Revert "cmd/compile: redo arm64 LR/FP save and restore"

This reverts commit 719dfcf.

Reason for revert: Causing crashes.

Change-Id: I0b8526dd03d82fa074ce4f97f1789eeac702b3eb
Reviewed-on: https://go-review.googlesource.com/c/go/+/709755
Reviewed-by: Keith Randall <[email protected]>
Reviewed-by: David Chase <[email protected]>
LUCI-TryBot-Result: Go LUCI <[email protected]>
Auto-Submit: Keith Randall <[email protected]>
Reviewed-by: Cherry Mui <[email protected]>

Author: randall77

src/cmd/compile/abi-internal.md

@@ -576,19 +576,19 @@ A function's stack frame, after the frame is created, is laid out as
 follows:
 
     +------------------------------+
-    | return PC                    |
-    | frame pointer on entry       | ← R29 points to
     | ... locals ...               |
     | ... outgoing arguments ...   |
-    | unused word                  | ← RSP points to
+    | return PC                    | ← RSP points to
+    | frame pointer on entry       |
     +------------------------------+ ↓ lower addresses
 
 The "return PC" is loaded to the link register, R30, as part of the
 arm64 `CALL` operation.
 
-On entry, a function pushes R30 (the return address) and R29
-(the caller's frame pointer) onto the bottom of the stack. It then
-subtracts a constant from RSP to open its stack frame.
+On entry, a function subtracts from RSP to open its stack frame, and
+saves the values of R30 and R29 at the bottom of the frame.
+Specifically, R30 is saved at 0(RSP) and R29 is saved at -8(RSP),
+after RSP is updated.
 
 A leaf function that does not require any stack space may omit the
 saved R30 and R29.

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

@@ -11,12 +11,10 @@ import (
 )
 
 func padframe(frame int64) int64 {
-	// arm64 requires frame sizes here that are 8 mod 16.
-	// With the additional (unused) slot at the bottom of the frame,
-	// that makes an aligned 16 byte frame.
-	// Adding a save region for LR+FP does not change the alignment.
-	if frame != 0 {
-		frame += (-(frame + 8)) & 15
+	// arm64 requires that the frame size (not counting saved FP&LR)
+	// be 16 bytes aligned. If not, pad it.
+	if frame%16 != 0 {
+		frame += 16 - (frame % 16)
 	}
 	return frame
 }

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

@@ -221,7 +221,7 @@ func ssaGenValue(s *ssagen.State, v *ssa.Value) {
 
 		for i := 0; i < len(args); i++ {
 			a := args[i]
-			// Offset by size of the unused slot before start of args.
+			// Offset by size of the saved LR slot.
 			addr := ssagen.SpillSlotAddr(a, arm64.REGSP, base.Ctxt.Arch.FixedFrameSize)
 			// Look for double-register operations if we can.
 			if i < len(args)-1 {

src/cmd/compile/internal/ssagen/pgen.go

@@ -393,16 +393,10 @@ func StackOffset(slot ssa.LocalSlot) int32 {
 	case ir.PAUTO:
 		off = n.FrameOffset()
 		if base.Ctxt.Arch.FixedFrameSize == 0 {
-			// x86 return address
 			off -= int64(types.PtrSize)
 		}
 		if buildcfg.FramePointerEnabled {
-			// frame pointer
 			off -= int64(types.PtrSize)
-			if buildcfg.GOARCH == "arm64" {
-				// arm64 return address also
-				off -= int64(types.PtrSize)
-			}
 		}
 	}
 	return int32(off + slot.Off)

src/cmd/compile/internal/ssagen/ssa.go

@@ -7150,15 +7150,14 @@ func defframe(s *State, e *ssafn, f *ssa.Func) {
 	// Insert code to zero ambiguously live variables so that the
 	// garbage collector only sees initialized values when it
 	// looks for pointers.
-	// Note: lo/hi are offsets from varp and will be negative.
 	var lo, hi int64
 
 	// Opaque state for backend to use. Current backends use it to
 	// keep track of which helper registers have been zeroed.
 	var state uint32
 
 	// Iterate through declarations. Autos are sorted in decreasing
-	// frame offset order (least negative to most negative).
+	// frame offset order.
 	for _, n := range e.curfn.Dcl {
 		if !n.Needzero() {
 			continue

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

@@ -51,6 +51,7 @@ type ctxt7 struct {
 	blitrl     *obj.Prog
 	elitrl     *obj.Prog
 	autosize   int32
+	extrasize  int32
 	instoffset int64
 	pc         int64
 	pool       struct {
@@ -1121,7 +1122,8 @@ func span7(ctxt *obj.Link, cursym *obj.LSym, newprog obj.ProgAlloc) {
 		ctxt.Diag("arm64 ops not initialized, call arm64.buildop first")
 	}
 
-	c := ctxt7{ctxt: ctxt, newprog: newprog, cursym: cursym, autosize: int32(p.To.Offset)}
+	c := ctxt7{ctxt: ctxt, newprog: newprog, cursym: cursym, autosize: int32(p.To.Offset & 0xffffffff), extrasize: int32(p.To.Offset >> 32)}
+	p.To.Offset &= 0xffffffff // extrasize is no longer needed
 
 	// Process literal pool and allocate initial program counter for each Prog, before
 	// generating branch veneers.
@@ -2117,8 +2119,8 @@ func (c *ctxt7) aclass(a *obj.Addr) int {
 				// a.Offset is still relative to pseudo-SP.
 				a.Reg = obj.REG_NONE
 			}
-			// The frame top 16 bytes are for LR/FP
-			c.instoffset = int64(c.autosize) + a.Offset - extrasize
+			// The frame top 8 or 16 bytes are for FP
+			c.instoffset = int64(c.autosize) + a.Offset - int64(c.extrasize)
 			return autoclass(c.instoffset)
 
 		case obj.NAME_PARAM:
@@ -2178,8 +2180,8 @@ func (c *ctxt7) aclass(a *obj.Addr) int {
 				// a.Offset is still relative to pseudo-SP.
 				a.Reg = obj.REG_NONE
 			}
-			// The frame top 16 bytes are for LR/FP
-			c.instoffset = int64(c.autosize) + a.Offset - extrasize
+			// The frame top 8 or 16 bytes are for FP
+			c.instoffset = int64(c.autosize) + a.Offset - int64(c.extrasize)
 
 		case obj.NAME_PARAM:
 			if a.Reg == REGSP {