feat(asm): emit the loong64 stack-split guard for big frames

asm/arm64_frame.go

@@ -182,7 +182,7 @@ func arm64Prologue(fi arm64FrameInfo) []byte {
 
 // arm64SubImmWords emits SUB $imm, SP, Rd: the immediate form when the value
 // fits the imm12 field, otherwise the toolchain materialises it into REGTMP
-// (R27) and subtracts the register.
+// (R27) and subtracts the register in the extended-register form.
 func arm64SubImmWords(imm uint32, rd uint32) []uint32 {
 	if imm <= 0xFFF {
 		return []uint32{a64AddSub(1, 1, 0, 0, imm, 31, rd)}
@@ -191,7 +191,7 @@ func arm64SubImmWords(imm uint32, rd uint32) []uint32 {
 	if err != nil {
 		mov = nil
 	}
-	return append(wordsOf(mov), arm64DPSRWords(arm64OpSub, 27, 31, rd))
+	return append(wordsOf(mov), arm64DPExtWords(arm64OpSub, 27, 31, rd))
 }
 
 // arm64AddImmWords emits ADD $imm, SP, Rd with the same REGTMP fallback.
@@ -203,7 +203,7 @@ func arm64AddImmWords(imm uint32, rd uint32) []uint32 {
 	if err != nil {
 		mov = nil
 	}
-	return append(wordsOf(mov), arm64DPSRWords(arm64OpAdd, 27, 31, rd))
+	return append(wordsOf(mov), arm64DPExtWords(arm64OpAdd, 27, 31, rd))
 }
 
 // arm64Return returns the bytes for a RET: the epilogue (restore FP/LR and
@@ -326,6 +326,13 @@ func arm64DPSRWords(base uint32, rm, rn, rd uint32) uint32 {
 	return base | rm<<16 | rn<<5 | rd
 }
 
+// arm64DPExtWords builds one data-processing (extended register) word, the
+// form the toolchain picks when a large immediate was materialised into
+// REGTMP before the operation: base | 1<<21 | Rm<<16 | UXTX<<13 | Rn<<5 | Rd.
+func arm64DPExtWords(base, rm, rn, rd uint32) uint32 {
+	return base | 1<<21 | rm<<16 | 3<<13 | rn<<5 | rd
+}
+
 // wordsOf converts little-endian instruction bytes back to words.
 func wordsOf(b []byte) []uint32 {
 	ws := make([]uint32, 0, len(b)/4)
@@ -359,7 +366,7 @@ func arm64GuardBytes(fi arm64FrameInfo, blockStart int) []byte {
 		}
 		ws = append(ws, wordsOf(mov)...)
 		ml := len(mov) / 4
-		ws = append(ws, arm64DPSRWords(arm64OpSubs, 27, 31, 17)) // SUBS R27, RSP, R17
+		ws = append(ws, arm64DPExtWords(arm64OpSubs, 27, 31, 17)) // SUBS R17, RSP, R27
 		ws = append(ws, br(8+ml, a64CondLO))
 		ws = append(ws, arm64DPSRWords(arm64OpSubs, 16, 17, 31)) // CMP R16, R17
 		ws = append(ws, br(8+ml+8, a64CondLS))

asm/elfloong64.go

@@ -16,6 +16,7 @@ const (
 	// LoongArch relocation types (the ELF psABI).
 	rLarchPCALAHI20 = 71 // R_LARCH_PCALA_HI20 (pcalau12i)
 	rLarchPCALALO12 = 72 // R_LARCH_PCALA_LO12 (addi.d/ld/st)
+	rLarchB26       = 66 // R_LARCH_B26 (b/bl, matches the Go linker's mapping)
 )
 
 // ELFLOONG64Object returns the image as an ELF64 relocatable object file for
@@ -95,8 +96,11 @@ func (img *Image) ELFLOONG64Object() ([]byte, error) {
 				return nil, fmt.Errorf("relocation references unknown symbol %q", r.Name)
 			}
 			typ := uint32(rLarchPCALAHI20)
-			if r.Kind == RelLoong64AddrLo {
+			switch r.Kind {
+			case RelLoong64AddrLo:
 				typ = rLarchPCALALO12
+			case RelLoong64Branch:
+				typ = rLarchB26
 			}
 			relas = append(relas, elfRela{
 				off:    uint64(fn.Offset + r.Off),

