Baremetal multicore support

builder/sizes.go

@@ -490,7 +490,7 @@ func loadProgramSize(path string, packagePathMap map[string]string) (*programSiz
 				continue
 			}
 			if section.Type == elf.SHT_NOBITS {
-				if section.Name == ".stack" {
+				if strings.HasPrefix(section.Name, ".stack") {
 					// TinyGo emits stack sections on microcontroller using the
 					// ".stack" name.
 					// This is a bit ugly, but I don't think there is a way to

builder/sizes_test.go

@@ -42,9 +42,9 @@ func TestBinarySize(t *testing.T) {
 	// This is a small number of very diverse targets that we want to test.
 	tests := []sizeTest{
 		// microcontrollers
-		{"hifive1b", "examples/echo", 4560, 280, 0, 2268},
-		{"microbit", "examples/serial", 2924, 388, 8, 2272},
-		{"wioterminal", "examples/pininterrupt", 7383, 1489, 116, 6912},
+		{"hifive1b", "examples/echo", 4556, 280, 0, 2264},
+		{"microbit", "examples/serial", 2920, 388, 8, 2272},
+		{"wioterminal", "examples/pininterrupt", 7379, 1489, 116, 6912},
 
 		// TODO: also check wasm. Right now this is difficult, because
 		// wasm binaries are run through wasm-opt and therefore the

compileopts/config.go

@@ -110,6 +110,11 @@ func (c *Config) BuildTags() []string {
 		"math_big_pure_go",                           // to get math/big to work
 		"gc." + c.GC(), "scheduler." + c.Scheduler(), // used inside the runtime package
 		"serial." + c.Serial()}...) // used inside the machine package
+	switch c.Scheduler() {
+	case "threads", "cores":
+	default:
+		tags = append(tags, "tinygo.unicore")
+	}
 	for i := 1; i <= c.GoMinorVersion; i++ {
 		tags = append(tags, fmt.Sprintf("go1.%d", i))
 	}

compileopts/options.go

@@ -10,7 +10,7 @@ import (
 var (
 	validBuildModeOptions     = []string{"default", "c-shared", "wasi-legacy"}
 	validGCOptions            = []string{"none", "leaking", "conservative", "custom", "precise", "boehm"}
-	validSchedulerOptions     = []string{"none", "tasks", "asyncify", "threads"}
+	validSchedulerOptions     = []string{"none", "tasks", "asyncify", "threads", "cores"}
 	validSerialOptions        = []string{"none", "uart", "usb", "rtt"}
 	validPrintSizeOptions     = []string{"none", "short", "full", "html"}
 	validPanicStrategyOptions = []string{"print", "trap"}

compileopts/options_test.go

@@ -10,7 +10,7 @@ import (
 func TestVerifyOptions(t *testing.T) {
 
 	expectedGCError := errors.New(`invalid gc option 'incorrect': valid values are none, leaking, conservative, custom, precise, boehm`)
-	expectedSchedulerError := errors.New(`invalid scheduler option 'incorrect': valid values are none, tasks, asyncify, threads`)
+	expectedSchedulerError := errors.New(`invalid scheduler option 'incorrect': valid values are none, tasks, asyncify, threads, cores`)
 	expectedPrintSizeError := errors.New(`invalid size option 'incorrect': valid values are none, short, full, html`)
 	expectedPanicStrategyError := errors.New(`invalid panic option 'incorrect': valid values are print, trap`)
 

compiler/compiler.go

@@ -1856,15 +1856,7 @@ func (b *builder) createBuiltin(argTypes []types.Type, argValues []llvm.Value, c
 //
 // This is also where compiler intrinsics are implemented.
 func (b *builder) createFunctionCall(instr *ssa.CallCommon) (llvm.Value, error) {
-	var params []llvm.Value
-	for _, param := range instr.Args {
-		params = append(params, b.getValue(param, getPos(instr)))
-	}
-
-	// Try to call the function directly for trivially static calls.
-	var callee, context llvm.Value
-	var calleeType llvm.Type
-	exported := false
+	// See if this is an intrinsic function that is handled specially.
 	if fn := instr.StaticCallee(); fn != nil {
 		// Direct function call, either to a named or anonymous (directly
 		// applied) function call. If it is anonymous, it may be a closure.
@@ -1900,13 +1892,29 @@ func (b *builder) createFunctionCall(instr *ssa.CallCommon) (llvm.Value, error)
 			return llvm.ConstInt(b.ctx.Int8Type(), panicStrategy, false), nil
 		case name == "runtime/interrupt.New":
 			return b.createInterruptGlobal(instr)
+		case name == "runtime.exportedFuncPtr":
+			_, ptr := b.getFunction(instr.Args[0].(*ssa.Function))
+			return b.CreatePtrToInt(ptr, b.uintptrType, ""), nil
+		case name == "(*runtime/interrupt.Checkpoint).Save":
+			return b.createInterruptCheckpoint(instr.Args[0]), nil
 		case name == "internal/abi.FuncPCABI0":
 			retval := b.createDarwinFuncPCABI0Call(instr)
 			if !retval.IsNil() {
 				return retval, nil
 			}
 		}
+	}
 
