cmd/internal/obj/arm64: fix VMOVQ instruction encoding error

The VMOVQ instruction moves a 128-bit constant into a V register, as 128-bit
constant can't be loaded into a register directly, we split it into two 64-bit
constants and load it from constant pool. Currently we add the 128-bit constant
to literal pool by calling the 'addpool' function twice, this is not the right
way because it doesn't guarantee the two DWORD instructions are consecutive,
and the second call of addpool will overwrite the p.Pool field,resulting in a
wrong PC-relative offset value of the Prog.

This CL renames the flag LFROM3 to LFROM128, and adds a new function addpool128
to add a 128-bit constant to the literal pool.

Change-Id: I616f043c99a9a18a663f8768842cc980de2e6f79
Reviewed-on: https://go-review.googlesource.com/c/go/+/282334
Reviewed-by: eric fang <eric.fang@arm.com>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
Run-TryBot: eric fang <eric.fang@arm.com>
Trust: eric fang <eric.fang@arm.com>

src/cmd/internal/obj/arm64/asm7.go

@@ -280,7 +280,7 @@ func MOVCONST(d int64, s int, rt int) uint32 {
 const (
 	// Optab.flag
 	LFROM     = 1 << 0 // p.From uses constant pool
-	LFROM3    = 1 << 1 // p.From3 uses constant pool
+	LFROM128  = 1 << 1 // p.From3<<64+p.From forms a 128-bit constant in literal pool
 	LTO       = 1 << 2 // p.To uses constant pool
 	NOTUSETMP = 1 << 3 // p expands to multiple instructions, but does NOT use REGTMP
 )
@@ -419,7 +419,7 @@ var optab = []Optab{
 	{AMOVD, C_LACON, C_NONE, C_NONE, C_RSP, 34, 8, REGSP, LFROM, 0},
 
 	// Move a large constant to a vector register.
-	{AVMOVQ, C_VCON, C_NONE, C_VCON, C_VREG, 101, 4, 0, LFROM | LFROM3, 0},
+	{AVMOVQ, C_VCON, C_NONE, C_VCON, C_VREG, 101, 4, 0, LFROM128, 0},
 	{AVMOVD, C_VCON, C_NONE, C_NONE, C_VREG, 101, 4, 0, LFROM, 0},
 	{AVMOVS, C_LCON, C_NONE, C_NONE, C_VREG, 101, 4, 0, LFROM, 0},
 
@@ -995,8 +995,8 @@ func span7(ctxt *obj.Link, cursym *obj.LSym, newprog obj.ProgAlloc) {
 		if o.flag&LFROM != 0 {
 			c.addpool(p, &p.From)
 		}
-		if o.flag&LFROM3 != 0 {
-			c.addpool(p, p.GetFrom3())
+		if o.flag&LFROM128 != 0 {
+			c.addpool128(p, &p.From, p.GetFrom3())
 		}
 		if o.flag&LTO != 0 {
 			c.addpool(p, &p.To)
@@ -1201,6 +1201,36 @@ func (c *ctxt7) flushpool(p *obj.Prog, skip int) {
 	}
 }
 
+// addpool128 adds a 128-bit constant to literal pool by two consecutive DWORD
+// instructions, the 128-bit constant is formed by ah.Offset<<64+al.Offset.
+func (c *ctxt7) addpool128(p *obj.Prog, al, ah *obj.Addr) {
+	lit := al.Offset
+	q := c.newprog()
+	q.As = ADWORD
+	q.To.Type = obj.TYPE_CONST
+	q.To.Offset = lit
+	q.Pc = int64(c.pool.size)
+
+	lit = ah.Offset
+	t := c.newprog()
+	t.As = ADWORD
+	t.To.Type = obj.TYPE_CONST
+	t.To.Offset = lit
+	t.Pc = int64(c.pool.size + 8)
+	q.Link = t
+
+	if c.blitrl == nil {
+		c.blitrl = q
+		c.pool.start = uint32(p.Pc)
+	} else {
+		c.elitrl.Link = q
+	}
+
+	c.elitrl = t
+	c.pool.size += 16
+	p.Pool = q
+}
+
 /*
  * MOVD foo(SB), R is actually
  *   MOVD addr, REGTMP

