feat(asm): emit the loong64 stack-split guard for small and medium frames

asm/loong64_frame.go

@@ -36,6 +36,11 @@ type loong64FrameInfo struct {
 	args     int  // the declared -argsize
 	noSplit  bool // the NOSPLIT flag
 	leaf     bool // no call instructions in the body
+
+	// Stack-split guard state: like amd64 and arm64, a leaf function with a
+	// small autosize is auto-marked NOSPLIT by the toolchain.
+	needSplit  bool
+	splitClass int // 0: <=StackSmall, 1: <=StackBig, 2: >StackBig
 }
 
 // loong64ComputeFrame derives the frame layout for a TEXT function.
@@ -59,9 +64,110 @@ func loong64ComputeFrame(t *ast.Text) loong64FrameInfo {
 		// A zero-frame non-leaf function still opens an 8-byte frame for LR.
 		fi.autosize = 8
 	}
+	switch {
+	case fi.noSplit:
+	case fi.autosize < stackSmall && fi.leaf:
+		// Auto-NOSPLIT, as the toolchain's leaf mark concludes.
+	default:
+		fi.needSplit = true
+		switch {
+		case fi.autosize <= stackSmall:
+			fi.splitClass = 0
+		case fi.autosize <= stackBig:
+			fi.splitClass = 1
+		default:
+			fi.splitClass = 2
+		}
+	}
 	return fi
 }
 
+// loong64GuardLen returns the byte length of the stack-split guard prefix
+// (zero when the function needs no guard).  The big class materialises two
+// constants through R30.
+func loong64GuardLen(fi loong64FrameInfo) int {
+	if !fi.needSplit {
+		return 0
+	}
+	switch fi.splitClass {
+	case 0:
+		return 12
+	case 1:
+		return 16
+	default:
+		return 40 // MOVV + [LU12IW+ORI] + SGTU + BNE + [LU12IW+ORI] + ADDV + SGTU + BEQ
+	}
+}
+
+// loong64Lu12iOri materialises the 32-bit constant v in rd with the
+// toolchain's LU12IW/ORI pair (the ORI reads and writes rd itself).
+func loong64Lu12iOri(rd int, v int64) []uint32 {
+	hi := int32(v >> 12)
+	lo := int32(v & 0xFFF)
+	return []uint32{
+		0x0a<<25 | uint32(hi&0xFFFFF)<<5 | uint32(rd),
+		0x0e<<22 | uint32(lo)<<10 | uint32(rd)<<5 | uint32(rd),
+	}
+}
+
+// loong64GuardBytes emits the stack-split guard prefix.  blockStart is the
+// function-relative address of the morestack call at the end of the function;
+// branch displacements are in instructions.
+func loong64GuardBytes(fi loong64FrameInfo, blockStart int) []byte {
+	// MOVV 16(g), R20 (g.stackguard0), g = R22.
+	ws := []uint32{l64irr(l64loadStoreTable["MOVV"].ld, 16, 22, 20)}
+	switch fi.splitClass {
+	case 0:
+		// SGTU SP, R20, R20; BEQ R20, more
+		ws = append(ws, l64rrr(l64DualTable["SGTU"].rrr, 3, 20, 20))
+		ws = append(ws, loong64Beqz(20, int32((blockStart-8)>>2)))
+	case 1:
+		off := int32(fi.autosize - stackSmall)
+		ws = append(ws, l64irr(l64DualTable["ADDV"].imm, int(-off), 3, 24))
+		ws = append(ws, l64rrr(l64DualTable["SGTU"].rrr, 24, 20, 20))
+		ws = append(ws, loong64Beqz(20, int32((blockStart-12)>>2)))
+	default:
+		off := int64(fi.autosize - stackSmall)
+		movLen := 8 // LU12IW + ORI
+		ws = append(ws, loong64Lu12iOri(30, off)...)
+		ws = append(ws, l64rrr(l64DualTable["SGTU"].rrr, 30, 3, 24))
+		ws = append(ws, loong64Bnez(24, int32((blockStart-(8+movLen))>>2)))
+		ws = append(ws, loong64Lu12iOri(30, -off)...)
+		ws = append(ws, l64rrr(l64DualTable["ADDV"].rrr, 30, 3, 24))
+		ws = append(ws, l64rrr(l64DualTable["SGTU"].rrr, 24, 20, 20))
+		ws = append(ws, loong64Beqz(20, int32((blockStart-loong64GuardLen(fi)+12)>>2)))
+	}
+	return l64WordsLE(ws...)
+}
+
+// loong64Beqz/loong64Bnez build the 21-bit conditional branches against R0
+// that the toolchain emits for its guard compares.
+func loong64Beqz(rj int, dispInstr int32) uint32 {
+	return l64ir21(l64branch21Table["BEQZ"], int(dispInstr), rj)
+}
+
+func loong64Bnez(rj int, dispInstr int32) uint32 {
+	return l64ir21(l64branch21Table["BNEZ"], int(dispInstr), rj)
+}
+
+// loong64MoreStackBlock emits the trailing block: MOVV R1, R31 (save LR),
+// BL runtime.morestack_noctxt, B back to the function entry.
+func loong64MoreStackBlock(blockStart int) ([]byte, Reloc) {
+	ws := []uint32{
+		l64rrr(l64DualTable["ADD"].rrr, 0, 1, 31), // MOVV R1, R31 (ADD R1, R0, R31)
+		l64bbl(l64jumpTable["BL"], 0),             // BL, patched by the linker
+	}
+	disp := (-(blockStart + 8)) >> 2
+	ws = append(ws, l64bbl(l64jumpTable["B"], int(disp)))
+	reloc := Reloc{
+		Off:   blockStart + 4,
+		After: blockStart + 8,
+		Name:  "runtime\u00b7morestack_noctxt",
+		Kind:  RelLoong64Branch,
+	}
+	return l64WordsLE(ws...), reloc
+}
+
 // loong64IsLeaf reports whether a function contains no call instructions
 // (JAL/BL/CALL), matching the toolchain's LEAF mark, which drives the frame
 // and the epilogue shape.