[clang] [llvm] [FMV][AIX] Implement target_clones part 2 (target-features) (PR #206786)

[zhijian lin via llvm-commits]

================
@@ -727,18 +727,55 @@ llvm::APInt PPCTargetInfo::getFMVPriority(ArrayRef<StringRef> Features) const {
     return llvm::APInt(32, 0);
   assert(Features.size() == 1 && "one feature/cpu per clone on PowerPC");
   ParsedTargetAttr ParsedAttr = parseTargetAttr(Features[0]);
+
+  // Priority scheme:
+  // CPU specifications: 100-500 (pwr7=100, pwr8=200, ..., pwr11=500)
+  // For target-features, they can be divided into 3 categories:
+  // 1) non-CPU properties (e.g. invariant-function-descriptors); those cannot
+  //    be tested at runtime, and are currently excluded from target_clones.
+  // 2) CPU properties that cannot be disabled (e.g. mma); these features map to
+  //    CPUs directly:
+  //     +feature => __builtin_cpu_supports("<minimum-cpu>")
+  //              => true for CPU <minimum-cpu> and above.
+  //     -feature => !__builtin_cpu_supports("<minimum-cpu>")
+  //              => true for CPU <minimum-cpu-minus-one> and below.
+  // 3) CPU properties that can be disabled (e.g. vsx); those can map to CPUs
+  // directly for the positive requirement (same as (2));
+  // for the negative requirement checking the CPU is incorrect:
+  //   target_clones(no-vsx, cpu=pwr8)
+  // should pick no-vsx when vsx is disabled at runtime, so we need to test
+  // negative form of category 3 first, and will only allow one negative form
+  // from this category on a target_clones.
   if (!ParsedAttr.CPU.empty()) {
     int Priority = llvm::StringSwitch<int>(ParsedAttr.CPU)
-                       .Case("pwr7", 1)
-                       .Case("pwr8", 2)
-                       .Case("pwr9", 3)
-                       .Case("pwr10", 4)
-                       .Case("pwr11", 5)
+#define PPC_AIX_CLONES_CPU(CPU_NAME, _, PRIORITY) .Case(CPU_NAME, PRIORITY)
+#include "llvm/TargetParser/PPCTargetParser.def"
+                       .Default(0);
+    return llvm::APInt(32, Priority);
+  }
+
+  // Feature strings
+  if (ParsedAttr.Features.size() == 1) {
+    StringRef Feature = ParsedAttr.Features[0];
+    bool IsNegated = Feature.starts_with("-");
+    // Remove leading '+' or '-'
+    if (Feature.starts_with("+") || Feature.starts_with("-"))
+      Feature = Feature.drop_front(1);
+
+    // Check if this is a negative category 3 feature (highest priority)
+    // Sema guarantees there's only one such version on a target_clones.
+    if (IsNegated && llvm::PPC::canDisableFeatureOnAIX(Feature))
+      return llvm::APInt(32, NEGATIVE_FEATURE_PRIORITY);
+
+    // Regular feature priority (positive or negative category 2)
+    int Priority = llvm::StringSwitch<int>(Feature)
+#define PPC_AIX_CLONES_FEATURE(FEATURE_NAME, _, PRIORITY)                      \
+  .Case(FEATURE_NAME, PRIORITY)
+#include "llvm/TargetParser/PPCTargetParser.def"
                        .Default(0);
     return llvm::APInt(32, Priority);
   }
-  assert(false && "unimplemented");
-  return llvm::APInt(32, 0);
+  llvm_unreachable("Invalid target_clones parameter");
----------------
diggerlin wrote:

agree getFMVPriority is called during Codegen, after Sema has verified.  but if use input void __attribute__((target_clones("mma,vsx", "default"))), I do not think you Diag "mma,vsx" is invalid in the somewhere ? 
it will reach the   llvm_unreachable("Invalid target_clones parameter"); is it correct ?

https://github.com/llvm/llvm-project/pull/206786