Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[NFC][clang][FMV][TargetInfo] Refactor API for FMV feature priority. #116257

Open
wants to merge 2 commits into
base: main
Choose a base branch
from

Conversation

labrinea
Copy link
Collaborator

Currently we have code with target hooks in CodeGenModule shared between X86 and AArch64 for sorting MultiVersionResolverOptions. Those are used when generating IFunc resolvers for FMV. The RISCV target has different criteria for sorting, therefore it repeats sorting after calling CodeGenFunction::EmitMultiVersionResolver.

I am moving the FMV priority logic in TargetInfo, so that it can be implemented by the TargetParser which then makes it possible to query it from llvm. Here is an example why this is handy: #87939

Currently we have code with target hooks in CodeGenModule shared
between X86 and AArch64 for sorting MultiVersionResolverOptions.
Those are used when generating IFunc resolvers for FMV. The RISCV
target has different criteria for sorting, therefore it repeats
sorting after calling CodeGenFunction::EmitMultiVersionResolver.

I am moving the FMV priority logic in TargetInfo, so that it can
be implemented by the TargetParser which then makes it possible
to query it from llvm. Here is an example why this is handy:
llvm#87939
@llvmbot llvmbot added clang Clang issues not falling into any other category backend:AArch64 backend:RISC-V clang:frontend Language frontend issues, e.g. anything involving "Sema" clang:codegen labels Nov 14, 2024
@llvmbot
Copy link

llvmbot commented Nov 14, 2024

@llvm/pr-subscribers-clang-codegen
@llvm/pr-subscribers-backend-risc-v
@llvm/pr-subscribers-backend-aarch64

@llvm/pr-subscribers-clang

Author: Alexandros Lamprineas (labrinea)

Changes

Currently we have code with target hooks in CodeGenModule shared between X86 and AArch64 for sorting MultiVersionResolverOptions. Those are used when generating IFunc resolvers for FMV. The RISCV target has different criteria for sorting, therefore it repeats sorting after calling CodeGenFunction::EmitMultiVersionResolver.

I am moving the FMV priority logic in TargetInfo, so that it can be implemented by the TargetParser which then makes it possible to query it from llvm. Here is an example why this is handy: #87939


Full diff: https://github.com/llvm/llvm-project/pull/116257.diff

12 Files Affected:

  • (modified) clang/include/clang/Basic/TargetInfo.h (+1-5)
  • (modified) clang/lib/Basic/Targets/AArch64.cpp (+2-12)
  • (modified) clang/lib/Basic/Targets/AArch64.h (+1-2)
  • (modified) clang/lib/Basic/Targets/RISCV.cpp (+15)
  • (modified) clang/lib/Basic/Targets/RISCV.h (+1)
  • (modified) clang/lib/Basic/Targets/X86.cpp (+22-13)
  • (modified) clang/lib/Basic/Targets/X86.h (+1-1)
  • (modified) clang/lib/CodeGen/ABIInfo.cpp (+2-2)
  • (modified) clang/lib/CodeGen/CodeGenFunction.cpp (+5-33)
  • (modified) clang/lib/CodeGen/CodeGenModule.cpp (+6-17)
  • (modified) llvm/include/llvm/TargetParser/AArch64TargetParser.h (+3)
  • (modified) llvm/lib/TargetParser/AArch64TargetParser.cpp (+13)
diff --git a/clang/include/clang/Basic/TargetInfo.h b/clang/include/clang/Basic/TargetInfo.h
index 25eda907d20a7b..c27f7eb591f43d 100644
--- a/clang/include/clang/Basic/TargetInfo.h
+++ b/clang/include/clang/Basic/TargetInfo.h
@@ -1514,14 +1514,10 @@ class TargetInfo : public TransferrableTargetInfo,
 