+	var params []llvm.Value
+	for _, param := range instr.Args {
+		params = append(params, b.getValue(param, getPos(instr)))
+	}
+
+	// Try to call the function directly for trivially static calls.
+	var callee, context llvm.Value
+	var calleeType llvm.Type
+	exported := false
+	if fn := instr.StaticCallee(); fn != nil {
 		calleeType, callee = b.getFunction(fn)
 		info := b.getFunctionInfo(fn)
 		if callee.IsNil() {

compiler/defer.go

@@ -103,10 +103,11 @@ func (b *builder) createLandingPad() {
 	b.CreateBr(b.blockEntries[b.fn.Recover])
 }
 
-// createInvokeCheckpoint saves the function state at the given point, to
-// continue at the landing pad if a panic happened. This is implemented using a
-// setjmp-like construct.
-func (b *builder) createInvokeCheckpoint() {
+// Create a checkpoint (similar to setjmp). This emits inline assembly that
+// stores the current program counter inside the ptr address (actually
+// ptr+sizeof(ptr)) and then returns a boolean indicating whether this is the
+// normal flow (false) or we jumped here from somewhere else (true).
+func (b *builder) createCheckpoint(ptr llvm.Value) llvm.Value {
 	// Construct inline assembly equivalents of setjmp.
 	// The assembly works as follows:
 	//   * All registers (both callee-saved and caller saved) are clobbered
@@ -217,11 +218,19 @@ li a0, 0
 		// This case should have been handled by b.supportsRecover().
 		b.addError(b.fn.Pos(), "unknown architecture for defer: "+b.archFamily())
 	}
-	asmType := llvm.FunctionType(resultType, []llvm.Type{b.deferFrame.Type()}, false)
+	asmType := llvm.FunctionType(resultType, []llvm.Type{b.dataPtrType}, false)
 	asm := llvm.InlineAsm(asmType, asmString, constraints, false, false, 0, false)
-	result := b.CreateCall(asmType, asm, []llvm.Value{b.deferFrame}, "setjmp")
+	result := b.CreateCall(asmType, asm, []llvm.Value{ptr}, "setjmp")
 	result.AddCallSiteAttribute(-1, b.ctx.CreateEnumAttribute(llvm.AttributeKindID("returns_twice"), 0))
 	isZero := b.CreateICmp(llvm.IntEQ, result, llvm.ConstInt(resultType, 0, false), "setjmp.result")
+	return isZero
+}
+
+// createInvokeCheckpoint saves the function state at the given point, to
+// continue at the landing pad if a panic happened. This is implemented using a
+// setjmp-like construct.
+func (b *builder) createInvokeCheckpoint() {
+	isZero := b.createCheckpoint(b.deferFrame)
 	continueBB := b.insertBasicBlock("")
 	b.CreateCondBr(isZero, continueBB, b.landingpad)
 	b.SetInsertPointAtEnd(continueBB)

compiler/inlineasm.go

@@ -249,3 +249,15 @@ func (b *builder) emitCSROperation(call *ssa.CallCommon) (llvm.Value, error) {
 		return llvm.Value{}, b.makeError(call.Pos(), "unknown CSR operation: "+name)
 	}
 }
+
+// Implement runtime/interrupt.Checkpoint.Save. It needs to be implemented
+// directly at the call site. If it isn't implemented directly at the call site
+// (but instead through a function call), it might result in an overwritten
+// stack in the non-jump return case.
+func (b *builder) createInterruptCheckpoint(ptr ssa.Value) llvm.Value {
+	addr := b.getValue(ptr, ptr.Pos())
+	b.createNilCheck(ptr, addr, "deref")
+	stackPointer := b.readStackPointer()
+	b.CreateStore(stackPointer, addr)
+	return b.createCheckpoint(addr)
+}

