diff --git a/cl/_testrt/concat/in.go b/cl/_testrt/concat/in.go
index 811a0c8d..149bfbae 100644
--- a/cl/_testrt/concat/in.go
+++ b/cl/_testrt/concat/in.go
@@ -13,5 +13,5 @@ func concat(args ...string) (ret string) {
 
 func main() {
 	result := concat("Hello", " ", "World")
-	c.Fprintf(c.Stderr, c.Str("Hello %s\n"), c.AllocaCStr(result))
+	c.Fprintf(c.Stderr, c.Str("Hi, %s\n"), c.AllocaCStr(result))
 }
diff --git a/cl/_testrt/concat/out.ll b/cl/_testrt/concat/out.ll
index 06e49967..764decd6 100644
--- a/cl/_testrt/concat/out.ll
+++ b/cl/_testrt/concat/out.ll
@@ -10,30 +10,38 @@ source_filename = "main"
 @2 = private unnamed_addr constant [2 x i8] c" \00", align 1
 @3 = private unnamed_addr constant [6 x i8] c"World\00", align 1
 @__stderrp = external global ptr
-@4 = private unnamed_addr constant [10 x i8] c"Hello %s\0A\00", align 1
+@4 = private unnamed_addr constant [8 x i8] c"Hi, %s\0A\00", align 1
 
 define %"github.com/goplus/llgo/internal/runtime.String" @main.concat(%"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 = call %"github.com/goplus/llgo/internal/runtime.String" @"github.com/goplus/llgo/internal/runtime.NewString"(ptr @0, i64 0)
+  %3 = extractvalue %"github.com/goplus/llgo/internal/runtime.String" %2, 0
+  %4 = extractvalue %"github.com/goplus/llgo/internal/runtime.String" %2, 1
   br label %_llgo_1
 
 _llgo_1:                                          ; preds = %_llgo_2, %_llgo_0
-  %2 = phi %"github.com/goplus/llgo/internal/runtime.String" [ %10, %_llgo_0 ], [ %9, %_llgo_2 ]
-  %3 = phi i64 [ -1, %_llgo_0 ], [ %4, %_llgo_2 ]
-  %4 = add i64 %3, 1
-  %5 = icmp slt i64 %4, %1
-  br i1 %5, label %_llgo_2, label %_llgo_3
+  %5 = phi ptr [ %3, %_llgo_0 ], [ %14, %_llgo_2 ]
+  %6 = phi i64 [ %4, %_llgo_0 ], [ %15, %_llgo_2 ]
+  %7 = phi i64 [ -1, %_llgo_0 ], [ %8, %_llgo_2 ]
+  %mrv = insertvalue %"github.com/goplus/llgo/internal/runtime.String" poison, ptr %5, 0
+  %mrv1 = insertvalue %"github.com/goplus/llgo/internal/runtime.String" %mrv, i64 %6, 1
+  ret %"github.com/goplus/llgo/internal/runtime.String" %mrv1
+  %8 = add i64 %7, 1
+  %9 = icmp slt i64 %8, %1
+  br i1 %9, label %_llgo_2, label %_llgo_3
 
 _llgo_2:                                          ; preds = %_llgo_1
-  %6 = call ptr @"github.com/goplus/llgo/internal/runtime.SliceData"(%"github.com/goplus/llgo/internal/runtime.Slice" %0)
-  %7 = getelementptr inbounds %"github.com/goplus/llgo/internal/runtime.String", ptr %6, i64 %4
-  %8 = load %"github.com/goplus/llgo/internal/runtime.String", ptr %7, align 8
-  %9 = call %"github.com/goplus/llgo/internal/runtime.String" @"github.com/goplus/llgo/internal/runtime.StringCat"(%"github.com/goplus/llgo/internal/runtime.String" %2, %"github.com/goplus/llgo/internal/runtime.String" %8)
+  %10 = call ptr @"github.com/goplus/llgo/internal/runtime.SliceData"(%"github.com/goplus/llgo/internal/runtime.Slice" %0)
+  %11 = getelementptr inbounds %"github.com/goplus/llgo/internal/runtime.String", ptr %10, i64 %8
+  %12 = load %"github.com/goplus/llgo/internal/runtime.String", ptr %11, align 8
+  %13 = call %"github.com/goplus/llgo/internal/runtime.String" @"github.com/goplus/llgo/internal/runtime.StringCat"(void <badref>, %"github.com/goplus/llgo/internal/runtime.String" %12)
+  %14 = extractvalue %"github.com/goplus/llgo/internal/runtime.String" %13, 0
+  %15 = extractvalue %"github.com/goplus/llgo/internal/runtime.String" %13, 1
   br label %_llgo_1
 
