Improve FMA detection.

Don't clone on platforms that always support it, and more accurately check CPU features.
JuliaLang · Nov 16, 2021 · 36a7f6f · 36a7f6f
1 parent cdb2c3c
commit 36a7f6f
Show file tree

Hide file tree

Showing 2 changed files with 41 additions and 12 deletions.
diff --git a/src/llvm-cpufeatures.cpp b/src/llvm-cpufeatures.cpp
@@ -7,7 +7,8 @@
 // The following intrinsics are supported:
 // - julia.cpu.have_fma: returns 1 if the platform supports hardware-accelerated FMA
 //
-// XXX: can / do we want to make this a codegen pass to enable querying TargetPassConfig?
+// XXX: can / do we want to make this a codegen pass to enable querying TargetPassConfig
+//      instead of using the global target machine?
 
 #include "llvm-version.h"
 
@@ -26,28 +27,48 @@ using namespace llvm;
 
 extern TargetMachine *jl_TargetMachine;
 
-namespace {
+// whether this platform unconditionally (i.e. without needing multiversioning) supports FMA
+Optional<bool> always_have_fma() {
+#ifdef _CPU_AARCH64_
+    return true;
+#else
+    return {};
+#endif
+}
 
-static void lowerHaveFMA(Function &F, Instruction *I) {
-    Triple TheTriple = Triple(jl_TargetMachine->getTargetTriple());
+bool have_fma(Function &F) {
+    auto unconditional = always_have_fma();
+    if (unconditional.hasValue())
+        return unconditional.getValue();
 
-    Attribute CPUAttr = F.getFnAttribute("target-cpu");
     Attribute FSAttr = F.getFnAttribute("target-features");
-
-    StringRef CPU =
-        CPUAttr.isValid() ? CPUAttr.getValueAsString() : jl_TargetMachine->getTargetCPU();
     StringRef FS =
         FSAttr.isValid() ? FSAttr.getValueAsString() : jl_TargetMachine->getTargetFeatureString();
 
-    if (TheTriple.getArch() == Triple::x86_64 && FS.find("+fma") != StringRef::npos)
+    SmallVector<StringRef, 6> Features;
+    FS.split(Features, ',');
+    for (StringRef Feature : Features)
+#if defined _CPU_ARM_
+      if (Feature == "+vfp4")
+        return true;
+#else
+      if (Feature == "+fma" || Feature == "+fma4")
+        return true;
+#endif
+
+    return false;
+}
+
+void lowerHaveFMA(Function &F, Instruction *I) {
+    if (have_fma(F))
         I->replaceAllUsesWith(ConstantInt::get(I->getType(), 1));
     else
         I->replaceAllUsesWith(ConstantInt::get(I->getType(), 0));
 
     return;
 }
 
-static bool lowerCPUFeatures(Module &M)
+bool lowerCPUFeatures(Module &M)
 {
     SmallVector<Instruction*,6> Materialized;
     if (auto have_fma = M.getFunction("julia.cpu.have_fma")) {
@@ -68,7 +89,6 @@ static bool lowerCPUFeatures(Module &M)
         return false;
     }
 }
-}
 
 struct CPUFeatures : PassInfoMixin<CPUFeatures> {
     PreservedAnalyses run(Module &M, ModuleAnalysisManager &AM);

diff --git a/src/llvm-multiversioning.cpp b/src/llvm-multiversioning.cpp
@@ -43,6 +43,8 @@ using namespace llvm;
 extern std::pair<MDNode*,MDNode*> tbaa_make_child(const char *name, MDNode *parent=nullptr,
                                                   bool isConstant=false);
 
+extern Optional<bool> always_have_fma();
+
 namespace {
 
 // These are valid detail cloning conditions in the target flags.
@@ -470,7 +472,14 @@ uint32_t CloneCtx::collect_func_info(Function &F)
                         flag |= JL_TARGET_CLONE_MATH;
                     }
                     else if (name.startswith("julia.cpu.")) {
-                        flag |= JL_TARGET_CLONE_CPU;
+                        if (name == "julia.cpu.have_fma") {
+                            // for some platforms we know they always do (or don't) support
+                            // FMA. in those cases we don't need to clone the function.
+                            if (!always_have_fma().hasValue())
+                                flag |= JL_TARGET_CLONE_CPU;
+                        } else {
+                            flag |= JL_TARGET_CLONE_CPU;
+                        }
                     }
                 }
             }