src/device/riscv/start.S

@@ -3,8 +3,47 @@
 .type _start,@function
 
 _start:
+    // If we're on a multicore system, we need to wait for hart 0 to wake us up.
+#if TINYGO_CORES > 1
+    csrr a0, mhartid
+
+    // Hart 0 stack
+    bnez a0, 1f
+    la sp,      _stack_top
+
+1:
+    // Hart 1 stack
+    li a1, 1
+    bne a0, a1, 2f
+    la sp,      _stack1_top
+
+2:
+    // Hart 2 stack
+    #if TINYGO_CORES >= 3
+    li a1, 2
+    bne a0, a1, 3f
+    la sp,      _stack2_top
+    #endif
+
+3:
+    // Hart 3 stack
+    #if TINYGO_CORES >= 4
+    li a1, 3
+    bne a0, a1, 4f
+    la sp,      _stack3_top
+    #endif
+
+4:
+    // done
+
+#if TINYGO_CORES > 4
+#error only up to 4 cores are supported at the moment!
+#endif
+
+#else
     // Load the stack pointer.
     la sp,      _stack_top
+#endif
 
     // Load the globals pointer. The program will load pointers relative to this
     // register, so it must be set to the right value on startup.

src/internal/task/atomic-cooperative.go

@@ -1,4 +1,4 @@
-//go:build !scheduler.threads
+//go:build tinygo.unicore
 
 package task
 

src/internal/task/atomic-preemptive.go

@@ -1,4 +1,4 @@
-//go:build scheduler.threads
+//go:build !tinygo.unicore
 
 package task
 

src/internal/task/futex-cooperative.go

@@ -1,4 +1,4 @@
-//go:build !scheduler.threads
+//go:build tinygo.unicore
 
 package task
 

src/internal/task/futex-cores.go

@@ -0,0 +1,64 @@
+//go:build scheduler.cores
+
+package task
+
+import "runtime/interrupt"
+
+// A futex is a way for userspace to wait with the pointer as the key, and for
+// another thread to wake one or all waiting threads keyed on the same pointer.
+//
+// A futex does not change the underlying value, it only reads it before to prevent
+// lost wake-ups.
+type Futex struct {
+	Uint32
+
+	waiters Stack
+}
+
+// Atomically check for cmp to still be equal to the futex value and if so, go
+// to sleep. Return true if we were definitely awoken by a call to Wake or
+// WakeAll, and false if we can't be sure of that.
+func (f *Futex) Wait(cmp uint32) (awoken bool) {
+	mask := lockFutex()
+
+	if f.Uint32.Load() != cmp {
+		unlockFutex(mask)
+		return false
+	}
+
+	// Push the current goroutine onto the waiter stack.
+	f.waiters.Push(Current())
+
+	unlockFutex(mask)
+
+	// Pause until this task is awoken by Wake/WakeAll.
+	Pause()
+
+	// We were awoken by a call to Wake or WakeAll. There is no chance for
+	// spurious wakeups.
+	return true
+}
+
+// Wake a single waiter.
+func (f *Futex) Wake() {
+	mask := lockFutex()
+	if t := f.waiters.Pop(); t != nil {
+		scheduleTask(t)
+	}
+	unlockFutex(mask)
+}
+
+// Wake all waiters.
+func (f *Futex) WakeAll() {
+	mask := lockFutex()
+	for t := f.waiters.Pop(); t != nil; t = f.waiters.Pop() {
+		scheduleTask(t)
+	}
+	unlockFutex(mask)
+}
+
+//go:linkname lockFutex runtime.lockFutex
+func lockFutex() interrupt.State
+
+//go:linkname unlockFutex runtime.unlockFutex
+func unlockFutex(interrupt.State)

src/internal/task/mutex-cooperative.go

@@ -1,4 +1,4 @@
-//go:build !scheduler.threads
+//go:build tinygo.unicore
 
 package task
 

src/internal/task/mutex-preemptive.go

@@ -1,4 +1,4 @@
-//go:build scheduler.threads
+//go:build !tinygo.unicore
 
 package task
 

src/internal/task/pmutex-cooperative.go

@@ -1,4 +1,4 @@
-//go:build !scheduler.threads
+//go:build tinygo.unicore
 
 package task
 

src/internal/task/pmutex-preemptive.go

@@ -1,4 +1,4 @@
-//go:build scheduler.threads
+//go:build !tinygo.unicore
 
 package task