diff --git a/cl/_testdata/print/in.go b/cl/_testdata/print/in.go
new file mode 100644
index 00000000..8ab1f11c
--- /dev/null
+++ b/cl/_testdata/print/in.go
@@ -0,0 +1,174 @@
+package main
+
+import (
+	"unsafe"
+
+	"github.com/goplus/llgo/internal/runtime/c"
+)
+
+func gwrite(b []byte) {
+	if len(b) == 0 {
+		return
+	}
+	for _, v := range b {
+		c.Printf(c.Str("%c"), v)
+	}
+}
+
+// func printfloat(v float64) {
+// 	switch {
+// 	case v != v:
+// 		printstring("NaN")
+// 		return
+// 	case v+v == v && v > 0:
+// 		printstring("+Inf")
+// 		return
+// 	case v+v == v && v < 0:
+// 		printstring("-Inf")
+// 		return
+// 	}
+
+// 	const n = 7 // digits printed
+// 	var buf [n + 7]byte
+// 	buf[0] = '+'
+// 	e := 0 // exp
+// 	if v == 0 {
+// 		if 1/v < 0 {
+// 			buf[0] = '-'
+// 		}
+// 	} else {
+// 		if v < 0 {
+// 			v = -v
+// 			buf[0] = '-'
+// 		}
+
+// 		// normalize
+// 		for v >= 10 {
+// 			e++
+// 			v /= 10
+// 		}
+// 		for v < 1 {
+// 			e--
+// 			v *= 10
+// 		}
+
+// 		// round
+// 		h := 5.0
+// 		for i := 0; i < n; i++ {
+// 			h /= 10
+// 		}
+// 		v += h
+// 		if v >= 10 {
+// 			e++
+// 			v /= 10
+// 		}
+// 	}
+
+// 	// format +d.dddd+edd
+// 	for i := 0; i < n; i++ {
+// 		s := int(v)
+// 		buf[i+2] = byte(s + '0')
+// 		v -= float64(s)
+// 		v *= 10
+// 	}
+// 	buf[1] = buf[2]
+// 	buf[2] = '.'
+
+// 	buf[n+2] = 'e'
+// 	buf[n+3] = '+'
+// 	if e < 0 {
+// 		e = -e
+// 		buf[n+3] = '-'
+// 	}
+
+// 	buf[n+4] = byte(e/100) + '0'
+// 	buf[n+5] = byte(e/10)%10 + '0'
+// 	buf[n+6] = byte(e%10) + '0'
+// 	gwrite(buf[:])
+// }
+
+func printuint(v uint64) {
+	var buf [100]byte
+	i := len(buf)
+	for i--; i > 0; i-- {
+		buf[i] = byte(v%10 + '0')
+		if v < 10 {
+			break
+		}
+		v /= 10
+	}
+	gwrite(buf[i:])
+}
+
+// func printint(v int64) {
+// 	if v < 0 {
+// 		printstring("-")
+// 		v = -v
+// 	}
+// 	printuint(uint64(v))
+// }
+
+var minhexdigits = 0
+
+func printhex(v uint64) {
+	const dig = "0123456789abcdef"
+	var buf [100]byte
+	i := len(buf)
+	for i--; i > 0; i-- {
+		buf[i] = dig[v%16]
+		if v < 16 && len(buf)-i >= minhexdigits {
+			break
+		}
+		v /= 16
+	}
+	i--
+	buf[i] = 'x'
+	i--
+	buf[i] = '0'
+	gwrite(buf[i:])
+}
+
+func printsp() {
+	printstring(" ")
+}
+
+func printnl() {
+	printstring("\n")
+}
+
+func printstring(s string) {
+	gwrite(bytes(s))
+}
+
+type slice struct {
+	array unsafe.Pointer
+	len   int
+	cap   int
+}
+
+type stringStruct struct {
+	str unsafe.Pointer
+	len int
+}
+
+func stringStructOf(sp *string) *stringStruct {
+	return (*stringStruct)(unsafe.Pointer(sp))
+}
+
+func bytes(s string) (ret []byte) {
+	rp := (*slice)(unsafe.Pointer(&ret))
+	sp := stringStructOf(&s)
+	rp.array = sp.str
+	rp.len = sp.len
+	rp.cap = sp.len
+	return
+}
+
+func main() {
+	printstring("llgo")
+	printnl()
+	printuint(1024)
+	printnl()
+	printhex(0x1234abcf)
+	printnl()
+}
