[rascani]

On toolchains where int32_t and uint32_t are long (the Xtensa and RISC-V
ESP-IDF toolchains, and arm-none-eabi-gcc), log and check messages that pass
them to %d or %u trip -Wformat. ESP-IDF builds with -Werror, so its component
currently compiles the vendored ExecuTorch sources with a blanket -Wno-error;
these fixes let the next change narrow that.

The runtime, flat_tensor and portable kernel sources now use PRId32/PRIu32, or
cast to int where the rest of a message already does. ET_PRI_SIZES_AND_STRIDES,
the format for elements of sizes() and strides(), was "d" in lean mode, which is
wrong for int32_t on these toolchains, so its users such as extension/tensor's
tensor_ptr.cpp failed -Werror=format there. It is now PRId32, which is "d"
wherever int32_t is int. The size and stride messages in op_fft_r2c and in the
ET_CHECK_CONTIGUOUS and ET_CHECK_SAME_STRIDES macros of tensor_util.h use it,
so the STRIDES macros no longer truncate ATen-mode strides to int32_t, and those
macros loop over and print dimensions as size_t, which kernels outside this
tree also get.

Other diagnostics that ESP-IDF turns into errors are fixed in the same sources:
an always-true index >= 0 on a uint32_t in flat_tensor, a cast to a
const-qualified type in native_dropout, prod's *= on bool outputs
(-Wint-in-bool-context), missing parentheses in ET_CHECK_SAME_SHAPE3, and
always_inline on non-inline templates in the Cadence generic im2row operators,
which now use ET_INLINE. prod's bool product, in both prod_out and prod_int_out,
is now && through a template helper, because if constexpr only discards a
branch inside a template; results are unchanged for every dtype.

Verified by compiling every vendored runtime and kernel source (281) as the
generated component builds them for esp32s3 and esp32c6, under ESP-IDF 6.1's
-Wall -Werror, at both -Os and -Og; by compiling extension/tensor/tensor_ptr.cpp
and the tensor_util.h check macros with both ESP-IDF toolchains under
-Werror=format and on host in ATen mode; and by running the component's
emulator tests. On host, portable_kernels_test and the exec_aten util tests
pass.

Authored with Claude Code.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/24702

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

✅ No Failures

As of commit 95cef92 with merge base 4411b8d ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[github-actions]

This PR needs a release notes: label

If your change should be included in the release notes (i.e. would users of this library care about this change?), please use a label starting with release notes:. This helps us keep track and include your important work in the next release notes.

To add a label, you can comment to pytorchbot, for example
@pytorchbot label "release notes: none"

For more information, see
https://github.com/pytorch/pytorch/wiki/PyTorch-AutoLabel-Bot#why-categorize-for-release-notes-and-how-does-it-work.

[JakeStevens]

Thanks for cleaning these up! A few things I'd like changed before this lands.

Use ET_PRI_SIZES_AND_STRIDES for sizes/strides instead of per-site casts

There's already a format specifier for SizesType/StridesType elements. It's documented as "must be used instead of using d directly to support ATen mode", but in lean mode it's still "d" (runtime/core/portable_type/tensor_impl.h:337). That makes it wrong on exactly the toolchains this PR targets. Instead of fixing it, the PR casts at each call site, two different ways in the same header:

• ET_CHECK_CONTIGUOUS and op_fft_r2c.cpp: static_cast<int64_t> + PRId64
• ET_CHECK_SAME_STRIDES2/3: (int32_t) + PRId32. In ATen mode this truncates int64_t strides, while the macro just above prints the same type at full width.

The macro's existing users are still broken. At this PR's head, extension/tensor/tensor_ptr.cpp:106-108 fails -Werror=format with 3 errors under ESP-IDF GCC 15.2 (xtensa-esp32s3-elf-g++ and riscv32-esp-elf-g++). extension/wasm/wasm_bindings.cpp:171-173 uses the same pattern. extension/tensor isn't vendored into the IDF component, but your out-of-tree argument for tensor_util.h applies to it too.

Suggested fix:

// runtime/core/portable_type/tensor_impl.h
#include <cinttypes>
...
#define ET_PRI_SIZES_AND_STRIDES PRId32

Then use "%" ET_PRI_SIZES_AND_STRIDES with no casts at the four sites. strides[i] * sizes[i] has the element type in both modes, so it fits too. Where int32_t is int, PRId32 is "d", so host builds don't change. With this change, tensor_ptr.cpp and the rewritten macro messages compile cleanly under -Werror=format on both IDF toolchains and on host clang.

Smaller things

• tensor_util.h:286-290, 319-325: the STRIDES2/3 loop index now prints with ET_PRIsize_t, but i comes from c10::irange(a__.dim()), so it's ssize_t in lean mode and int64_t in ATen mode. The first check in each macro (:276, :306) also still passes dim() to %zu. In ATen mode wherever int64_t is long long (e.g. macOS), plain -Wformat flags both. To fix, loop with size_t as ET_CHECK_CONTIGUOUS does, and print dim() as a size_t as ET_CHECK_SAME_SHAPE3 does.

• The %zu → ET_PRIsize_t edits don't affect -Werror, since %zu with a size_t is already clean on all the GCC targets named. They do leave the touched functions mixed:

  • flat_tensor_data_map.cpp:359 prints flat_tensor_data->size() with %zu, 14 lines above :373, which prints the same value with ET_PRIsize_t.
  • tensor_parser_exec_aten.cpp:115 sits between the two edited checks.
  • ET_CHECK_SAME_SHAPE3's message (tensor_util.h:108) still uses %zu.

  I'd either convert the touched functions fully or leave the Cadence %zu migration to its own change.

• op_prod.cpp:105: prod_int_out still computes the bool product as acc * outv, and only prod_out uses multiply(). It avoids the warning only because the conversion happens inside std::tuple's constructor. Using multiply(acc, outv) there would keep one spelling in the file.

• op_im2row.cpp:13-17 / op_transposed_im2row.cpp:16-20: the patched local ALWAYS_INLINE is now the same as ET_INLINE (runtime/platform/compiler.h:91-94, which also covers MSVC). Nothing defines DISABLE_ALWAYS_INLINE, so both blocks could be replaced with ET_INLINE.

[rascani]

Thanks Jake — all addressed in e8c8420:

• ET_PRI_SIZES_AND_STRIDES is now PRId32 in lean mode (with <cinttypes>), and ET_CHECK_CONTIGUOUS, ET_CHECK_SAME_STRIDES2/3 and op_fft_r2c use it with no casts. tensor_ptr.cpp now compiles cleanly under -Werror=format with both IDF toolchains. It reproduced your 3 errors on each toolchain before the change.
• The STRIDES macros loop over size_t and print dim() as size_t, like ET_CHECK_SAME_SHAPE3. A host ATen-mode compile with -Wformat-signedness flagged 10 diagnostics before the change and none after.
• I dropped the %zu → ET_PRIsize_t migration and left it for its own change, so the touched functions are consistent again.
• prod_int_out uses multiply(), and the im2row operators use ET_INLINE.