   // Return the target-specific priority for features/cpus/vendors so
   // that they can be properly sorted for checking.
-  virtual unsigned multiVersionSortPriority(StringRef Name) const {
+  virtual unsigned getFMVPriority(ArrayRef<StringRef> Features) const {
     return 0;
   }
 
-  // Return the target-specific cost for feature
-  // that taken into account in priority sorting.
-  virtual unsigned multiVersionFeatureCost() const { return 0; }
-
   // Validate the contents of the __builtin_cpu_is(const char*)
   // argument.
   virtual bool validateCpuIs(StringRef Name) const { return false; }
diff --git a/clang/lib/Basic/Targets/AArch64.cpp b/clang/lib/Basic/Targets/AArch64.cpp
index eb8a3ada034482..4efc1841c836d0 100644
--- a/clang/lib/Basic/Targets/AArch64.cpp
+++ b/clang/lib/Basic/Targets/AArch64.cpp
@@ -705,18 +705,8 @@ AArch64TargetInfo::getVScaleRange(const LangOptions &LangOpts) const {
   return std::nullopt;
 }
 
-unsigned AArch64TargetInfo::multiVersionSortPriority(StringRef Name) const {
-  if (Name == "default")
-    return 0;
-  if (auto Ext = llvm::AArch64::parseFMVExtension(Name))
-    return Ext->Priority;
-  return 0;
-}
-
-unsigned AArch64TargetInfo::multiVersionFeatureCost() const {
-  // Take the maximum priority as per feature cost, so more features win.
-  constexpr unsigned MaxFMVPriority = 1000;
-  return MaxFMVPriority;
+unsigned AArch64TargetInfo::getFMVPriority(ArrayRef<StringRef> Features) const {
+  return llvm::AArch64::getFMVPriority(Features);
 }
 
 bool AArch64TargetInfo::doesFeatureAffectCodeGen(StringRef Name) const {
diff --git a/clang/lib/Basic/Targets/AArch64.h b/clang/lib/Basic/Targets/AArch64.h
index 4c25bdb5bb16df..68a8b1ebad8cde 100644
--- a/clang/lib/Basic/Targets/AArch64.h
+++ b/clang/lib/Basic/Targets/AArch64.h
@@ -137,8 +137,7 @@ class LLVM_LIBRARY_VISIBILITY AArch64TargetInfo : public TargetInfo {
   void fillValidCPUList(SmallVectorImpl<StringRef> &Values) const override;
   bool setCPU(const std::string &Name) override;
 
-  unsigned multiVersionSortPriority(StringRef Name) const override;
-  unsigned multiVersionFeatureCost() const override;
+  unsigned getFMVPriority(ArrayRef<StringRef> Features) const override;
 
   bool useFP16ConversionIntrinsics() const override {
     return false;
diff --git a/clang/lib/Basic/Targets/RISCV.cpp b/clang/lib/Basic/Targets/RISCV.cpp
index eaaba7642bd7b2..d5312bb171af6c 100644
--- a/clang/lib/Basic/Targets/RISCV.cpp
+++ b/clang/lib/Basic/Targets/RISCV.cpp
@@ -476,6 +476,21 @@ ParsedTargetAttr RISCVTargetInfo::parseTargetAttr(StringRef Features) const {
   return Ret;
 }
 