diff --git a/cl/_testdata/print/out.ll b/cl/_testdata/print/out.ll
new file mode 100644
index 00000000..b8256d6e
--- /dev/null
+++ b/cl/_testdata/print/out.ll
@@ -0,0 +1,213 @@
+; ModuleID = 'main'
+source_filename = "main"
+
+%"github.com/goplus/llgo/internal/runtime.Slice" = type { ptr, i64, i64 }
+%"github.com/goplus/llgo/internal/runtime.String" = type { ptr, i64 }
+%main.stringStruct = type { ptr, i64 }
+%main.slice = type { ptr, i64, i64 }
+
+@"main.init$guard" = global ptr null
+@main.minhexdigits = global ptr null
+@0 = private unnamed_addr constant [3 x i8] c"%c\00", align 1
+@1 = private unnamed_addr constant [5 x i8] c"llgo\00", align 1
+@2 = private unnamed_addr constant [17 x i8] c"0123456789abcdef\00", align 1
+@3 = private unnamed_addr constant [2 x i8] c"\0A\00", align 1
+@4 = private unnamed_addr constant [2 x i8] c" \00", align 1
+
+define %"github.com/goplus/llgo/internal/runtime.Slice" @main.bytes(%"github.com/goplus/llgo/internal/runtime.String" %0) {
+_llgo_0:
+  %1 = call ptr @"github.com/goplus/llgo/internal/runtime.AllocZ"(i64 16)
+  store %"github.com/goplus/llgo/internal/runtime.String" %0, ptr %1, align 8
+  %2 = call ptr @"github.com/goplus/llgo/internal/runtime.AllocZ"(i64 24)
+  %3 = call ptr @main.stringStructOf(ptr %1)
+  %4 = getelementptr inbounds %main.stringStruct, ptr %3, i32 0, i32 0
+  %5 = load ptr, ptr %4, align 8
+  %6 = getelementptr inbounds %main.slice, ptr %2, i32 0, i32 0
+  store ptr %5, ptr %6, align 8
+  %7 = getelementptr inbounds %main.stringStruct, ptr %3, i32 0, i32 1
+  %8 = load i64, ptr %7, align 4
+  %9 = getelementptr inbounds %main.slice, ptr %2, i32 0, i32 1
+  store i64 %8, ptr %9, align 4
+  %10 = getelementptr inbounds %main.stringStruct, ptr %3, i32 0, i32 1
+  %11 = load i64, ptr %10, align 4
+  %12 = getelementptr inbounds %main.slice, ptr %2, i32 0, i32 2
+  store i64 %11, ptr %12, align 4
+  %13 = load %"github.com/goplus/llgo/internal/runtime.Slice", ptr %2, align 8
+  ret %"github.com/goplus/llgo/internal/runtime.Slice" %13
+}
+
+define void @main.gwrite(%"github.com/goplus/llgo/internal/runtime.Slice" %0) {
+_llgo_0:
+  %1 = call i64 @"github.com/goplus/llgo/internal/runtime.SliceLen"(%"github.com/goplus/llgo/internal/runtime.Slice" %0)
+  %2 = icmp eq i64 %1, 0
+  br i1 %2, label %_llgo_1, label %_llgo_2
+
+_llgo_1:                                          ; preds = %_llgo_0
+  ret void
+
+_llgo_2:                                          ; preds = %_llgo_0
+  %3 = call i64 @"github.com/goplus/llgo/internal/runtime.SliceLen"(%"github.com/goplus/llgo/internal/runtime.Slice" %0)
+  br label %_llgo_3
+
+_llgo_3:                                          ; preds = %_llgo_4, %_llgo_2
+  %4 = phi i64 [ -1, %_llgo_2 ], [ %5, %_llgo_4 ]
+  %5 = add i64 %4, 1
+  %6 = icmp slt i64 %5, %3
+  br i1 %6, label %_llgo_4, label %_llgo_5
+
+_llgo_4:                                          ; preds = %_llgo_3