asm/elfloong64_test.go

@@ -198,3 +198,45 @@ TEXT ·nop(SB), NOSPLIT, $0
 		t.Error("function symbol nop not found")
 	}
 }
+
+// TestELFLOONG64BranchRelocation checks that the morestack call and an
+// internal CALL both carry R_LARCH_B26 in the emitted object, matching the
+// Go linker's mapping of its call relocation.
+func TestELFLOONG64BranchRelocation(t *testing.T) {
+	f, errs := parser.Parse("k_loong64.s", "TEXT \u00b7callbig(SB), $8192-0\n\tCALL \u00b7other(SB)\n\tRET\nTEXT \u00b7other(SB), NOSPLIT, $0\n\tRET\n")
+	if len(errs) > 0 {
+		t.Fatalf("parse: %v", errs)
+	}
+	img, err := AssembleFileLOONG64(f)
+	if err != nil {
+		t.Fatalf("AssembleFileLOONG64: %v", err)
+	}
+	obj, err := img.ELFLOONG64Object()
+	if err != nil {
+		t.Fatalf("ELFLOONG64Object: %v", err)
+	}
+	ef, err := elf.NewFile(bytes.NewReader(obj))
+	if err != nil {
+		t.Fatalf("parse emitted object: %v", err)
+	}
+	defer ef.Close()
+	relaSec := ef.Section(".rela.text")
+	if relaSec == nil {
+		t.Fatal("missing .rela.text")
+	}
+	raw, err := relaSec.Data()
+	if err != nil {
+		t.Fatal(err)
+	}
+	// The guard's morestack call plus the body's CALL to other.
+	if len(raw)%24 != 0 || len(raw)/24 != 2 {
+		t.Fatalf(".rela.text has %d bytes, want two 24-byte entries", len(raw))
+	}
+	le := binary.LittleEndian
+	for i := range 2 {
+		info := le.Uint64(raw[i*24+8:])
+		if elf.R_LARCH(info&0xffffffff) != elf.R_LARCH_B26 {
+			t.Errorf("relocation %d type = %v, want R_LARCH_B26", i, elf.R_LARCH(info&0xffffffff))
+		}
+	}
+}

asm/goobj.go

@@ -156,6 +156,16 @@ const goobjBuiltinMorestackNoctxt = 246
 // stack-guard call.
 var goobjBuiltinMorestack = "runtime\u00b7morestack_noctxt"
 
+// isCallReloc reports whether k is one of the per-arch call relocations a
+// direct branch to a TEXT symbol carries.
+func isCallReloc(k RelocKind) bool {
+	switch k {
+	case RelCall, RelRISCVJal, RelArm64Branch, RelLoong64Branch:
+		return true
+	}
+	return false
+}
+
 const goobjMagic = "\x00go120ld"
 
 // goSym is one symbol definition under construction.
@@ -436,7 +446,7 @@ func (img *Image) emitGOObject(pkgPath, srcPath string, pre []byte, minLC int, r
 				// A call to a TEXT function of the same file references the
 				// non-package definition table.
 				ni, isText := textNpIdx[r.Name]
-				if !isText || r.Kind != RelCall {
+				if !isText || !isCallReloc(r.Kind) {
 					return nil, fmt.Errorf("GOOBJ emission: reference to unknown symbol %q", r.Name)
 				}
 				pkg = pkgIdxNone

asm/guard_test.go

@@ -8,6 +8,7 @@ import (
 	"encoding/hex"
 	"testing"
 
+	"sourcedock.dev/petrbalvin/gasm-devkit/ast"
 	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
 )
 