src/cmd/internal/obj/arm64/asm_arm64_test.go

@@ -47,7 +47,7 @@ func TestLarge(t *testing.T) {
 
 	// assemble generated file
 	cmd := exec.Command(testenv.GoToolPath(t), "tool", "asm", "-S", "-o", filepath.Join(dir, "test.o"), tmpfile)
-	cmd.Env = append(os.Environ(), "GOARCH=arm64", "GOOS=linux")
+	cmd.Env = append(os.Environ(), "GOOS=linux")
 	out, err := cmd.CombinedOutput()
 	if err != nil {
 		t.Errorf("Assemble failed: %v, output: %s", err, out)
@@ -62,7 +62,7 @@ func TestLarge(t *testing.T) {
 
 	// build generated file
 	cmd = exec.Command(testenv.GoToolPath(t), "tool", "asm", "-o", filepath.Join(dir, "x.o"), tmpfile)
-	cmd.Env = append(os.Environ(), "GOARCH=arm64", "GOOS=linux")
+	cmd.Env = append(os.Environ(), "GOOS=linux")
 	out, err = cmd.CombinedOutput()
 	if err != nil {
 		t.Errorf("Build failed: %v, output: %s", err, out)
@@ -96,7 +96,7 @@ func TestNoRet(t *testing.T) {
 		t.Fatal(err)
 	}
 	cmd := exec.Command(testenv.GoToolPath(t), "tool", "asm", "-o", filepath.Join(dir, "x.o"), tmpfile)
-	cmd.Env = append(os.Environ(), "GOARCH=arm64", "GOOS=linux")
+	cmd.Env = append(os.Environ(), "GOOS=linux")
 	if out, err := cmd.CombinedOutput(); err != nil {
 		t.Errorf("%v\n%s", err, out)
 	}
@@ -134,7 +134,7 @@ func TestPCALIGN(t *testing.T) {
 			t.Fatal(err)
 		}
 		cmd := exec.Command(testenv.GoToolPath(t), "tool", "asm", "-S", "-o", tmpout, tmpfile)
-		cmd.Env = append(os.Environ(), "GOARCH=arm64", "GOOS=linux")
+		cmd.Env = append(os.Environ(), "GOOS=linux")
 		out, err := cmd.CombinedOutput()
 		if err != nil {
 			t.Errorf("The %s build failed: %v, output: %s", test.name, err, out)
@@ -150,3 +150,13 @@ func TestPCALIGN(t *testing.T) {
 		}
 	}
 }
+
+func testvmovq() (r1, r2 uint64)
+
+// TestVMOVQ checks if the arm64 VMOVQ instruction is working properly.
+func TestVMOVQ(t *testing.T) {
+	a, b := testvmovq()
+	if a != 0x7040201008040201 || b != 0x3040201008040201 {
+		t.Errorf("TestVMOVQ got: a=0x%x, b=0x%x, want: a=0x7040201008040201, b=0x3040201008040201", a, b)
+	}
+}

src/cmd/internal/obj/arm64/asm_arm64_test.s

@@ -0,0 +1,14 @@
+// Copyright 2021 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+#include "textflag.h"
+
+// testvmovq() (r1, r2 uint64)
+TEXT ·testvmovq(SB), NOSPLIT, $0-16
+	VMOVQ   $0x7040201008040201, $0x3040201008040201, V1
+	VMOV    V1.D[0], R0
+	VMOV    V1.D[1], R1
+	MOVD    R0, r1+0(FP)
+	MOVD    R1, r2+8(FP)
+	RET