[Target] Use LLVM target parser for determining Arm(R) A-Profile Architecture features

lhutton1 · lhutton1 · commit dfcdb3ec75a6 · 2024-01-17T17:28:16.000Z
Currently, target features are determined by a set of fixed checks on
the target string. This works well for checking support of a small
number of simple features, but it doesn't scale. Some problems include:
- There are many non-trivial conditions for which a feature may(not) be
  available. It is easy to miss these with the current implementation.
- The inclusion of some features in a target string can imply other
  features. For example, "+sve" implies "+neon". This currently isn't
  taken into account.
- The tests in tests/cpp/target/parsers/aprofile_test.c suggest that
  targets such as "llvm -mcpu=cortex-a+neon" and "llvm -mattr=+noneon"
  are supported target strings. The features will be correctly parsed in
  TVM, however, they are not valid in LLVM. Therefore, it's possible
  that TVM and LLVM have different understanding of the features
  available.

This commit uses the more robust LLVM target parser to determine support
for the features in TVM. It leverages previous infrastructure added to
TVM for obtaining a list of all supported features given an input
target, and uses this to check the existance of certain features we're
interested in. It should be trivial to grow this list over time. As a
result of this change, the problems mentioned above are solved.

In the current form, this commit drops support for target strings such
as "llvm -mcpu=cortex-a+neon" and "llvm -mattr=+noneon". A scan of the
codebase suggests this functionality is not in use (only in test cases).
Should we feel the need to support them, or have a smoother migration
for downstream users of TVM we can add a translator to the parser to
convert these into LLVM compatible targets.

Change-Id: Ic2bf3b68c8af74025ec388d304bd014624c0c585
diff --git a/src/target/llvm/llvm_instance.cc b/src/target/llvm/llvm_instance.cc
@@ -199,21 +199,23 @@ std::ostream& operator<<(std::ostream& os, const LLVMTargetInfo::Option& opt) {
   return os;
 }
 
-LLVMTargetInfo::LLVMTargetInfo(LLVMInstance& instance, const Target& target) {
-  triple_ = target->GetAttr<String>("mtriple").value_or("default");
+LLVMTargetInfo::LLVMTargetInfo(LLVMInstance& instance, const Target& target)
+    : LLVMTargetInfo(instance, target->Export()) {}
 
+LLVMTargetInfo::LLVMTargetInfo(LLVMInstance& instance, const TargetJSON& target) {
+  triple_ = Downcast<String>(target.Get("mtriple").value_or(String("default")));
   if (triple_.empty() || triple_ == "default") {
     triple_ = llvm::sys::getDefaultTargetTriple();
   }
-  cpu_ = target->GetAttr<String>("mcpu").value_or(defaults::cpu);
+  cpu_ = Downcast<String>(target.Get("mcpu").value_or(String(defaults::cpu)));
 
-  if (const Optional<Array<String>>& v = target->GetAttr<Array<String>>("mattr")) {
+  if (const auto& v = Downcast<Optional<Array<String>>>(target.Get("mattr"))) {
     for (const String& s : v.value()) {
       attrs_.push_back(s);
     }
   }
   // llvm module target
-  if (target->kind->name == "llvm") {
+  if (Downcast<String>(target.Get("kind")) == "llvm") {
     // legalize -mcpu with the target -mtriple
     auto arches = GetAllLLVMTargetArches();
     bool has_arch =
@@ -224,7 +226,7 @@ LLVMTargetInfo::LLVMTargetInfo(LLVMInstance& instance, const Target& target) {
     }
   }
 