+unsigned RISCVTargetInfo::getFMVPriority(ArrayRef<StringRef> Features) const {
+  SmallVector<StringRef, 8> Attrs;
+  Features[0].split(Attrs, ';');
+  // Default Priority is zero.
+  unsigned Priority = 0;
+  for (auto Attr : Attrs) {
+    if (Attr.consume_front("priority=")) {
+      unsigned Result;
+      if (!Attr.getAsInteger(0, Result))
+        Priority = Result;
+    }
+  }
+  return Priority;
+}
+
 TargetInfo::CallingConvCheckResult
 RISCVTargetInfo::checkCallingConvention(CallingConv CC) const {
   switch (CC) {
diff --git a/clang/lib/Basic/Targets/RISCV.h b/clang/lib/Basic/Targets/RISCV.h
index 3b418585ab4a39..0a6caeb54bbd5d 100644
--- a/clang/lib/Basic/Targets/RISCV.h
+++ b/clang/lib/Basic/Targets/RISCV.h
@@ -122,6 +122,7 @@ class RISCVTargetInfo : public TargetInfo {
   void fillValidTuneCPUList(SmallVectorImpl<StringRef> &Values) const override;
   bool supportsTargetAttributeTune() const override { return true; }
   ParsedTargetAttr parseTargetAttr(StringRef Str) const override;
+  unsigned getFMVPriority(ArrayRef<StringRef> Features) const override;
 
   std::pair<unsigned, unsigned> hardwareInterferenceSizes() const override {
     return std::make_pair(32, 32);
diff --git a/clang/lib/Basic/Targets/X86.cpp b/clang/lib/Basic/Targets/X86.cpp
index 086b4415412e67..ce5ccd0447b437 100644
--- a/clang/lib/Basic/Targets/X86.cpp
+++ b/clang/lib/Basic/Targets/X86.cpp
@@ -1363,19 +1363,28 @@ static llvm::X86::ProcessorFeatures getFeature(StringRef Name) {
   // correct, so it asserts if the value is out of range.
 }
 
-unsigned X86TargetInfo::multiVersionSortPriority(StringRef Name) const {
-  // Valid CPUs have a 'key feature' that compares just better than its key
-  // feature.
-  using namespace llvm::X86;
-  CPUKind Kind = parseArchX86(Name);
-  if (Kind != CK_None) {
-    ProcessorFeatures KeyFeature = getKeyFeature(Kind);
-    return (getFeaturePriority(KeyFeature) << 1) + 1;
-  }
-
-  // Now we know we have a feature, so get its priority and shift it a few so
-  // that we have sufficient room for the CPUs (above).
-  return getFeaturePriority(getFeature(Name)) << 1;
+unsigned X86TargetInfo::getFMVPriority(ArrayRef<StringRef> Features) const {
+  auto getPriority = [this](StringRef Feature) -> unsigned {
+    if (Feature.empty())
+      return 0;
+
+    // Valid CPUs have a 'key feature' that compares just better than its key
+    // feature.
+    using namespace llvm::X86;
+    CPUKind Kind = parseArchX86(Feature);
+    if (Kind != CK_None) {
+      ProcessorFeatures KeyFeature = getKeyFeature(Kind);
+      return (getFeaturePriority(KeyFeature) << 1) + 1;
+    }
+    // Now we know we have a feature, so get its priority and shift it a few so
+    // that we have sufficient room for the CPUs (above).
+    return getFeaturePriority(getFeature(Feature)) << 1;
+  };
+
+  unsigned Priority = 0;
+  for (StringRef Feature : Features)
+    Priority = std::max(Priority, getPriority(Feature));
+  return Priority;
 }
 
 bool X86TargetInfo::validateCPUSpecificCPUDispatch(StringRef Name) const {
diff --git a/clang/lib/Basic/Targets/X86.h b/clang/lib/Basic/Targets/X86.h
index 06a7eed8177cb2..3ed36c8fa724b5 100644
--- a/clang/lib/Basic/Targets/X86.h
+++ b/clang/lib/Basic/Targets/X86.h
@@ -384,7 +384,7 @@ class LLVM_LIBRARY_VISIBILITY X86TargetInfo : public TargetInfo {
     return CPU != llvm::X86::CK_None;
   }
 