+  %7 = call ptr @"github.com/goplus/llgo/internal/runtime.SliceData"(%"github.com/goplus/llgo/internal/runtime.Slice" %0)
+  %8 = getelementptr inbounds i8, ptr %7, i64 %5
+  %9 = load i8, ptr %8, align 1
+  %10 = call i32 (ptr, ...) @printf(ptr @0, i8 %9)
+  br label %_llgo_3
+
+_llgo_5:                                          ; preds = %_llgo_3
+  ret void
+}
+
+define void @main.init() {
+_llgo_0:
+  %0 = load i1, ptr @"main.init$guard", align 1
+  br i1 %0, label %_llgo_2, label %_llgo_1
+
+_llgo_1:                                          ; preds = %_llgo_0
+  store i1 true, ptr @"main.init$guard", align 1
+  store i64 0, ptr @main.minhexdigits, align 4
+  br label %_llgo_2
+
+_llgo_2:                                          ; preds = %_llgo_1, %_llgo_0
+  ret void
+}
+
+define void @main() {
+_llgo_0:
+  call void @"github.com/goplus/llgo/internal/runtime.init"()
+  call void @main.init()
+  %0 = call %"github.com/goplus/llgo/internal/runtime.String" @"github.com/goplus/llgo/internal/runtime.NewString"(ptr @1, i64 4)
+  call void @main.printstring(%"github.com/goplus/llgo/internal/runtime.String" %0)
+  call void @main.printnl()
+  call void @main.printuint(i64 1024)
+  call void @main.printnl()
+  call void @main.printhex(i64 305441743)
+  call void @main.printnl()
+  ret void
+}
+
+define void @main.printhex(i64 %0) {
+_llgo_0:
+  %1 = call ptr @"github.com/goplus/llgo/internal/runtime.AllocZ"(i64 100)
+  br label %_llgo_3
+
+_llgo_1:                                          ; preds = %_llgo_3
+  %2 = urem i64 %14, 16
+  %3 = call %"github.com/goplus/llgo/internal/runtime.String" @"github.com/goplus/llgo/internal/runtime.NewString"(ptr @2, i64 16)
+  %4 = call ptr @"github.com/goplus/llgo/internal/runtime.StringData"(%"github.com/goplus/llgo/internal/runtime.String" %3)
+  %5 = getelementptr inbounds i8, ptr %4, i64 %2
+  %6 = load i8, ptr %5, align 1
+  %7 = getelementptr inbounds i8, ptr %1, i64 %15
+  store i8 %6, ptr %7, align 1
+  %8 = icmp ult i64 %14, 16
+  br i1 %8, label %_llgo_5, label %_llgo_4
+
+_llgo_2:                                          ; preds = %_llgo_5, %_llgo_3
+  %9 = sub i64 %15, 1
+  %10 = getelementptr inbounds i8, ptr %1, i64 %9
+  store i8 120, ptr %10, align 1
+  %11 = sub i64 %9, 1
+  %12 = getelementptr inbounds i8, ptr %1, i64 %11
+  store i8 48, ptr %12, align 1
+  %13 = call %"github.com/goplus/llgo/internal/runtime.Slice" @"github.com/goplus/llgo/internal/runtime.NewSlice3"(ptr %1, i64 1, i64 100, i64 %11, i64 100, i64 100)
+  call void @main.gwrite(%"github.com/goplus/llgo/internal/runtime.Slice" %13)
+  ret void
+
+_llgo_3:                                          ; preds = %_llgo_4, %_llgo_0
+  %14 = phi i64 [ %0, %_llgo_0 ], [ %17, %_llgo_4 ]
+  %15 = phi i64 [ 99, %_llgo_0 ], [ %18, %_llgo_4 ]
+  %16 = icmp sgt i64 %15, 0