-  if (const Optional<Array<String>>& v = target->GetAttr<Array<String>>("cl-opt")) {
+  if (const auto& v = Downcast<Optional<Array<String>>>(target.Get("cl-opt"))) {
     llvm::StringMap<llvm::cl::Option*>& options = llvm::cl::getRegisteredOptions();
     bool parse_error = false;
     for (const String& s : v.value()) {
@@ -245,7 +247,7 @@ LLVMTargetInfo::LLVMTargetInfo(LLVMInstance& instance, const Target& target) {
   }
 
   llvm::FloatABI::ABIType float_abi = llvm::FloatABI::Default;
-  if (const Optional<String>& v = target->GetAttr<String>("mfloat-abi")) {
+  if (const auto& v = Downcast<Optional<String>>(target.Get("mfloat-abi"))) {
     String value = v.value();
     if (value == "hard") {
       float_abi = llvm::FloatABI::Hard;
@@ -268,14 +270,14 @@ LLVMTargetInfo::LLVMTargetInfo(LLVMInstance& instance, const Target& target) {
   target_options_.NoInfsFPMath = false;
   target_options_.NoNaNsFPMath = true;
   target_options_.FloatABIType = float_abi;
-  if (const Optional<String>& v = target->GetAttr<String>("mabi")) {
-    target_options_.MCOptions.ABIName = v.value();
+  if (target.find("mabi") != target.end()) {
+    target_options_.MCOptions.ABIName = Downcast<String>(target.Get("mabi"));
   }
 
-  auto maybe_level = target->GetAttr<Integer>("opt-level");
+  auto maybe_level = Downcast<Integer>(target.Get("opt-level"));
 #if TVM_LLVM_VERSION <= 170
   if (maybe_level.defined()) {
-    int level = maybe_level.value()->value;
+    int level = maybe_level->value;
     if (level <= 0) {
       opt_level_ = llvm::CodeGenOpt::None;
     } else if (level == 1) {
@@ -312,7 +314,7 @@ LLVMTargetInfo::LLVMTargetInfo(LLVMInstance& instance, const Target& target) {
   // Fast math options
 
   auto GetBoolFlag = [&target](llvm::StringRef flag) -> bool {
-    return target->GetAttr<Bool>(flag.str()).value_or(Bool(false));
+    return Downcast<Bool>(target.Get(flag.str()).value_or(Bool(false)));
   };
   if (GetBoolFlag("fast-math")) {
 #if TVM_LLVM_VERSION >= 60
@@ -831,7 +833,7 @@ const Array<String> LLVMTargetInfo::GetAllLLVMTargetArches() const {
   return cpu_arches;
 }
 
-const Array<String> LLVMTargetInfo::GetAllLLVMCpuFeatures() const {
+const Map<String, String> LLVMTargetInfo::GetAllLLVMCpuFeatures() const {
   std::string feats = "";
   for (const auto& attr : attrs_) {
     feats += feats.empty() ? attr : ("," + attr);
@@ -845,10 +847,11 @@ const Array<String> LLVMTargetInfo::GetAllLLVMCpuFeatures() const {
 #else
       MCInfo->getAllProcessorFeatures();
 #endif
-  Array<String> cpu_features;
+  // TVM doesn't have an FFI friendly Set, so use a Map instead for now
+  Map<String, String> cpu_features;
   for (const auto& feat : llvm_features) {
     if (MCInfo->checkFeatures("+" + std::string(feat.Key))) {
-      cpu_features.push_back(feat.Key);
+      cpu_features.Set(feat.Key, "");
     }
   }
 
@@ -858,9 +861,7 @@ const Array<String> LLVMTargetInfo::GetAllLLVMCpuFeatures() const {
 const bool LLVMTargetInfo::TargetHasCPUFeature(const std::string& feature) const {
   // lookup features for `-mcpu`
   auto feats = GetAllLLVMCpuFeatures();
-  bool has_feature =
-      std::any_of(feats.begin(), feats.end(), [&](const auto& var) { return var == feature; });
-
+  bool has_feature = feats.find(feature) != feats.end();
   return has_feature;
 }
 
diff --git a/src/target/llvm/llvm_instance.h b/src/target/llvm/llvm_instance.h
@@ -156,6 +156,14 @@ class LLVMTargetInfo {
    */
   // NOLINTNEXTLINE(runtime/references)
   LLVMTargetInfo(LLVMInstance& scope, const std::string& target_str);
+  /*!
+   * \brief Constructs LLVMTargetInfo from `Target`
+   * \param scope LLVMInstance object
+   * \param target TVM JSON Target object for target "llvm"
+   */
+  // NOLINTNEXTLINE(runtime/references)
+  LLVMTargetInfo(LLVMInstance& instance, const TargetJSON& target);
+
   /*!
    * \brief Destroys LLVMTargetInfo object
    */
@@ -285,11 +293,12 @@ class LLVMTargetInfo {
 