 _llgo_3:                                          ; preds = %_llgo_1
-  ret %"github.com/goplus/llgo/internal/runtime.String" %2
-  %10 = call %"github.com/goplus/llgo/internal/runtime.String" @"github.com/goplus/llgo/internal/runtime.NewString"(ptr @0, i64 0)
+  ret void <badref>
 }
 
 define void @main.init() {
diff --git a/cl/compile.go b/cl/compile.go
index 8f67a0cb..3c7e2fe4 100644
--- a/cl/compile.go
+++ b/cl/compile.go
@@ -237,7 +237,8 @@ func (p *context) compileBlock(b llssa.Builder, block *ssa.BasicBlock, doInit bo
 		callRuntimeInit(b, pkg)
 		b.Call(pkg.FuncOf("main.init").Expr)
 	}
-	for _, instr := range block.Instrs {
+	instrs := p.compilePhis(b, block.Instrs)
+	for _, instr := range instrs {
 		p.compileInstr(b, instr)
 	}
 	return ret
@@ -327,6 +328,50 @@ func (p *context) allocaCStr(b llssa.Builder, args []ssa.Value) (ret llssa.Expr)
 	panic("allocaCStr(s string): invalid arguments")
 }
 
+func isPhi(i ssa.Instruction) bool {
+	_, ok := i.(*ssa.Phi)
+	return ok
+}
+
+func (p *context) compilePhis(b llssa.Builder, instrs []ssa.Instruction) []ssa.Instruction {
+	if ninstr := len(instrs); ninstr > 0 {
+		if isPhi(instrs[0]) {
+			n := 1
+			for n < ninstr && isPhi(instrs[n]) {
+				n++
+			}
+			rets := make([]llssa.Expr, n)
+			for i := 0; i < n; i++ {
+				iv := instrs[i].(*ssa.Phi)
+				rets[i] = p.compilePhi(b, iv)
+			}
+			for i := 0; i < n; i++ {
+				iv := instrs[i].(*ssa.Phi)
+				p.bvals[iv] = rets[i].Do(b)
+			}
+			return instrs[n:]
+		}
+	}
+	return instrs
+}
+
+func (p *context) compilePhi(b llssa.Builder, v *ssa.Phi) (ret llssa.Expr) {
+	phi := b.Phi(p.prog.Type(v.Type()))
+	ret = phi.Expr
+	p.phis = append(p.phis, func() {
+		preds := v.Block().Preds
+		bblks := make([]llssa.BasicBlock, len(preds))
+		for i, pred := range preds {
+			bblks[i] = p.fn.Block(pred.Index)
+		}
+		edges := v.Edges
+		phi.AddIncoming(b, bblks, func(i int) llssa.Expr {
+			return p.compileValue(b, edges[i])
+		})
+	})
+	return
+}
+
 func (p *context) compileInstrOrValue(b llssa.Builder, iv instrOrValue, asValue bool) (ret llssa.Expr) {
 	if asValue {
 		if v, ok := p.bvals[iv]; ok {
@@ -390,18 +435,6 @@ func (p *context) compileInstrOrValue(b llssa.Builder, iv instrOrValue, asValue
 	case *ssa.UnOp:
 		x := p.compileValue(b, v.X)
 		ret = b.UnOp(v.Op, x)
-	case *ssa.Phi:
-		phi := b.Phi(p.prog.Type(v.Type()))
-		ret = phi.Expr
-		p.phis = append(p.phis, func() {
-			vals := p.compileValues(b, v.Edges, 0)
-			preds := v.Block().Preds
-			bblks := make([]llssa.BasicBlock, len(preds))
-			for i, pred := range preds {
-				bblks[i] = p.fn.Block(pred.Index)
-			}
-			phi.AddIncoming(b, vals, bblks)
-		})
 	case *ssa.ChangeType:
 		t := v.Type()
 		x := p.compileValue(b, v.X)
diff --git a/cl/compile_test.go b/cl/compile_test.go
index 6efd7574..b9c3443f 100644
--- a/cl/compile_test.go
+++ b/cl/compile_test.go
@@ -29,7 +29,7 @@ func testCompile(t *testing.T, src, expected string) {
 }
 