+  br i1 %16, label %_llgo_1, label %_llgo_2
+
+_llgo_4:                                          ; preds = %_llgo_5, %_llgo_1
+  %17 = udiv i64 %14, 16
+  %18 = sub i64 %15, 1
+  br label %_llgo_3
+
+_llgo_5:                                          ; preds = %_llgo_1
+  %19 = sub i64 100, %15
+  %20 = load i64, ptr @main.minhexdigits, align 4
+  %21 = icmp sge i64 %19, %20
+  br i1 %21, label %_llgo_2, label %_llgo_4
+}
+
+define void @main.printnl() {
+_llgo_0:
+  %0 = call %"github.com/goplus/llgo/internal/runtime.String" @"github.com/goplus/llgo/internal/runtime.NewString"(ptr @3, i64 1)
+  call void @main.printstring(%"github.com/goplus/llgo/internal/runtime.String" %0)
+  ret void
+}
+
+define void @main.printsp() {
+_llgo_0:
+  %0 = call %"github.com/goplus/llgo/internal/runtime.String" @"github.com/goplus/llgo/internal/runtime.NewString"(ptr @4, i64 1)
+  call void @main.printstring(%"github.com/goplus/llgo/internal/runtime.String" %0)
+  ret void
+}
+
+define void @main.printstring(%"github.com/goplus/llgo/internal/runtime.String" %0) {
+_llgo_0:
+  %1 = call %"github.com/goplus/llgo/internal/runtime.Slice" @main.bytes(%"github.com/goplus/llgo/internal/runtime.String" %0)
+  call void @main.gwrite(%"github.com/goplus/llgo/internal/runtime.Slice" %1)
+  ret void
+}
+
+define void @main.printuint(i64 %0) {
+_llgo_0:
+  %1 = call ptr @"github.com/goplus/llgo/internal/runtime.AllocZ"(i64 100)
+  br label %_llgo_3
+
+_llgo_1:                                          ; preds = %_llgo_3
+  %2 = urem i64 %8, 10
+  %3 = add i64 %2, 48
+  %4 = trunc i64 %3 to i8
+  %5 = getelementptr inbounds i8, ptr %1, i64 %9
+  store i8 %4, ptr %5, align 1
+  %6 = icmp ult i64 %8, 10
+  br i1 %6, label %_llgo_2, label %_llgo_4
+
+_llgo_2:                                          ; preds = %_llgo_3, %_llgo_1
+  %7 = call %"github.com/goplus/llgo/internal/runtime.Slice" @"github.com/goplus/llgo/internal/runtime.NewSlice3"(ptr %1, i64 1, i64 100, i64 %9, i64 100, i64 100)
+  call void @main.gwrite(%"github.com/goplus/llgo/internal/runtime.Slice" %7)
+  ret void
+
+_llgo_3:                                          ; preds = %_llgo_4, %_llgo_0
+  %8 = phi i64 [ %0, %_llgo_0 ], [ %11, %_llgo_4 ]
+  %9 = phi i64 [ 99, %_llgo_0 ], [ %12, %_llgo_4 ]
+  %10 = icmp sgt i64 %9, 0
+  br i1 %10, label %_llgo_1, label %_llgo_2
+
+_llgo_4:                                          ; preds = %_llgo_1
+  %11 = udiv i64 %8, 10
+  %12 = sub i64 %9, 1
+  br label %_llgo_3
+}
+
+define ptr @main.stringStructOf(ptr %0) {
+_llgo_0:
+  ret ptr %0
+}
+
+declare ptr @"github.com/goplus/llgo/internal/runtime.AllocZ"(i64)
+
+declare i64 @"github.com/goplus/llgo/internal/runtime.SliceLen"(%"github.com/goplus/llgo/internal/runtime.Slice")
+
+declare ptr @"github.com/goplus/llgo/internal/runtime.SliceData"(%"github.com/goplus/llgo/internal/runtime.Slice")
+
+declare i32 @printf(ptr, ...)