   /*!
    * \brief Get all CPU features from target
-   * \return list with all valid cpu features
+   * \return Map with all valid cpu features as keys and empty string as value. The Map
+   *         is intended to be used as a Set, which TVM does not currently support.
    * \note The features are fetched from the LLVM backend using the target `-mtriple`
    *       and the `-mcpu` architecture, but also consider the `-mattr` attributes.
    */
-  const Array<String> GetAllLLVMCpuFeatures() const;
+  const Map<String, String> GetAllLLVMCpuFeatures() const;
 
   /*!
    * \brief Check the target if has a specific cpu feature
diff --git a/src/target/llvm/llvm_module.cc b/src/target/llvm/llvm_module.cc
@@ -545,12 +545,12 @@ TVM_REGISTER_GLOBAL("target.llvm_get_cpu_archlist")
     });
 
 TVM_REGISTER_GLOBAL("target.llvm_get_cpu_features")
-    .set_body_typed([](const Target& target) -> Array<String> {
+    .set_body_typed([](const Target& target) -> Map<String, String> {
       auto use_target = target.defined() ? target : Target::Current(false);
       // ignore non "llvm" target
       if (target.defined()) {
         if (target->kind->name != "llvm") {
-          return Array<String>{};
+          return {};
         }
       }
       auto llvm_instance = std::make_unique<LLVMInstance>();
@@ -570,8 +570,7 @@ TVM_REGISTER_GLOBAL("target.llvm_cpu_has_feature")
       auto llvm_instance = std::make_unique<LLVMInstance>();
       LLVMTargetInfo llvm_backend(*llvm_instance, use_target);
       auto cpu_features = llvm_backend.GetAllLLVMCpuFeatures();
-      bool has_feature = std::any_of(cpu_features.begin(), cpu_features.end(),
-                                     [&](auto& var) { return var == feature; });
+      bool has_feature = cpu_features.find(feature) != cpu_features.end();
       return has_feature;
     });
 
diff --git a/src/target/parsers/aprofile.cc b/src/target/parsers/aprofile.cc
@@ -24,9 +24,11 @@
 
 #include "aprofile.h"
 
+#include <memory>
 #include <string>
 
 #include "../../support/utils.h"
+#include "../llvm/llvm_instance.h"
 
 namespace tvm {
 namespace target {
@@ -52,33 +54,6 @@ double GetArchVersion(Optional<Array<String>> attr) {
   return GetArchVersion(attr.value());
 }
 
-static inline bool HasFlag(String attr, std::string flag) {
-  std::string attr_str = attr;
-  return attr_str.find(flag) != std::string::npos;
-}
-
-static inline bool HasFlag(Optional<String> attr, std::string flag) {
-  if (!attr) {
-    return false;
-  }
-  return HasFlag(attr.value(), flag);
-}
-
-static inline bool HasFlag(Optional<Array<String>> attr, std::string flag) {
-  if (!attr) {
-    return false;
-  }
-  Array<String> attr_array = attr.value();
-
-  auto matching_attr = std::find_if(attr_array.begin(), attr_array.end(),
-                                    [flag](String attr_str) { return HasFlag(attr_str, flag); });
-  return matching_attr != attr_array.end();
-}
-
-static bool HasFlag(Optional<String> mcpu, Optional<Array<String>> mattr, std::string flag) {
-  return HasFlag(mcpu, flag) || HasFlag(mattr, flag);
-}
-
 bool IsAArch32(Optional<String> mtriple, Optional<String> mcpu) {
   if (mtriple) {
     bool is_mprofile = mcpu && support::StartsWith(mcpu.value(), "cortex-m");
@@ -102,38 +77,25 @@ bool IsArch(TargetJSON attrs) {
 }
 
