[llvm] [clang] support avx512fp16 target multiversioning (PR #151320)

[via llvm-commits]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-backend-x86

Author: Jinjie Huang  (Jinjie-Huang)

<details>
<summary>Changes</summary>

Currently, multiversioning for the avx512fp16 target seems not yet supported due to a validation check in the frontend (SemaDecl). 

While compiler-rt already supports detecting this feature bit from a previous [patch](https://github.com/llvm/llvm-project/pull/78384/files), so this patch tries to enables the full multiversioning functionality of avx512fp16 in Clang by defining a priority for avx512fp16 in X86TargetParser.def.

---
Full diff: https://github.com/llvm/llvm-project/pull/151320.diff


2 Files Affected:

- (modified) llvm/include/llvm/TargetParser/X86TargetParser.def (+1-1) 
- (modified) llvm/lib/TargetParser/X86TargetParser.cpp (+2-2) 


``````````diff
diff --git a/llvm/include/llvm/TargetParser/X86TargetParser.def b/llvm/include/llvm/TargetParser/X86TargetParser.def
index 5b719a2b0eb39..0833d3c2ad860 100644
--- a/llvm/include/llvm/TargetParser/X86TargetParser.def
+++ b/llvm/include/llvm/TargetParser/X86TargetParser.def
@@ -239,7 +239,7 @@ X86_FEATURE_COMPAT(KL,              "kl",                     0)
 X86_FEATURE       (FXSR,            "fxsr")
 X86_FEATURE_COMPAT(WIDEKL,          "widekl",                 0)
 X86_FEATURE_COMPAT(AVXVNNI,         "avxvnni",                0)
-X86_FEATURE_COMPAT(AVX512FP16,      "avx512fp16",             0)
+X86_FEATURE_COMPAT(AVX512FP16,      "avx512fp16",            38)
 X86_FEATURE       (CCMP,            "ccmp")
 X86_FEATURE       (Push2Pop2,       "push2pop2")
 X86_FEATURE       (PPX,             "ppx")
diff --git a/llvm/lib/TargetParser/X86TargetParser.cpp b/llvm/lib/TargetParser/X86TargetParser.cpp
index 9cd35e35d4bc9..3e752d49eb9b0 100644
--- a/llvm/lib/TargetParser/X86TargetParser.cpp
+++ b/llvm/lib/TargetParser/X86TargetParser.cpp
@@ -776,13 +776,13 @@ unsigned llvm::X86::getFeaturePriority(ProcessorFeatures Feat) {
 #ifndef NDEBUG
   // Check that priorities are set properly in the .def file. We expect that
   // "compat" features are assigned non-duplicate consecutive priorities
-  // starting from one (1, ..., 37) and multiple zeros.
+  // starting from one (1, ..., 38) and multiple zeros.
 #define X86_FEATURE_COMPAT(ENUM, STR, PRIORITY) PRIORITY,
   unsigned Priorities[] = {
 #include "llvm/TargetParser/X86TargetParser.def"
   };
   std::array<unsigned, std::size(Priorities)> HelperList;
-  const size_t MaxPriority = 37;
+  const size_t MaxPriority = 38;
   std::iota(HelperList.begin(), HelperList.begin() + MaxPriority + 1, 0);
   for (size_t i = MaxPriority + 1; i != std::size(Priorities); ++i)
     HelperList[i] = 0;

``````````

</details>


https://github.com/llvm/llvm-project/pull/151320