-  unsigned multiVersionSortPriority(StringRef Name) const override;
+  unsigned getFMVPriority(ArrayRef<StringRef> Features) const override;
 
   bool setFPMath(StringRef Name) override;
 
diff --git a/clang/lib/CodeGen/ABIInfo.cpp b/clang/lib/CodeGen/ABIInfo.cpp
index edd7146dc1ac76..8e76cf15b642c6 100644
--- a/clang/lib/CodeGen/ABIInfo.cpp
+++ b/clang/lib/CodeGen/ABIInfo.cpp
@@ -218,8 +218,8 @@ void ABIInfo::appendAttributeMangling(StringRef AttrStr,
     // only have "+" prefixes here.
     assert(LHS.starts_with("+") && RHS.starts_with("+") &&
            "Features should always have a prefix.");
-    return TI.multiVersionSortPriority(LHS.substr(1)) >
-           TI.multiVersionSortPriority(RHS.substr(1));
+    return TI.getFMVPriority({LHS.substr(1)}) >
+           TI.getFMVPriority({RHS.substr(1)});
   });
 
   bool IsFirst = true;
diff --git a/clang/lib/CodeGen/CodeGenFunction.cpp b/clang/lib/CodeGen/CodeGenFunction.cpp
index 6ead45793742d6..7ac60e8dc0508e 100644
--- a/clang/lib/CodeGen/CodeGenFunction.cpp
+++ b/clang/lib/CodeGen/CodeGenFunction.cpp
@@ -2904,24 +2904,6 @@ void CodeGenFunction::EmitMultiVersionResolver(
   }
 }
 
-static unsigned getPriorityFromAttrString(StringRef AttrStr) {
-  SmallVector<StringRef, 8> Attrs;
-
-  AttrStr.split(Attrs, ';');
-
-  // Default Priority is zero.
-  unsigned Priority = 0;
-  for (auto Attr : Attrs) {
-    if (Attr.consume_front("priority=")) {
-      unsigned Result;
-      if (!Attr.getAsInteger(0, Result))
-        Priority = Result;
-    }
-  }
-
-  return Priority;
-}
-
 void CodeGenFunction::EmitRISCVMultiVersionResolver(
     llvm::Function *Resolver, ArrayRef<MultiVersionResolverOption> Options) {
 
@@ -2939,20 +2921,10 @@ void CodeGenFunction::EmitRISCVMultiVersionResolver(
   bool HasDefault = false;
   unsigned DefaultIndex = 0;
 
-  SmallVector<CodeGenFunction::MultiVersionResolverOption, 10> CurrOptions(
-      Options);
-
-  llvm::stable_sort(
-      CurrOptions, [](const CodeGenFunction::MultiVersionResolverOption &LHS,
-                      const CodeGenFunction::MultiVersionResolverOption &RHS) {
-        return getPriorityFromAttrString(LHS.Conditions.Features[0]) >
-               getPriorityFromAttrString(RHS.Conditions.Features[0]);
-      });
-
   // Check the each candidate function.
-  for (unsigned Index = 0; Index < CurrOptions.size(); Index++) {
+  for (unsigned Index = 0; Index < Options.size(); Index++) {
 
-    if (CurrOptions[Index].Conditions.Features[0].starts_with("default")) {
+    if (Options[Index].Conditions.Features[0].starts_with("default")) {
       HasDefault = true;
       DefaultIndex = Index;
       continue;
@@ -2963,7 +2935,7 @@ void CodeGenFunction::EmitRISCVMultiVersionResolver(
     std::vector<std::string> TargetAttrFeats =
         getContext()
             .getTargetInfo()
-            .parseTargetAttr(CurrOptions[Index].Conditions.Features[0])
+            .parseTargetAttr(Options[Index].Conditions.Features[0])
             .Features;
 
     if (TargetAttrFeats.empty())
@@ -3004,7 +2976,7 @@ void CodeGenFunction::EmitRISCVMultiVersionResolver(
     llvm::BasicBlock *RetBlock = createBasicBlock("resolver_return", Resolver);
     CGBuilderTy RetBuilder(*this, RetBlock);
     CreateMultiVersionResolverReturn(
-        CGM, Resolver, RetBuilder, CurrOptions[Index].Function, SupportsIFunc);
+        CGM, Resolver, RetBuilder, Options[Index].Function, SupportsIFunc);
     llvm::BasicBlock *ElseBlock = createBasicBlock("resolver_else", Resolver);
 
