Optimize dot product using restrict keyword (Fixes #4393)

[Vallabh-1504]

Description

addresses issue #4393 by adding the restrict keyword to the dot product functions. This informs the compiler that the input arrays do not overlap, enabling better SIMD vectorization and potential performance improvements.

Changes

• Defined a TESS_RESTRICT macro in src/arch/dotproduct.h to handle compiler differences:
  • Uses __restrict for MSVC.
  • Uses __restrict__ for GCC/Clang.
• Updated function signatures to use TESS_RESTRICT on pointer arguments (u and v) in:
  • DotProductNative
  • DotProductAVX / AVX512F
  • DotProductSSE
  • DotProductFMA
  • DotProductNEON

Verification

• Code compiles successfully.
• Passed existing tests (CI checks pending).

[stweil]

Thanks for addressing this issue. Did you compare the generated code? Before merging this pull request, I want to be sure that it really improves the code.

[Vallabh-1504]

hi, thanks for the review!

I haven't compared the generated assembly for this specific build because I am working in a limited environment and relying on CI.

However, I implemented this based on the request in the issue. The standard behavior of restrict (and __restrict on MSVC) is to tell the compiler that u and v do not overlap. This typically allows the compiler to skip runtime alias checks (loop versioning) and vectorize the loop more aggressively.

Since I cannot generate the assembly locally, would you be able to verify if the output looks correct on your end?

[stweil]

Test results:

• aarch64 with g++ (Debian 14.2.0-19): no effect on generated code
• aarch64 with Debian clang version 19.1.7 (3+b1): no effect on generated code
• x86_64 with Debian clang version 19.1.7 (3+b1): no effect on generated code

[stweil]

A hardcoded __restrict is accepted by MSC++, g++ and clang++. Therefore I don't think we need TESS_RESTRICT. I suggest to wait with an update of the pull request until we know that it really has an effect.

[Vallabh-1504]

Thanks for the feedback!

1. I agree. I can update the PR to remove TESS_RESTRICT and use __restrict directly to simplify the code.

2. I ran a local verification on MSVC and the generated assembly was identical for both versions. The compiler is smart enough to handle the const pointers without the explicit hint.

@amitdo – Since you opened the original issue, did you have a specific scenario (specific architecture or older compiler version) where this provided a benefit?

I will wait for confirmation before updating the PR.

[amitdo]

I suggest to test this also on a machine with an amd64 CPU.

The test run time should be long enough to reduce the influence of the initial data loading.

[stweil]

The modified code has no effect on the generated binaries in all settings which I tested up to now. Therefore a runtime test is currently not needed.