@@ -116,7 +117,7 @@ func TestStackGuardBytesARM64(t *testing.T) {
 		{"leafmed", "TEXT \u00b7leafmed(SB), $256-0\n\tRET\n",
 			"900b40f9f14302d13f0210eb09010054f44304d19dfa3fa99f020091fd2300d1fd230491ff430491c0035fd6e3031eaa00000000f3ffff17"},
 		{"leafbig", "TEXT \u00b7leafbig(SB), $8192-0\n\tRET\n",
-			"900b40f91bf283d2f1031beba30100543f0210eb690100541b0284d2f4031bcb9dfa3fa99f020091fd2300d11b0184d2fd031b8b1b0284d2ff031b8bc0035fd6e3031eaa00000000eeffff17"},
+			"900b40f91bf283d2f1633beba30100543f0210eb690100541b0284d2f4633bcb9dfa3fa99f020091fd2300d11b0184d2fd633b8b1b0284d2ff633b8bc0035fd6e3031eaa00000000eeffff17"},
 		{"callsmall", "TEXT \u00b7callsmall(SB), $16-0\n\tCALL \u00b7other(SB)\n\tRET\nTEXT \u00b7other(SB), NOSPLIT, $0\n\tRET\n",
 			"900b40f9ff6330eb09010054fe0f1ef8fd831ff8fd2300d100000000fd835ff8fe0742f8c0035fd6e3031eaa00000000f4ffff17"},
 		{"nosplit", "TEXT \u00b7nosplit(SB), NOSPLIT, $16-0\n\tRET\n",
@@ -187,9 +188,10 @@ 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
+// loong64): every guard class (including the medium class with the
-// big class (>StackBig) and the huge-frame body prologue remain toolchain
+// materialised constant and the big class with the ORI-less constants), the
-// divergences tracked separately.
+// auto-NOSPLIT leaf behaviour, the large-frame R30 prologue/epilogue forms
+// and the morestack block.  Relocation fields are masked.
 func TestStackGuardBytesLOONG64(t *testing.T) {
 	for _, tt := range []struct {
 		name string
@@ -199,9 +201,37 @@ func TestStackGuardBytesLOONG64(t *testing.T) {
 		{"leafsmall", "TEXT \u00b7leafsmall(SB), $16-0\n\tRET\n",
 			"61a0ff2963a0ff026100c0296360c0022000004c"},
 		{"leafmed", "TEXT \u00b7leafmed(SB), $256-0\n\tRET\n",
-			"d442c02878e0fd0294e21200801a004061e0fb2963e0fb026100c0296320c4022000004c3f00100000000000ffd7ff53"},
+			"d442c02878e0fd0294e21200801a004061e0fb2963e0fb026100c0296320c4022000004c3f00150000000000ffd7ff53"},
 		{"nosplit", "TEXT \u00b7nosplit(SB), NOSPLIT, $16-0\n\tRET\n",
 			"61a0ff2963a0ff026100c0296360c0022000004c"},
+		// The LR store leaves the 12-bit store-offset range while the SP
+		// adjust immediate still fits, and the epilogue adjusts through a
+		// single ORI.
+		{"fit2048", "TEXT \u00b7fit2048(SB), $2040-0\n\tRET\n",
+			"d442c0287800e20294e21200802600401e000014de8f1000c103e0296300e0026100c0291e00a00363f810002000004c3f00150000000000ffcbff53"},
+		// Medium class at the materialisation boundary (off = 2048 still
+		// immediate, 2049+ goes through R30).
+		{"med2048off", "TEXT \u00b7med2048off(SB), $2168-0\n\tRET\n",
+			"d442c0287800e00294e21200802e0040feffff15de8f1000c103de29feffff15de039e0363f810006100c0291e00a20363f810002000004c3f00150000000000ffc3ff53"},
+		{"medmat", "TEXT \u00b7medmat(SB), $2176-0\n\tRET\n",
+			"d442c028feffff15dee39f0378f8100094e21200802e0040feffff15de8f1000c1e3dd29feffff15dee39d0363f810006100c0291e20a20363f810002000004c3f00150000000000ffbbff53"},
+		// Big class with the rounding-split store and the floor-split adjust.
+		{"leafbig", "TEXT \u00b7leafbig(SB), $8192-0\n\tRET\n",
+			"d442c0283e000014de23be0378f8120000470044deffff15dee3810378f8100094e2120080320040deffff15de8f1000c1e3ff29beffff15dee3bf0363f810006100c0295e000014de23800363f810002000004c3f00150000000000ffa7ff53"},
+		// Zero low 12 bits drop the ORI from the store, the adjust and the
+		// epilogue materialisation.
+		{"bigzero", "TEXT \u00b7bigzero(SB), $4088-0\n\tRET\n",
+			"d442c028feffff15de03820378f8100094e21200802a0040feffff15de8f1000c103c029feffff1563f810006100c0293e00001463f810002000004c3f00150000000000ffbfff53"},
+		// Big class whose first constant has a zero high part: a single ORI.
+		{"big3976", "TEXT \u00b7big3976(SB), $4096-0\n\tRET\n",
+			"d442c0281e20be0378f8120000470044feffff15dee3810378f8100094e2120080320040feffff15de8f1000c1e3ff29deffff15dee3bf0363f810006100c0293e000014de23800363f810002000004c3f00150000000000ffabff53"},
+		// Big class at a multiple of 4096: both guard constants lose their
+		// ORI word.
+		{"giantlo0", "TEXT \u00b7giantlo0(SB), $4216-0\n\tRET\n",
+			"d442c0283e00001478f8120000430044feffff1578f8100094e2120080320040feffff15de8f1000c103fe29deffff15de03be0363f810006100c0293e000014de03820363f810002000004c3f00150000000000ffafff53"},
+		// Non-leaf big frame: the body call plus the LR restore epilogue.
+		{"callbig", "TEXT \u00b7callbig(SB), $8192-0\n\tCALL \u00b7other(SB)\n\tRET\nTEXT \u00b7other(SB), NOSPLIT, $0\n\tRET\n",
+			"d442c0283e000014de23be0378f81200004f0044deffff15dee3810378f8100094e21200803a0040deffff15de8f1000c1e3ff29beffff15dee3bf0363f810006100c029000000006100c0285e000014de23800363f810002000004c3f00150000000000ff9fff53"},
 	} {
 		f, errs := parser.Parse("g_loong64.s", tt.src)
 		if len(errs) > 0 {
@@ -224,3 +254,30 @@ func TestStackGuardBytesLOONG64(t *testing.T) {
 		}
 	}
 }
+
+// TestStackGuardGOObjInternalCall checks that GOOBJ emission succeeds when a
+// guarded function calls a TEXT symbol of the same file, for every arch's
+// call relocation kind.
+func TestStackGuardGOObjInternalCall(t *testing.T) {
+	for _, tt := range []struct {
+		src      string
+		assemble func(*ast.File) (*Image, error)
+	}{
+		{"g_amd64.s", AssembleFile},
+		{"g_arm64.s", AssembleFileARM64},
+		{"g_riscv64.s", AssembleFileRISCV},
+		{"g_loong64.s", AssembleFileLOONG64},
+	} {
+		f, errs := parser.Parse(tt.src, "TEXT \u00b7callsmall(SB), $16-0\n\tCALL \u00b7other(SB)\n\tRET\nTEXT \u00b7other(SB), NOSPLIT, $0\n\tRET\n")
+		if len(errs) > 0 {
+			t.Fatalf("%s: parse: %v", tt.src, errs)
+		}
+		img, err := tt.assemble(f)
+		if err != nil {
+			t.Fatalf("%s: assemble: %v", tt.src, err)
+		}
+		if _, err := img.GOObject("testpkg", tt.src); err != nil {
+			t.Errorf("%s: GOObject: %v", tt.src, err)
+		}
+	}
+}

