printer: use tabwriter to align instructions (#8)

README.md

@@ -140,6 +140,7 @@ TEXT ·Sum(SB), NOSPLIT, $0-32
 	MOVQ xs_base(FP), AX
 	MOVQ xs_len+8(FP), CX
 	XORQ DX, DX
+
 loop:
 	CMPQ CX, $0x00
 	JE   done
@@ -147,6 +148,7 @@ loop:
 	ADDQ $0x08, AX
 	DECQ CX
 	JMP  loop
+
 done:
 	MOVQ DX, ret+24(FP)
 	RET

examples/dot/dot.s

@@ -13,6 +13,7 @@ TEXT ·Dot(SB), NOSPLIT, $0-52
 	VXORPS Y3, Y3, Y3
 	VXORPS Y4, Y4, Y4
 	VXORPS Y5, Y5, Y5
+
 blockloop:
 	CMPQ        DX, $0x00000030
 	JL          tail
@@ -32,8 +33,10 @@ blockloop:
 	ADDQ        $0x000000c0, CX
 	SUBQ        $0x00000030, DX
 	JMP         blockloop
+
 tail:
 	VXORPS X12, X12, X12
+
 tailloop:
 	CMPQ        DX, $0x00000000
 	JE          reduce
@@ -43,6 +46,7 @@ tailloop:
 	ADDQ        $0x00000004, CX
 	DECQ        DX
 	JMP         tailloop
+
 reduce:
 	VADDPS       Y0, Y1, Y0
 	VADDPS       Y0, Y2, Y0

examples/fnv1a/fnv1a.s

@@ -8,6 +8,7 @@ TEXT ·Hash64(SB), NOSPLIT, $0-32
 	MOVQ data_len+8(FP), BX
 	MOVQ $0xcbf29ce484222325, AX
 	MOVQ $0x00000100000001b3, BP
+
 loop:
 	CMPQ    BX, $0x00
 	JE      done
@@ -17,6 +18,7 @@ loop:
 	INCQ    CX
 	DECQ    BX
 	JMP     loop
+
 done:
 	MOVQ AX, ret+24(FP)
 	RET

examples/stadtx/stadtx.s

@@ -31,6 +31,7 @@ TEXT ·Hash(SB), NOSPLIT, $0-40
 	JE    shortCore2
 	CMPQ  SI, $0x00000003
 	JE    shortCore3
+
 shortCore3:
 	MOVQ  (CX), SI
 	MOVQ  $0x9c1b8e1e9628323f, DI
@@ -42,6 +43,7 @@ shortCore3:
 	ADDQ  BX, BP
 	ADDQ  $0x00000008, CX
 	SUBQ  $0x00000008, DX
+
 shortCore2:
 	MOVQ  (CX), SI
 	MOVQ  $0x9c1b8e1e9628323f, DI
@@ -53,6 +55,7 @@ shortCore2:
 	ADDQ  BX, BP
 	ADDQ  $0x00000008, CX
 	SUBQ  $0x00000008, DX
+
 shortCore1:
 	MOVQ  (CX), SI
 	MOVQ  $0x9c1b8e1e9628323f, DI
@@ -64,6 +67,7 @@ shortCore1:
 	ADDQ  BX, BP
 	ADDQ  $0x00000008, CX
 	SUBQ  $0x00000008, DX
+
 shortCore0:
 	CMPQ DX, $0x00000000
 	JE   shortTail0
@@ -81,36 +85,45 @@ shortCore0:
 	JE   shortTail6
 	CMPQ DX, $0x00000007
 	JE   shortTail7
+
 shortTail7:
 	MOVBQZX 6(CX), SI
 	SHLQ    $0x20, SI
 	ADDQ    SI, BX
+
 shortTail6:
 	MOVBQZX 5(CX), SI
 	SHLQ    $0x30, SI
 	ADDQ    SI, BP
+
 shortTail5:
 	MOVBQZX 4(CX), SI
 	SHLQ    $0x10, SI
 	ADDQ    SI, BX
+
 shortTail4:
 	MOVLQZX (CX), SI
 	ADDQ    SI, BP
 	JMP     shortAfter
+
 shortTail3:
 	MOVBQZX 2(CX), SI
 	SHLQ    $0x30, SI
 	ADDQ    SI, BX
+
 shortTail2:
 	MOVWQZX (CX), SI
 	ADDQ    SI, BP
 	JMP     shortAfter
+
 shortTail1:
 	MOVBQZX (CX), SI
 	ADDQ    SI, BX
+
 shortTail0:
 	RORQ $0x20, BP
 	XORQ $0x000000ff, BP
+
 shortAfter:
 	XORQ BX, BP
 	RORQ $0x21, BX
@@ -138,6 +151,7 @@ shortAfter:
 	XORQ BP, BX
 	MOVQ BX, ret+32(FP)
 	RET
+
 coreLong:
 	MOVQ  16(AX), DI
 	MOVQ  24(AX), AX
@@ -151,6 +165,7 @@ coreLong:
 	MOVQ  $0x9c1b8e1e9628323f, R8
 	IMULQ R8, SI
 	XORQ  SI, AX
+
 block:
 	MOVQ  (CX), SI
 	MOVQ  $0x00000000802910e3, R8
@@ -191,6 +206,7 @@ block:
 	JE    longCore2
 	CMPQ  SI, $0x00000003
 	JE    longCore3
+
 longCore3:
 	MOVQ  (CX), SI
 	MOVQ  $0x00000000802910e3, R9
@@ -200,6 +216,7 @@ longCore3:
 	XORQ  AX, BX
 	ADDQ  $0x00000008, CX
 	SUBQ  $0x00000008, DX
+
 longCore2:
 	MOVQ  (CX), SI
 	MOVQ  $0x00000000819b13af, R9
@@ -209,6 +226,7 @@ longCore2:
 	XORQ  DI, BP
 	ADDQ  $0x00000008, CX
 	SUBQ  $0x00000008, DX
+
 longCore1:
 	MOVQ  (CX), SI
 	MOVQ  $0x0000000091cb27e5, R9
@@ -218,6 +236,7 @@ longCore1:
 	ADDQ  BX, DI
 	ADDQ  $0x00000008, CX
 	SUBQ  $0x00000008, DX
+
 longCore0:
 	RORQ  $0x0b, AX
 	SUBQ  BP, AX
@@ -241,35 +260,44 @@ longCore0:
 	JE    longTail6
 	CMPQ  DX, $0x00000007
 	JE    longTail7
+
 longTail7:
 	MOVBQZX 6(CX), SI
 	ADDQ    SI, BP
+
 longTail6:
 	MOVWQZX 4(CX), SI
 	ADDQ    SI, DI
 	MOVLQZX (CX), SI
 	ADDQ    SI, AX
 	JMP     longAfter
+
 longTail5:
 	MOVBQZX 4(CX), SI
 	ADDQ    SI, BP
+
 longTail4:
 	MOVLQZX (CX), SI
 	ADDQ    SI, DI
 	JMP     longAfter
+
 longTail3:
 	MOVBQZX 2(CX), SI
 	ADDQ    SI, AX
+
 longTail2:
 	MOVWQZX (CX), SI
 	ADDQ    SI, BP
 	JMP     longAfter
+
 longTail1:
 	MOVBQZX (CX), SI
 	ADDQ    SI, DI
+
 longTail0:
 	ROLQ $0x20, AX
 	XORQ $0x000000ff, AX
+
 longAfter:
 	SUBQ DI, BP
 	RORQ $0x13, BX

