Intel prepares GCC for Razor Lake and Coral Rapids with a new operation for cryptographic calculations
Intel has sent the GCC project a series of patches that prepare for the recognition of Razor Lake and Coral Rapids. The messages, published on November 13 by Haochen Jiang and with contributions from Qin Sunallow us to review how it proposes to identify these future families and what instructions it intends to enable for them.
The simplest change corresponds to Razor Lake, which appears as an alias for Nova Lake because both share the set of functions included in this proposal. Coral Rapids receives a more extensive addition, with self-identification, compilation options, and support for a new vector operation called VPMLDSABFLYDQ.
Razor Lake initially shares Nova Lake profile
Razor Lake patch adds its name to the GCC processor table, but associates it with the profile already used by Nova Lake. The explanation accompanying the shipment expressly states that they have the same set of functions. The code also incorporates the model values necessary to recognize the family.
This equivalence has a specific scope: describes the instructions and treatment proposed by this compiler. From this it cannot be deduced that the processors have the same cores, frequencies, caches or consumption. Nor does it allow us to calculate a difference in performance between both generations.
The modified documentation brings together the two names under a common list. Among the capabilities listed are AVX10.1, AVX10.2, APX_F and MOVRSalong with previous extensions. Thus, the work presented allows the Razor Lake name to be used within an already existing code generation infrastructure.
For software developers, this distinction avoids interpreting the new name as an obligation to rewrite each optimization. In this initial proposal, GCC reuses the Nova Lake profile. Any subsequent differences would have to appear in other changes or in additional documentation from the manufacturer.
Coral Rapids receives independent identification
The addition of Coral Rapids modifies fourteen files. Adds a processor identifier, recognition using family and model data, specific macros, and documentation to select this architecture. It also introduces cases in the compiler tests that check the handling of the new target.
The intended option is -march=coralrapids. This type of selection allows compiling taking as reference the capabilities of the chosen processor. Its use must correspond to the hardware where the program will be executed, especially when instructions are enabled that other equipment may not support.
Feature profile departs from Diamond Rapids and adds AVX10VPMLDSABFLYDQ. The documentary list includes AVX10.2 and various AMX capabilities, including AMX-FP8, AMX-TF32, AMX-AVX512 and AMX-MOVRS. It is convenient to read this list as the support proposed in the patch, without converting it into a complete commercial file.
The shipment also reuses existing cost tables and incorporates instruction planning adjustments. They are internal decisions of the compiler to organize the generated code. These tables, by themselves, do not provide a measure of processor performance nor are they a substitute for testing real-world applications.
A cryptographic operation with 128, 256 and 512 bit variants
The series includes a separate patch for VPMLDSABFLYDQ. Its scope includes detection of the capability, an option to enable it, and intrinsic functions with which the programmer can request the operation from C or C++ code, without directly writing the entire sequence in assembler.
The added file contains variants for 128, 256 and 512 bit vectors. The tests verify that GCC issues the corresponding instruction with registers of each width. Another block compares its results with reference calculations, using modular multiplication operations and addition and subtraction combinations.
This detail specifies what the support adds: the compiler needs to recognize the capability and have a way to generate the instruction, as well as identify the processor.
The operation is linked to ML-DSA calculations, a digital signature algorithm designed to resist attacks with quantum computers. Its function is limited to a part of mathematical processing; the patch does not announce a measured speedup of the entire algorithm.
To take advantage of it, cryptographic software will have to incorporate a compatible implementation and run on hardware that supports it. The messages do not provide percentages of improvement, latencies or consumption comparisons that allow quantifying their advantage over other instructions.
Patch submission still requires integration
The messages consulted are identified as initial support proposals. This phase allows you to review the code and discuss changes before incorporating it. Publishing a series to the development list does not, by itself, prove that a distributed version of GCC already includes all those modifications.
The submissions also do not specify which final version of the compiler will be supported. For developers and distributors, the next verifiable data will be its integration and the documentation of the version that incorporates it. The brand does not provide prices, commercial models or availability dates for Razor Lake or Coral Rapids in these patches.