 static TargetFeatures GetFeatures(TargetJSON target) {
-  Optional<String> mcpu = Downcast<Optional<String>>(target.Get("mcpu"));
-  Optional<String> mtriple = Downcast<Optional<String>>(target.Get("mtriple"));
-  Optional<Array<String>> mattr = Downcast<Optional<Array<String>>>(target.Get("mattr"));
-
-  const double arch_version = GetArchVersion(mattr);
-
-  const bool is_aarch64 = IsAArch64(mtriple);
+  String kind = Downcast<String>(target.Get("kind"));
+  ICHECK_EQ(kind, "llvm") << "Expected target kind 'llvm', but got '" << kind << "'";
 
-  const bool simd_flag = HasFlag(mcpu, mattr, "+neon") || HasFlag(mcpu, mattr, "+simd");
-  const bool has_asimd = is_aarch64 || simd_flag;
-  const bool has_sve = HasFlag(mcpu, mattr, "+sve");
-
-  const bool i8mm_flag = HasFlag(mcpu, mattr, "+i8mm");
-  const bool i8mm_disable = HasFlag(mcpu, mattr, "+noi8mm");
-  const bool i8mm_default = arch_version >= 8.6;
-  const bool i8mm_support = arch_version >= 8.2 && arch_version <= 8.5;
-  const bool has_i8mm = (i8mm_default && !i8mm_disable) || (i8mm_support && i8mm_flag);
+  Optional<String> mtriple = Downcast<Optional<String>>(target.Get("mtriple"));
 
-  const bool dotprod_flag = HasFlag(mcpu, mattr, "+dotprod");
-  const bool dotprod_disable = HasFlag(mcpu, mattr, "+nodotprod");
-  const bool dotprod_default = arch_version >= 8.4;
-  const bool dotprod_support = arch_version >= 8.2 && arch_version <= 8.3;
-  const bool has_dotprod =
-      (dotprod_default && !dotprod_disable) || (dotprod_support && dotprod_flag);
+  auto llvm_instance = std::make_unique<codegen::LLVMInstance>();
+  codegen::LLVMTargetInfo llvm_target(*llvm_instance, target);
+  Map<String, String> features = llvm_target.GetAllLLVMCpuFeatures();
 
-  const bool fp16_flag = HasFlag(mcpu, mattr, "+fullfp16");
-  const bool fp16_support = arch_version >= 8.2;
-  const bool has_fp16_simd = fp16_support && (fp16_flag || has_sve);
+  auto has_feature = [features](const String& feature) {
+    return features.find(feature) != features.end();
+  };
 
-  return {{"is_aarch64", Bool(is_aarch64)},  {"has_asimd", Bool(has_asimd)},
-          {"has_sve", Bool(has_sve)},        {"has_dotprod", Bool(has_dotprod)},
-          {"has_matmul_i8", Bool(has_i8mm)}, {"has_fp16_simd", Bool(has_fp16_simd)}};
+  return {{"is_aarch64", Bool(IsAArch64(mtriple))},
+          {"has_asimd", Bool(has_feature("neon"))},
+          {"has_sve", Bool(has_feature("sve"))},
+          {"has_dotprod", Bool(has_feature("dotprod"))},
+          {"has_matmul_i8", Bool(has_feature("i8mm"))},
+          {"has_fp16_simd", Bool(has_feature("fullfp16"))}};
 }
 
 static Array<String> MergeKeys(Optional<Array<String>> existing_keys) {
diff --git a/tests/cpp/target/parsers/aprofile_test.cc b/tests/cpp/target/parsers/aprofile_test.cc
diff --git a/tests/python/relay/strategy/test_select_implementation.py b/tests/python/relay/strategy/test_select_implementation.py