     Builder.SetInsertPoint(CurBlock);
@@ -3017,7 +2989,7 @@ void CodeGenFunction::EmitRISCVMultiVersionResolver(
   if (HasDefault) {
     Builder.SetInsertPoint(CurBlock);
     CreateMultiVersionResolverReturn(CGM, Resolver, Builder,
-                                     CurrOptions[DefaultIndex].Function,
+                                     Options[DefaultIndex].Function,
                                      SupportsIFunc);
     return;
   }
diff --git a/clang/lib/CodeGen/CodeGenModule.cpp b/clang/lib/CodeGen/CodeGenModule.cpp
index ba376f9ecfacde..fa569810da6296 100644
--- a/clang/lib/CodeGen/CodeGenModule.cpp
+++ b/clang/lib/CodeGen/CodeGenModule.cpp
@@ -4216,22 +4216,11 @@ static void ReplaceUsesOfNonProtoTypeWithRealFunction(llvm::GlobalValue *Old,
                                                       llvm::Function *NewFn);
 
 static unsigned
-TargetMVPriority(const TargetInfo &TI,
-                 const CodeGenFunction::MultiVersionResolverOption &RO) {
-  unsigned Priority = 0;
-  unsigned NumFeatures = 0;
-  for (StringRef Feat : RO.Conditions.Features) {
-    Priority = std::max(Priority, TI.multiVersionSortPriority(Feat));
-    NumFeatures++;
-  }
-
-  if (!RO.Conditions.Architecture.empty())
-    Priority = std::max(
-        Priority, TI.multiVersionSortPriority(RO.Conditions.Architecture));
-
-  Priority += TI.multiVersionFeatureCost() * NumFeatures;
-
-  return Priority;
+getFMVPriority(const TargetInfo &TI,
+               const CodeGenFunction::MultiVersionResolverOption &RO) {
+  llvm::SmallVector<StringRef, 8> Features{RO.Conditions.Features};
+  Features.push_back(RO.Conditions.Architecture);
+  return TI.getFMVPriority(Features);
 }
 
 // Multiversion functions should be at most 'WeakODRLinkage' so that a different
@@ -4362,7 +4351,7 @@ void CodeGenModule::emitMultiVersionFunctions() {
     llvm::stable_sort(
         Options, [&TI](const CodeGenFunction::MultiVersionResolverOption &LHS,
                        const CodeGenFunction::MultiVersionResolverOption &RHS) {
-          return TargetMVPriority(TI, LHS) > TargetMVPriority(TI, RHS);
+          return getFMVPriority(TI, LHS) > getFMVPriority(TI, RHS);
         });
     CodeGenFunction CGF(*this);
     CGF.EmitMultiVersionResolver(ResolverFunc, Options);
diff --git a/llvm/include/llvm/TargetParser/AArch64TargetParser.h b/llvm/include/llvm/TargetParser/AArch64TargetParser.h
index d7b1ba511f95d3..1311329821828f 100644
--- a/llvm/include/llvm/TargetParser/AArch64TargetParser.h
+++ b/llvm/include/llvm/TargetParser/AArch64TargetParser.h
@@ -268,6 +268,9 @@ void fillValidCPUArchList(SmallVectorImpl<StringRef> &Values);
 
 bool isX18ReservedByDefault(const Triple &TT);
 