asm/loong64_assemble.go

@@ -18,12 +18,10 @@ import (
 // then encoding with resolved branch targets).
 //
 // The emitted bytes match the Go toolchain's loong64 assembler, which is the
-// ground-truth oracle: prologue/epilogue, FP/SP frame mapping, branch
+// ground-truth oracle: prologue/epilogue (including the large-frame R30
-// encodings and the MOV immediate expansions all follow cmd/internal/obj/
+// materialisations), FP/SP frame mapping, the stack-split guard classes, and
-// loong64's asmout cases.  One deliberate difference: the stack-growth guard
+// branch encodings all follow cmd/internal/obj/loong64.  The morestack block
-// (the morestack check in the prologue and the call back into the runtime in
+// at the end of split functions carries the runtime.morestack_noctxt call.
-// the epilogue) is not emitted, so the bytes match only for NOSPLIT functions
-// or zero-frame leaves, where the toolchain emits no guard either.
 func assembleLOONG64(t *ast.Text) ([]byte, map[string]int, []Reloc, []LineEntry, []SpadjStep, error) {
 	fi := loong64ComputeFrame(t)
 	prologue := loong64Prologue(fi)
@@ -39,11 +37,12 @@ func assembleLOONG64(t *ast.Text) ([]byte, map[string]int, []Reloc, []LineEntry,
 	var relocs []Reloc
 	var spadj []SpadjStep
 
-	// The prologue (3 instructions when a frame is present) raises the SP
+	// The prologue raises the SP delta by autosize; the boundary is reported
-	// delta by autosize; the boundary is reported at the third instruction's
+	// after the SP adjust instruction, exactly as the toolchain's pctospadj
-	// pc, exactly as the toolchain's pctospadj does.
+	// does.  The prologue (3 instructions when a frame is present) may
+	// materialise its store or adjust through R30, which widens it.
 	if fi.autosize != 0 {
-		spadj = append(spadj, SpadjStep{PC: guardLen + 8, Value: fi.autosize})
+		spadj = append(spadj, SpadjStep{PC: guardLen + (loong64StoreWords(fi.autosize)+loong64AdjustWords(-int64(fi.autosize)))*4, Value: fi.autosize})
 	}
 