+
+declare void @"github.com/goplus/llgo/internal/runtime.init"()
+
+declare %"github.com/goplus/llgo/internal/runtime.String" @"github.com/goplus/llgo/internal/runtime.NewString"(ptr, i64)
+
+declare ptr @"github.com/goplus/llgo/internal/runtime.StringData"(%"github.com/goplus/llgo/internal/runtime.String")
+
+declare %"github.com/goplus/llgo/internal/runtime.Slice" @"github.com/goplus/llgo/internal/runtime.NewSlice3"(ptr, i64, i64, i64, i64, i64)
diff --git a/cl/compile.go b/cl/compile.go
index bee1a4f4..954412c9 100644
--- a/cl/compile.go
+++ b/cl/compile.go
@@ -242,8 +242,12 @@ func (p *context) compileFuncDecl(pkg llssa.Package, pkgTypes *types.Package, f
 			}
 			b := fn.NewBuilder()
 			p.bvals = make(map[ssa.Value]llssa.Expr)
+			off := make([]int, len(f.Blocks))
 			for i, block := range f.Blocks {
-				p.compileBlock(b, block, i == 0 && name == "main")
+				off[i] = p.compilePhis(b, block)
+			}
+			for i, block := range f.Blocks {
+				p.compileBlock(b, block, off[i], i == 0 && name == "main")
 			}
 			for _, phi := range p.phis {
 				phi()
@@ -281,7 +285,7 @@ func (p *context) funcOf(fn *ssa.Function) (ret llssa.Function, ftype int) {
 	return
 }
 
-func (p *context) compileBlock(b llssa.Builder, block *ssa.BasicBlock, doInit bool) llssa.BasicBlock {
+func (p *context) compileBlock(b llssa.Builder, block *ssa.BasicBlock, n int, doInit bool) llssa.BasicBlock {
 	ret := p.fn.Block(block.Index)
 	b.SetBlock(ret)
 	if doInit {
@@ -289,8 +293,7 @@ func (p *context) compileBlock(b llssa.Builder, block *ssa.BasicBlock, doInit bo
 		callRuntimeInit(b, pkg)
 		b.Call(pkg.FuncOf("main.init").Expr)
 	}
-	instrs := p.compilePhis(b, block.Instrs)
-	for _, instr := range instrs {
+	for _, instr := range block.Instrs[n:] {
 		p.compileInstr(b, instr)
 	}
 	return ret
@@ -386,26 +389,28 @@ func isPhi(i ssa.Instruction) bool {
 	return ok
 }
 
-func (p *context) compilePhis(b llssa.Builder, instrs []ssa.Instruction) []ssa.Instruction {
-	if ninstr := len(instrs); ninstr > 0 {
-		if isPhi(instrs[0]) {
+func (p *context) compilePhis(b llssa.Builder, block *ssa.BasicBlock) int {
+	ret := p.fn.Block(block.Index)
+	b.SetBlock(ret)
+	if ninstr := len(block.Instrs); ninstr > 0 {
+		if isPhi(block.Instrs[0]) {
 			n := 1
-			for n < ninstr && isPhi(instrs[n]) {
+			for n < ninstr && isPhi(block.Instrs[n]) {
 				n++
 			}
 			rets := make([]llssa.Expr, n)
 			for i := 0; i < n; i++ {
-				iv := instrs[i].(*ssa.Phi)
+				iv := block.Instrs[i].(*ssa.Phi)
 				rets[i] = p.compilePhi(b, iv)
 			}
 			for i := 0; i < n; i++ {
-				iv := instrs[i].(*ssa.Phi)
+				iv := block.Instrs[i].(*ssa.Phi)
 				p.bvals[iv] = rets[i].Do(b)
 			}
-			return instrs[n:]
+			return n
 		}
 	}
-	return instrs
+	return 0
 }
 
 func (p *context) compilePhi(b llssa.Builder, v *ssa.Phi) (ret llssa.Expr) {