+// Return the priority for a given set of FMV features.
+unsigned getFMVPriority(ArrayRef<StringRef> Features);
+
 // For given feature names, return a bitmask corresponding to the entries of
 // AArch64::CPUFeatures. The values in CPUFeatures are not bitmasks
 // themselves, they are sequential (0, 1, 2, 3, ...).
diff --git a/llvm/lib/TargetParser/AArch64TargetParser.cpp b/llvm/lib/TargetParser/AArch64TargetParser.cpp
index 45ecc4f24c2afd..fe5ab0fabefa6e 100644
--- a/llvm/lib/TargetParser/AArch64TargetParser.cpp
+++ b/llvm/lib/TargetParser/AArch64TargetParser.cpp
@@ -48,6 +48,19 @@ std::optional<AArch64::ArchInfo> AArch64::ArchInfo::findBySubArch(StringRef SubA
   return {};
 }
 
+unsigned AArch64::getFMVPriority(ArrayRef<StringRef> Features) {
+  constexpr unsigned MaxFMVPriority = 1000;
+  unsigned Priority = 0;
+  unsigned NumFeatures = 0;
+  for (StringRef Feature : Features) {
+    if (auto Ext = parseFMVExtension(Feature)) {
+      Priority = std::max(Priority, Ext->Priority);
+      NumFeatures++;
+    }
+  }
+  return Priority + MaxFMVPriority * NumFeatures;
+}
+
 uint64_t AArch64::getCpuSupportsMask(ArrayRef<StringRef> FeatureStrs) {
   uint64_t FeaturesMask = 0;
   for (const StringRef &FeatureStr : FeatureStrs) {

Copy link

github-actions bot commented Nov 14, 2024

✅ With the latest revision this PR passed the C/C++ code formatter.

const CodeGenFunction::MultiVersionResolverOption &RO) {
llvm::SmallVector<StringRef, 8> Features{RO.Conditions.Features};
Features.push_back(RO.Conditions.Architecture);
return TI.getFMVPriority(Features);
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Could it pass the explicit args for RO.Conditions.Features and RO.Conditions.Architecture?

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The struct MultiVersionResolverOption was originally tailored to x86 if I am not mistaken. RO.Conditions.Architecture isn't suitable for AArch64, and correct me if I am wrong but so far it doesn't seem that risc-v is using it either? Even on x86 Architecture is treated the same as any other feature string in the existing hook multiVersionSortPriority, so I don't see the need for explicit argument.

return TI.multiVersionSortPriority(LHS.substr(1)) >
TI.multiVersionSortPriority(RHS.substr(1));
return TI.getFMVPriority({LHS.substr(1)}) >
TI.getFMVPriority({RHS.substr(1)});
Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Perhaps there is value in adding another hook getFMVPriority(StringRef feature) which would be called on each feature of ArrayRef<StringRef> Features from the other hook, in order to avoid this inefficiency here of having to create a single element smallvector each time?

Copy link
Collaborator

@erichkeane erichkeane left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I wont be able to get back to this until after WG21, but some comments.


unsigned Priority = 0;
for (StringRef Feature : Features)
Priority = std::max(Priority, getPriority(Feature));
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I don't think 'max' does the right thing here, does it? The whole point was to put together a mask representation that matches the cpuid call (Best we could!) that is the collection of all the features. That way you could have a list of features and have them be differentiated (that is, some of the FMV supports just a list of features, and we want to be able to differentiate between high-feature, low-feature and high-feature, lower-feature.

NumFeatures++;
}
}
return Priority + MaxFMVPriority * NumFeatures;
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I don't understand the algorithm here. This seems like sorts resolvees first by how many conditions they've got, and then by the highest priority of each in that group. But what we actually need is a total order on two sets. Maybe the API should be shaped more like a comparison operator on feature sets, rather than a serialized index constructed from members's priorities?

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

(and yeah, I see that this is just a pure refactor, so it's not "new", but still...)

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
backend:AArch64 backend:RISC-V clang:codegen clang:frontend Language frontend issues, e.g. anything involving "Sema" clang Clang issues not falling into any other category
Projects
None yet
Development

Successfully merging this pull request may close these issues.

5 participants