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

asm/goobjloong64.go

@@ -25,11 +25,16 @@ func (img *Image) GOObjectLOONG64(pkgPath, srcPath string) ([]byte, error) {
 	}
 	return img.emitGOObject(pkgPath, srcPath, pre, 4, func(r Reloc) (uint16, uint8) {
 		// A pcalau12i+addi.d pair: the high part carries
-		// R_LOONG64_ADDR_HI, the low part R_LOONG64_ADDR_LO.
+		// R_LOONG64_ADDR_HI, the low part R_LOONG64_ADDR_LO; the guard's
-		if r.Kind == RelLoong64AddrLo {
+		// morestack call carries R_CALLLOONG64.
+		switch {
+		case r.Kind == RelLoong64AddrLo:
 			return relocLoong64AddrLo, 4
-		}
+		case r.Kind == RelLoong64Branch:
+			return relocCallLoong64, 4
+		default:
 			return relocLoong64AddrHi, 4
+		}
 	})
 }
 
@@ -39,6 +44,7 @@ func (img *Image) GOObjectLOONG64(pkgPath, srcPath string) ([]byte, error) {
 const (
 	relocLoong64AddrHi = 77 // R_LOONG64_ADDR_HI
 	relocLoong64AddrLo = 78 // R_LOONG64_ADDR_LO
+	relocCallLoong64   = 84 // R_CALLLOONG64
 )
 
 // toolchainObjectPreambleLOONG64 returns the "go object ...\n!\n" header

asm/guard_test.go

@@ -185,3 +185,42 @@ func TestStackGuardBytesRISCV64(t *testing.T) {
 		}
 	}
 }
+
+// The loong64 stack-split guard, pinned from `go tool asm` (Go 1.27,
+// loong64) for the small and medium frame classes plus auto-NOSPLIT.  The
+// big class (>StackBig) and the huge-frame body prologue remain toolchain
+// divergences tracked separately.
+func TestStackGuardBytesLOONG64(t *testing.T) {
+	for _, tt := range []struct {
+		name string
+		src  string
+		want string
+	}{
+		{"leafsmall", "TEXT \u00b7leafsmall(SB), $16-0\n\tRET\n",
+			"61a0ff2963a0ff026100c0296360c0022000004c"},
+		{"leafmed", "TEXT \u00b7leafmed(SB), $256-0\n\tRET\n",
+			"d442c02878e0fd0294e21200801a004061e0fb2963e0fb026100c0296320c4022000004c3f00100000000000ffd7ff53"},
+		{"nosplit", "TEXT \u00b7nosplit(SB), NOSPLIT, $16-0\n\tRET\n",
+			"61a0ff2963a0ff026100c0296360c0022000004c"},
+	} {
+		f, errs := parser.Parse("g_loong64.s", tt.src)
+		if len(errs) > 0 {
+			t.Fatalf("%s: parse: %v", tt.name, errs)
+		}
+		img, err := AssembleFileLOONG64(f)
+		if err != nil {
+			t.Fatalf("%s: assemble: %v", tt.name, err)
+		}
+		fn := img.Funcs[0]
+		code := append([]byte(nil), img.Code[fn.Offset:fn.Offset+fn.Size]...)
+		for _, r := range fn.Relocs {
+			for j := r.Off; j < r.Off+4 && j < len(code); j++ {
+				code[j] = 0
+			}
+		}
+		got := hex.EncodeToString(code)
+		if got != tt.want {
+			t.Errorf("%s:\n  got  %s\n  want %s", tt.name, got, tt.want)
+		}
+	}
+}

asm/link.go

@@ -102,6 +102,7 @@ const (
 	RelArm64Addr                        // R_ADDRARM64 (ADRP + ADD pair)
 	RelArm64Branch                      // R_CALLARM64 (BL instruction)
 	RelArm64LDST64                      // R_ARM64_PCREL_LDST64 (ADRP + 64-bit LDR/STR pair)
+	RelLoong64Branch                    // R_CALLLOONG64 (BL instruction)
 )
 
 type Reloc struct {

asm/loong64_assemble.go

@@ -27,6 +27,7 @@ import (
 func assembleLOONG64(t *ast.Text) ([]byte, map[string]int, []Reloc, []LineEntry, []SpadjStep, error) {
 	fi := loong64ComputeFrame(t)
 	prologue := loong64Prologue(fi)
+	guardLen := loong64GuardLen(fi)
 	chain := loong64JumpChain(t)
 	resolve := func(name string) string {
 		if r, ok := chain[name]; ok {
@@ -42,12 +43,12 @@ func assembleLOONG64(t *ast.Text) ([]byte, map[string]int, []Reloc, []LineEntry,
 	// delta by autosize; the boundary is reported at the third instruction's
 	// pc, exactly as the toolchain's pctospadj does.
 	if fi.autosize != 0 {
-		spadj = append(spadj, SpadjStep{PC: 8, Value: fi.autosize})
+		spadj = append(spadj, SpadjStep{PC: guardLen + 8, Value: fi.autosize})
 	}
 
 	// Pass 1: label offsets from the instruction sizes.
 	offsets := map[string]int{}
-	pos := len(prologue)
+	pos := guardLen + len(prologue)
 	for _, stmt := range t.Body {
 		switch s := stmt.(type) {
 		case *ast.Label:
@@ -57,9 +58,25 @@ func assembleLOONG64(t *ast.Text) ([]byte, map[string]int, []Reloc, []LineEntry,
 		}
 	}
 
-	// Pass 2: encode.  Relocation offsets are recorded function-relative.
+	// Pass 2: encode.  The guard prefix precedes the prologue; its branches
-	out := append([]byte(nil), prologue...)
+	// target the morestack block at the end of the function, which the first
-	pc := len(prologue)
+	// pass has sized.
+	bodyLen := 0
+	{
+		p := guardLen + len(prologue)
+		for _, stmt := range t.Body {
+			if in, ok := stmt.(*ast.Instr); ok {
+				p += loong64InstrSize(in, fi)
+			}
+		}
+		bodyLen = p - (guardLen + len(prologue))
+	}
+	var out []byte
+	if fi.needSplit {
+		out = append(out, loong64GuardBytes(fi, guardLen+len(prologue)+bodyLen)...)
+	}
+	out = append(out, prologue...)
+	pc := guardLen + len(prologue)
 	preCount := len(relocs)
 	var lines []LineEntry
 	for _, stmt := range t.Body {
@@ -94,6 +111,12 @@ func assembleLOONG64(t *ast.Text) ([]byte, map[string]int, []Reloc, []LineEntry,
 		out = append(out, code...)
 		pc += len(code)
 	}
+	if fi.needSplit {
+		block, blReloc := loong64MoreStackBlock(pc)
+		out = append(out, block...)
+		relocs = append(relocs, blReloc)
+		pc += len(block)
+	}
 	return out, offsets, relocs, lines, spadj, nil
 }
 

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.