 func TestFromTestrt(t *testing.T) {
-	cltest.FromDir(t, "concat", "./_testrt", true)
+	cltest.FromDir(t, "", "./_testrt", true)
 }
 
 func TestFromTestdata(t *testing.T) {
diff --git a/ssa/expr.go b/ssa/expr.go
index 495259fc..b69cb231 100644
--- a/ssa/expr.go
+++ b/ssa/expr.go
@@ -55,6 +55,7 @@ func (v Expr) Do(b Builder) Expr {
 		return vt.t.(delayExprTy)()
 	case vkPhisExpr:
 		e := vt.t.(*phisExprTy)
+		// TODO(xsw): to check CreateAggregateRet is correct or not
 		return Expr{b.impl.CreateAggregateRet(e.phis), e.Type}
 	}
 	return v
@@ -362,10 +363,10 @@ func llvmValues(vals []Expr) []llvm.Value {
 	return ret
 }
 
-func fieldValues(b llvm.Builder, vals []Expr, fldIdx int) []llvm.Value {
-	ret := make([]llvm.Value, len(vals))
-	for i, v := range vals {
-		ret[i] = llvm.CreateExtractValue(b, v.impl, fldIdx)
+func llvmDelayValues(f func(i int) Expr, n int) []llvm.Value {
+	ret := make([]llvm.Value, n)
+	for i := 0; i < n; i++ {
+		ret[i] = f(i).impl
 	}
 	return ret
 }
@@ -384,17 +385,30 @@ type Phi struct {
 }
 
 // AddIncoming adds incoming values to a phi node.
-func (p Phi) AddIncoming(b Builder, vals []Expr, bblks []BasicBlock) {
+func (p Phi) AddIncoming(b Builder, bblks []BasicBlock, f func(i int) Expr) {
 	bs := llvmBlocks(bblks)
 	if p.kind != vkPhisExpr { // normal phi node
-		vs := llvmValues(vals)
+		vs := llvmDelayValues(f, len(bblks))
 		p.impl.AddIncoming(vs, bs)
 		return
 	}
 	e := p.t.(*phisExprTy)
-	for i, phi := range e.phis {
-		flds := fieldValues(b.impl, vals, i)
-		phi.AddIncoming(flds, bs)
+	phis := e.phis
+	vals := make([][]llvm.Value, len(phis))
+	for iblk, blk := range bblks {
+		last := blk.impl.LastInstruction()
+		b.impl.SetInsertPointBefore(last)
+		impl := b.impl
+		val := f(iblk).impl
+		for i := range phis {
+			if iblk == 0 {
+				vals[i] = make([]llvm.Value, len(bblks))
+			}
+			vals[i][iblk] = llvm.CreateExtractValue(impl, val, i)
+		}
+	}
+	for i, phi := range phis {
+		phi.AddIncoming(vals[i], bs)
 	}
 }