diff --git a/compiler/compiler.go b/compiler/compiler.go
index db5d442c..a31c4d7d 100644
--- a/compiler/compiler.go
+++ b/compiler/compiler.go
@@ -195,7 +195,7 @@ func (c *Compiler) Compile(mainPath string) []error {
 		},
 		ShouldOverlay: func(path string) bool {
 			switch path {
-			case "machine", "os", "reflect", "runtime", "sync":
+			case "machine", "os", "reflect", "runtime", "runtime/volatile", "sync":
 				return true
 			default:
 				if strings.HasPrefix(path, "device/") || strings.HasPrefix(path, "examples/") {
@@ -1265,17 +1265,22 @@ func (c *Compiler) parseCall(frame *Frame, instr *ssa.CallCommon) (llvm.Value, e
 
 	// Try to call the function directly for trivially static calls.
 	if fn := instr.StaticCallee(); fn != nil {
-		switch fn.RelString(nil) {
-		case "device/arm.ReadRegister":
+		name := fn.RelString(nil)
+		switch {
+		case name == "device/arm.ReadRegister":
 			return c.emitReadRegister(instr.Args)
-		case "device/arm.Asm", "device/avr.Asm":
+		case name == "device/arm.Asm" || name == "device/avr.Asm":
 			return c.emitAsm(instr.Args)
-		case "device/arm.AsmFull", "device/avr.AsmFull":
+		case name == "device/arm.AsmFull" || name == "device/avr.AsmFull":
 			return c.emitAsmFull(frame, instr)
-		case "device/arm.SVCall0", "device/arm.SVCall1", "device/arm.SVCall2", "device/arm.SVCall3", "device/arm.SVCall4":
+		case strings.HasPrefix(name, "device/arm.SVCall"):
 			return c.emitSVCall(frame, instr.Args)
-		case "syscall.Syscall", "syscall.Syscall6", "syscall.Syscall9":
+		case strings.HasPrefix(name, "syscall.Syscall"):
 			return c.emitSyscall(frame, instr)
+		case strings.HasPrefix(name, "runtime/volatile.Load"):
+			return c.emitVolatileLoad(frame, instr)
+		case strings.HasPrefix(name, "runtime/volatile.Store"):
+			return c.emitVolatileStore(frame, instr)
 		}
 
 		targetFunc := c.ir.GetFunction(fn)
diff --git a/compiler/volatile.go b/compiler/volatile.go
new file mode 100644
index 00000000..7ee5b168
--- /dev/null
+++ b/compiler/volatile.go
@@ -0,0 +1,26 @@
+package compiler
+
+// This file implements volatile loads/stores in runtime/volatile.LoadT and
+// runtime/volatile.StoreT as compiler builtins.
+
+import (
+	"golang.org/x/tools/go/ssa"
+	"tinygo.org/x/go-llvm"
+)
+
+func (c *Compiler) emitVolatileLoad(frame *Frame, instr *ssa.CallCommon) (llvm.Value, error) {
+	addr := c.getValue(frame, instr.Args[0])
+	c.emitNilCheck(frame, addr, "deref")
+	val := c.builder.CreateLoad(addr, "")
+	val.SetVolatile(true)
+	return val, nil
+}
+
+func (c *Compiler) emitVolatileStore(frame *Frame, instr *ssa.CallCommon) (llvm.Value, error) {
+	addr := c.getValue(frame, instr.Args[0])
+	val := c.getValue(frame, instr.Args[1])
+	c.emitNilCheck(frame, addr, "deref")
+	store := c.builder.CreateStore(val, addr)
+	store.SetVolatile(true)
+	return llvm.Value{}, nil
+}
diff --git a/src/runtime/volatile/volatile.go b/src/runtime/volatile/volatile.go
new file mode 100644
index 00000000..47262f34
--- /dev/null
+++ b/src/runtime/volatile/volatile.go
@@ -0,0 +1,34 @@
+// Package volatile provides definitions for volatile loads and stores. These
+// are implemented as compiler builtins.
+//
+// The load operations load a volatile value. The store operations store to a
+// volatile value. The compiler will emit exactly one load or store operation
+// when possible and will not reorder volatile operations. However, the compiler
+// may move other operations across load/store operations, so make sure that all
+// relevant loads/stores are done in a volatile way if this is a problem.
+//
+// These loads and stores are commonly used to read/write values from memory
+// mapped peripheral devices. They do not provide atomicity, use the sync/atomic
+// package for that.
+//
+// For more details: https://llvm.org/docs/LangRef.html#volatile-memory-accesses
+// and https://blog.regehr.org/archives/28.
+package volatile
+
+// LoadUint8 loads the volatile value *addr.
+func LoadUint8(addr *uint8) (val uint8)
+
+// LoadUint16 loads the volatile value *addr.
+func LoadUint16(addr *uint16) (val uint16)
+
+// LoadUint32 loads the volatile value *addr.
+func LoadUint32(addr *uint32) (val uint32)
+
+// StoreUint8 stores val to the volatile value *addr.
+func StoreUint8(addr *uint8, val uint8)
+
+// StoreUint16 stores val to the volatile value *addr.
+func StoreUint16(addr *uint16, val uint16)
+
+// StoreUint32 stores val to the volatile value *addr.
+func StoreUint32(addr *uint32, val uint32)