarek]

Fix lives in: protobuf

What the workaround does

We empty-redefine the Abseil section attribute macro on every protobuf-related target so the custom section attribute disappears and descriptor tables fall back to default placement (.rodata):

# cmake/protobuf_setup.cmake:160-169
foreach(target ${vendored_targets})
    target_compile_definitions(${target} PRIVATE
        ABSL_ATTRIBUTE_SECTION_VARIABLE\(name\)=)
endforeach()

Root cause

protobuf v34's generated .pb.cc files mark descriptor tables with:

PROTOBUF_SECTION_VARIABLE(protodesc_cold)
// expands to: __attribute__((section("protodesc_cold")))

ESP-IDF's sections.ld linker script does not have a mapping for the protodesc_cold section. The unknown section breaks the .flash.appdesc / .flash.rodata contiguous-layout assertion at link time. Bare-metal / firmware platforms generally control their linker scripts tightly and don't expect libraries to invent new output sections.

Glibc / Linux tolerates this because the linker creates a default output section for any unknown input section; bare-metal linker scripts that use KEEP(*(.foo)) enumeration explicitly do not.

Proposed fix

Two options in protobuf:

1. Default to .rodata for descriptor tables. Gate the protodesc_cold attribute behind an explicit opt-in CMake option (PROTOBUF_USE_COLD_SECTION=ON). This is the cleanest fix — the section is a size-optimization for cold descriptors, not a correctness requirement.

2. Guard the attribute with __has_attribute(section) plus a portability check for bare-metal targets — but this would still need a way for the consumer to disable it, since the attribute compiles fine, the linker is the one that complains.

Approach 1 is preferable because it makes the default behavior portable and the optimization explicit.

What we delete once fixed

• The ABSL_ATTRIBUTE_SECTION_VARIABLE(name)= empty redefine in cmake/protobuf_setup.cmake:160-169

[ltowarek]

Update from PR #52 (ESP-IDF v6 upgrade):

ESP-IDF v6 enables --orphan-handling=error in the linker, which turns any unrecognised input section into a hard linker error rather than a warning. Abseil's LowLevelAlloc annotates functions with ABSL_ATTRIBUTE_SECTION (the function-level variant, distinct from the variable-level ABSL_ATTRIBUTE_SECTION_VARIABLE already suppressed). Without intervention, these produce named sections (.text.hot.NameHere) that the ESP-IDF linker script doesn't account for, causing the build to fail with --orphan-handling=error.

The existing suppression in _esp_opentelemetry_apply_int_override was extended to cover this macro as well:

"SHELL:-D ABSL_ATTRIBUTE_SECTION_VARIABLE(name)="   # descriptor tables → .rodata
"SHELL:-D ABSL_ATTRIBUTE_SECTION(name)="            # functions → .text (added for v6)

Both macros are now empty-defined for all vendored targets. The root cause and proposed upstream fix remain as described above.