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

Instead of storing LR (the return address) at 0(SP) and the FP
(parent's frame pointer) at -8(SP), store them at framesize-8(SP)
and framesize-16(SP), respectively.

We push and pop data onto the stack such that we're never accessing
anything below SP.

The prolog/epilog lengths are unchanged (3 insns for a typical prolog,
2 for a typical epilog).

We use 8 bytes more per frame.

Typical prologue:

    STP.W   (FP, LR), -16(SP)
    MOVD    SP, FP
    SUB	    $C, SP

Typical epilogue:

    ADD     $C, SP
    LDP.P   16(SP), (FP, LR)
    RET

The previous word where we stored LR, at 0(SP), is now unused.
We could repurpose that slot for storing a local variable.

The new prolog and epilog instructions are recognized by libunwind,
so pc-sampling tools like perf should now be accurate. (TODO: except
maybe after the first RET instruction? Have to look into that.)

Update golang#73753 (fixes, for arm64)
Update golang#57302 (Quim thinks this will help on that issue)

Change-Id: I4800036a9a9a08aaaf35d9f99de79a36cf37ebb8
Reviewed-on: https://go-review.googlesource.com/c/go/+/674615
Reviewed-by: David Chase <[email protected]>
LUCI-TryBot-Result: Go LUCI <[email protected]>
Reviewed-by: Keith Randall <[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 ...   |
-    | return PC                    | ← RSP points to
-    | frame pointer on entry       |
+    | unused word                  | ← RSP points to
     +------------------------------+ ↓ lower addresses
 
 The "return PC" is loaded to the link register, R30, as part of the
 arm64 `CALL` operation.
 
-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.
+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.
 
 A leaf function that does not require any stack space may omit the
 saved R30 and R29.

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

@@ -11,10 +11,12 @@ import (
 )
 
 func padframe(frame int64) int64 {
-	// 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)
+	// 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
 	}
 	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 saved LR slot.
+			// Offset by size of the unused slot before start of args.
 			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

@@ -393,10 +393,16 @@ 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

@@ -7150,14 +7150,15 @@ 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.
+	// frame offset order (least negative to most negative).
 	for _, n := range e.curfn.Dcl {
 		if !n.Needzero() {
 			continue

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

@@ -51,7 +51,6 @@ type ctxt7 struct {
 	blitrl     *obj.Prog
 	elitrl     *obj.Prog
 	autosize   int32
-	extrasize  int32
 	instoffset int64
 	pc         int64
 	pool       struct {
@@ -1122,8 +1121,7 @@ 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 & 0xffffffff), extrasize: int32(p.To.Offset >> 32)}
-	p.To.Offset &= 0xffffffff // extrasize is no longer needed
+	c := ctxt7{ctxt: ctxt, newprog: newprog, cursym: cursym, autosize: int32(p.To.Offset)}
 
 	// Process literal pool and allocate initial program counter for each Prog, before
 	// generating branch veneers.
@@ -2119,8 +2117,8 @@ func (c *ctxt7) aclass(a *obj.Addr) int {
 				// a.Offset is still relative to pseudo-SP.
 				a.Reg = obj.REG_NONE
 			}
-			// The frame top 8 or 16 bytes are for FP
-			c.instoffset = int64(c.autosize) + a.Offset - int64(c.extrasize)
+			// The frame top 16 bytes are for LR/FP
+			c.instoffset = int64(c.autosize) + a.Offset - extrasize
 			return autoclass(c.instoffset)
 
 		case obj.NAME_PARAM:
@@ -2180,8 +2178,8 @@ func (c *ctxt7) aclass(a *obj.Addr) int {
 				// a.Offset is still relative to pseudo-SP.
 				a.Reg = obj.REG_NONE
 			}
-			// The frame top 8 or 16 bytes are for FP
-			c.instoffset = int64(c.autosize) + a.Offset - int64(c.extrasize)
+			// The frame top 16 bytes are for LR/FP
+			c.instoffset = int64(c.autosize) + a.Offset - extrasize
 
 		case obj.NAME_PARAM:
 			if a.Reg == REGSP {