examples/sum/sum.s

@@ -7,6 +7,7 @@ TEXT ·Sum(SB), NOSPLIT, $0-32
 	MOVQ xs_base(FP), AX
 	MOVQ xs_len+8(FP), CX
 	XORQ DX, DX
+
 loop:
 	CMPQ CX, $0x00
 	JE   done
@@ -14,6 +15,7 @@ loop:
 	ADDQ $0x08, AX
 	DECQ CX
 	JMP  loop
+
 done:
 	MOVQ DX, ret+24(FP)
 	RET

internal/prnt/printer.go

@@ -4,6 +4,7 @@ package prnt
 import (
 	"bytes"
 	"fmt"
+	"io"
 )
 
 // Generator provides convenience methods for code generators. In particular it
@@ -15,6 +16,11 @@ type Generator struct {
 	err error
 }
 
+// Raw provides direct access to the underlying output stream.
+func (g *Generator) Raw() io.Writer {
+	return &g.buf
+}
+
 // Printf prints to the internal buffer.
 func (g *Generator) Printf(format string, args ...interface{}) {
 	if g.err != nil {

printer/goasm.go

@@ -1,8 +1,10 @@
 package printer
 
 import (
+	"fmt"
 	"strconv"
 	"strings"
+	"text/tabwriter"
 
 	"github.com/mmcloughlin/avo/internal/prnt"
 	"github.com/mmcloughlin/avo/ir"
@@ -81,20 +83,29 @@ func (p *goasm) function(f *ir.Function) {
 	}
 	p.Printf(", %s\n", textsize(f))
 
+	w := p.tabwriter()
 	for _, node := range f.Nodes {
 		switch n := node.(type) {
 		case *ir.Instruction:
+			leader := []byte{tabwriter.Escape, '\t', tabwriter.Escape}
+			fmt.Fprint(w, string(leader)+n.Opcode)
 			if len(n.Operands) > 0 {
-				p.Printf("\t%s\t%s\n", n.Opcode, joinOperands(n.Operands))
+				fmt.Fprintf(w, "\t%s", joinOperands(n.Operands))
-			} else {
-				p.Printf("\t%s\n", n.Opcode)
 			}
+			fmt.Fprint(w, "\n")
 		case ir.Label:
-			p.Printf("%s:\n", n)
+			w.Flush()
+			w = p.tabwriter()
+			p.Printf("\n%s:\n", n)
 		default:
 			panic("unexpected node type")
 		}
 	}
+	w.Flush()
+}
+
+func (p *goasm) tabwriter() *tabwriter.Writer {
+	return tabwriter.NewWriter(p.Raw(), 4, 4, 1, ' ', tabwriter.StripEscape)
 }
 
 func (p *goasm) global(g *ir.Global) {

printer/goasm_test.go

@@ -24,10 +24,10 @@ func TestBasic(t *testing.T) {
 		"",
 		"// func add(x uint64, y uint64) uint64",
 		"TEXT ·add(SB), $0-24",
-		"\tMOVQ\tx(FP), AX",
+		"\tMOVQ x(FP), AX",
-		"\tMOVQ\ty+8(FP), R9",
+		"\tMOVQ y+8(FP), R9",
-		"\tADDQ\tAX, R9",
+		"\tADDQ AX, R9",
-		"\tMOVQ\tR9, ret+16(FP)",
+		"\tMOVQ R9, ret+16(FP)",
 		"\tRET",
 		"",
 	})