[X86] computeKnownBitsForTargetNode - add X86ISD::GF2P8AFFINEQB handling · llvm/llvm-project#190520

(13 留言) (0 反應) (2 負責人)C++ (26,378 star) (10,782 fork)batch import

backend:X86good first issue

描述

X86ISD::GF2P8AFFINEQB nodes are being used for many shift/rotate style expansions, it'd be very useful if we can access the known bits to attempt additional folding afterward (e.g #190502).

This should only be attempted by somebody with access to a CPU with GFNI support to allow them to actually test this properly (although constexpr testing can help now that GFNI intrinsics support it).

Reference: https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#ig_expand=172,3053,2650,3028,3046,849,3516,3630,3054,3053,3054,4389,2239,2239,4602,205,196,4461,3954,1601,2140,417,2873,2851,2571,5840,2571,2927,3639&text=VGF2P8AFFINEQB

(see also GFNIAffine() in llvm-project\clang\lib\AST\ExprConstant.cpp)

貢獻者指南

技術棧: cpp
領域: tooling
議題類型: feature
難度: 5
預計時間: over 1 week
活動狀態: active
清晰度: clear
前置要求: LLVM codegen infrastructureX86 backend knowledgeknown bits analysisGFNI intrinsics
新手友善度: 15
研究方向: The issue requests adding handling for X86ISD::GF2P8AFFINEQB in the computeKnownBitsForTargetNode function. Start by studying the existing implementations for other X86ISD nodes in the X86 backend's target specific lowering. The reference implementation for GFNI intrinsics in clang's ExprConstant.cpp (GFNIAffine) can provide insight. The Intel intrinsics guide details the behavior of the VGF2P8AFFINEQB instruction. Ensure testing is done either on a CPU with GFNI support or using constexpr evaluation as mentioned.