avr: support ThinLTO

ThinLTO results in a small code size reduction, which is nice
(especially on these very small chips). It also brings us one step
closer to using ThinLTO everywhere.

compileopts/config.go

@@ -200,11 +200,6 @@ func (c *Config) UseThinLTO() bool {
 		// wasm-ld doesn't seem to support ThinLTO yet.
 		return false
 	}
-	if parts[0] == "avr" {
-		// These use external (GNU) linkers which might perhaps support ThinLTO
-		// through a plugin, but it's too much hassle to set up.
-		return false
-	}
 	// Other architectures support ThinLTO.
 	return true
 }

compiler/llvmutil/llvm.go

@@ -196,21 +196,31 @@ func SplitBasicBlock(builder llvm.Builder, afterInst llvm.Value, insertAfter llv
 	return newBlock
 }
 
-// Append the given values to the llvm.used array. The values can be any pointer
+// Append the given values to a global array like llvm.used. The global might
-// type, they will be bitcast to i8*.
+// not exist yet. The values can be any pointer type, they will be cast to i8*.
-func AppendToUsedGlobals(mod llvm.Module, values ...llvm.Value) {
+func AppendToGlobal(mod llvm.Module, globalName string, values ...llvm.Value) {
-	if !mod.NamedGlobal("llvm.used").IsNil() {
+	// Read the existing values in the llvm.used array (if it exists).
-		// Sanity check. TODO: we don't emit such a global at the moment, but
+	var usedValues []llvm.Value
-		// when we do we should append to it instead.
+	if used := mod.NamedGlobal(globalName); !used.IsNil() {
-		panic("todo: append to existing llvm.used")
+		builder := mod.Context().NewBuilder()
-	}
+		defer builder.Dispose()
+		usedInitializer := used.Initializer()
+		num := usedInitializer.Type().ArrayLength()
+		for i := 0; i < num; i++ {
+			usedValues = append(usedValues, builder.CreateExtractValue(usedInitializer, i, ""))
+		}
+		used.EraseFromParentAsGlobal()
+	}
+
+	// Add the new values.
 	i8ptrType := llvm.PointerType(mod.Context().Int8Type(), 0)
-	var castValues []llvm.Value
 	for _, value := range values {
-		castValues = append(castValues, llvm.ConstBitCast(value, i8ptrType))
+		usedValues = append(usedValues, llvm.ConstPointerCast(value, i8ptrType))
 	}
-	usedInitializer := llvm.ConstArray(i8ptrType, castValues)
+
-	used := llvm.AddGlobal(mod, usedInitializer.Type(), "llvm.used")
+	// Create a new array (with the old and new values).
+	usedInitializer := llvm.ConstArray(i8ptrType, usedValues)
+	used := llvm.AddGlobal(mod, usedInitializer.Type(), globalName)
 	used.SetInitializer(usedInitializer)
 	used.SetLinkage(llvm.AppendingLinkage)
 }

compiler/symbol.go

@@ -150,6 +150,19 @@ func (c *compilerContext) getFunction(fn *ssa.Function) (llvm.Type, llvm.Value)
 		// that the only thing we'll do is read the pointer.
 		llvmFn.AddAttributeAtIndex(1, c.ctx.CreateEnumAttribute(llvm.AttributeKindID("nocapture"), 0))
 		llvmFn.AddAttributeAtIndex(1, c.ctx.CreateEnumAttribute(llvm.AttributeKindID("readonly"), 0))
+	case "__mulsi3", "__divmodsi4", "__udivmodsi4":
+		if strings.Split(c.Triple, "-")[0] == "avr" {
+			// These functions are compiler-rt/libgcc functions that are
+			// currently implemented in Go. Assembly versions should appear in
+			// LLVM 16 hopefully. Until then, they need to be made available to
+			// the linker and the best way to do that is llvm.compiler.used.
+			// I considered adding a pragma for this, but the LLVM language
+			// reference explicitly says that this feature should not be exposed
+			// to source languages:
+			// > This is a rare construct that should only be used in rare
+			// > circumstances, and should not be exposed to source languages.
+			llvmutil.AppendToGlobal(c.mod, "llvm.compiler.used", llvmFn)
+		}
 	}
 
 	// External/exported functions may not retain pointer values.

src/runtime/arch_avr.go

@@ -40,6 +40,8 @@ func procUnpin() {
 
 // The following functions are workarounds for things missing in compiler-rt.
 // They will likely need special assembly implementations.
+// They are treated specially: they're added to @llvm.compiler.used so that the
+// linker won't eliminate them.
 
 //export __mulsi3
 func __mulsi3(a, b uint32) uint32 {

targets/avr.ld

@@ -5,7 +5,7 @@ MEMORY
     RAM (xrw)       : ORIGIN = 0x800000 + __ram_start, LENGTH = __ram_size
 }
 
-ENTRY(__vector_RESET)
+ENTRY(main)
 
 SECTIONS
 {

transform/stacksize.go

@@ -53,7 +53,7 @@ func CreateStackSizeLoads(mod llvm.Module, config *compileopts.Config) []string
 	stackSizesGlobal.SetInitializer(llvm.ConstArray(functions[0].Type(), defaultStackSizes))
 
 	// Add all relevant values to llvm.used (for LTO).
-	llvmutil.AppendToUsedGlobals(mod, append([]llvm.Value{stackSizesGlobal}, functionValues...)...)
+	llvmutil.AppendToGlobal(mod, "llvm.used", append([]llvm.Value{stackSizesGlobal}, functionValues...)...)
 
 	// Replace the calls with loads from the new global with stack sizes.
 	irbuilder := ctx.NewBuilder()