cmd/compile: simplify zerorange on arm64

Get rid of large zeroing cases. We only use this code
for small things now.

Change-Id: Iba0a98785c5b4b72cf031763edb69ff741ca41af
Reviewed-on: https://go-review.googlesource.com/c/go/+/678936
Reviewed-by: Keith Randall <[email protected]>
Auto-Submit: Keith Randall <[email protected]>
Reviewed-by: Jorropo <[email protected]>
LUCI-TryBot-Result: Go LUCI <[email protected]>
Reviewed-by: Mark Freeman <[email protected]>

Author: randall77

src/cmd/compile/internal/arm64/ggen.go

@@ -5,9 +5,7 @@
 package arm64
 
 import (
-	"cmd/compile/internal/ir"
 	"cmd/compile/internal/objw"
-	"cmd/compile/internal/types"
 	"cmd/internal/obj"
 	"cmd/internal/obj/arm64"
 )
@@ -22,47 +20,20 @@ func padframe(frame int64) int64 {
 }
 
 func zerorange(pp *objw.Progs, p *obj.Prog, off, cnt int64, _ *uint32) *obj.Prog {
-	if cnt == 0 {
-		return p
+	if cnt%8 != 0 {
+		panic("zeroed region not aligned")
 	}
-	if cnt < int64(4*types.PtrSize) {
-		for i := int64(0); i < cnt; i += int64(types.PtrSize) {
-			p = pp.Append(p, arm64.AMOVD, obj.TYPE_REG, arm64.REGZERO, 0, obj.TYPE_MEM, arm64.REGSP, 8+off+i)
-		}
-	} else if cnt <= int64(128*types.PtrSize) {
-		if cnt%(2*int64(types.PtrSize)) != 0 {
-			p = pp.Append(p, arm64.AMOVD, obj.TYPE_REG, arm64.REGZERO, 0, obj.TYPE_MEM, arm64.REGSP, 8+off)
-			off += int64(types.PtrSize)
-			cnt -= int64(types.PtrSize)
-		}
-		p = pp.Append(p, arm64.AMOVD, obj.TYPE_REG, arm64.REGSP, 0, obj.TYPE_REG, arm64.REG_R20, 0)
-		p = pp.Append(p, arm64.AADD, obj.TYPE_CONST, 0, 8+off, obj.TYPE_REG, arm64.REG_R20, 0)
-		p.Reg = arm64.REG_R20
-		p = pp.Append(p, obj.ADUFFZERO, obj.TYPE_NONE, 0, 0, obj.TYPE_MEM, 0, 0)
-		p.To.Name = obj.NAME_EXTERN
-		p.To.Sym = ir.Syms.Duffzero
-		p.To.Offset = 4 * (64 - cnt/(2*int64(types.PtrSize)))
-	} else {
-		// Not using REGTMP, so this is async preemptible (async preemption clobbers REGTMP).
-		// We are at the function entry, where no register is live, so it is okay to clobber
-		// other registers
-		const rtmp = arm64.REG_R20
-		p = pp.Append(p, arm64.AMOVD, obj.TYPE_CONST, 0, 8+off-8, obj.TYPE_REG, rtmp, 0)
-		p = pp.Append(p, arm64.AMOVD, obj.TYPE_REG, arm64.REGSP, 0, obj.TYPE_REG, arm64.REGRT1, 0)
-		p = pp.Append(p, arm64.AADD, obj.TYPE_REG, rtmp, 0, obj.TYPE_REG, arm64.REGRT1, 0)
-		p.Reg = arm64.REGRT1
-		p = pp.Append(p, arm64.AMOVD, obj.TYPE_CONST, 0, cnt, obj.TYPE_REG, rtmp, 0)
-		p = pp.Append(p, arm64.AADD, obj.TYPE_REG, rtmp, 0, obj.TYPE_REG, arm64.REGRT2, 0)
-		p.Reg = arm64.REGRT1
-		p = pp.Append(p, arm64.AMOVD, obj.TYPE_REG, arm64.REGZERO, 0, obj.TYPE_MEM, arm64.REGRT1, int64(types.PtrSize))
-		p.Scond = arm64.C_XPRE
-		p1 := p
-		p = pp.Append(p, arm64.ACMP, obj.TYPE_REG, arm64.REGRT1, 0, obj.TYPE_NONE, 0, 0)
-		p.Reg = arm64.REGRT2
-		p = pp.Append(p, arm64.ABNE, obj.TYPE_NONE, 0, 0, obj.TYPE_BRANCH, 0, 0)
-		p.To.SetTarget(p1)
+	off += 8 // return address was ignored in offset calculation
+	for cnt >= 16 && off < 512 {
+		p = pp.Append(p, arm64.ASTP, obj.TYPE_REGREG, arm64.REGZERO, arm64.REGZERO, obj.TYPE_MEM, arm64.REGSP, off)
+		off += 16
+		cnt -= 16
+	}
+	for cnt != 0 {
+		p = pp.Append(p, arm64.AMOVD, obj.TYPE_REG, arm64.REGZERO, 0, obj.TYPE_MEM, arm64.REGSP, off)
+		off += 8
+		cnt -= 8
 	}
-
 	return p
 }