 	// Pass 1: label offsets from the instruction sizes.
@@ -99,14 +98,9 @@ func assembleLOONG64(t *ast.Text) ([]byte, map[string]int, []Reloc, []LineEntry,
 		preCount = len(relocs)
 		lines = append(lines, LineEntry{Offset: pc, Line: in.Pos().Line})
 		// The RET's epilogue closes the frame: the SP delta returns to zero
-		// after the addi.d (one instruction for a leaf, two for a non-leaf
+		// after the frame-deallocating ADDV.
-		// with the LR restore).
 		if strings.ToUpper(in.Mnemonic.Text) == "RET" && fi.autosize != 0 {
-			epi := 4
+			spadj = append(spadj, SpadjStep{PC: pc + loong64EpilogueWords(fi)*4, Value: 0})
-			if !fi.leaf {
-				epi = 8
-			}
-			spadj = append(spadj, SpadjStep{PC: pc + epi, Value: 0})
 		}
 		out = append(out, code...)
 		pc += len(code)
@@ -255,9 +249,9 @@ func encodeLOONG64Instr(instr *ast.Instr, pc int, offsets map[string]int, fi loo
 		}
 		return l64wordLE(uint32(immFromOperand(ops[0]))), nil
 	case "JMP", "B":
-		return encodeLOONG64Branch(instr, mnem, pc, offsets, false, resolve)
+		return encodeLOONG64Branch(instr, mnem, pc, offsets, false, resolve, relocs)
 	case "JAL", "CALL", "BL":
-		return encodeLOONG64Branch(instr, mnem, pc, offsets, true, resolve)
+		return encodeLOONG64Branch(instr, mnem, pc, offsets, true, resolve, relocs)
 	case "MOV", "MOVB", "MOVH", "MOVW", "MOVV", "MOVBU", "MOVHU", "MOVWU", "MOVF", "MOVD":
 		return encodeLOONG64Mov(instr, mnem, fi, relocs)
 	}
@@ -516,7 +510,7 @@ func encodeLOONG64Instr(instr *ast.Instr, pc int, offsets map[string]int, fi loo
 //
 //	JMP/B label      → b label            JMP/B (rj) → jirl r0, rj, 0
 //	JAL/CALL/BL label → bl label          JAL/CALL/BL (rj) → jirl r1, rj, 0
-func encodeLOONG64Branch(instr *ast.Instr, mnem string, pc int, offsets map[string]int, link bool, resolve func(string) string) ([]byte, error) {
+func encodeLOONG64Branch(instr *ast.Instr, mnem string, pc int, offsets map[string]int, link bool, resolve func(string) string, relocs *[]Reloc) ([]byte, error) {
 	if len(instr.Operands) != 1 {
 		return nil, fmt.Errorf("%s expects 1 operand, got %d", mnem, len(instr.Operands))
 	}
@@ -533,6 +527,14 @@ func encodeLOONG64Branch(instr *ast.Instr, mnem string, pc int, offsets map[stri
 		}
 		return l64wordLE(l64irr16(l64branchTable["JIRL"], 0, rj, rd)), nil
 	}
+	// Direct symbol: sym+off(SB) → bl with an R_CALLLOONG64 relocation (the
+	// linker fills the offset), as the toolchain does for CALL/BL/JAL.
+	if op.Addr.Sym != nil && op.Addr.Sym.Pseudo == "SB" && link {
+		if relocs != nil {
+			*relocs = append(*relocs, Reloc{Off: 0, After: 4, Name: op.Addr.Sym.Name, Kind: RelLoong64Branch, Addend: op.Addr.Sym.Offset})
+		}
+		return l64wordLE(l64bbl(l64jumpTable["BL"], 0)), nil
+	}
 	// Direct: label → b/bl.
 	target := resolve(l64Label(op))
 	targetOff, ok := offsets[target]

asm/loong64_frame.go

@@ -20,14 +20,20 @@ import (
 // (the toolchain aligns frames with `if autosize&4 != 0 { autosize += 4 }`).
 // A leaf function (no calls) with a zero frame gets no prologue at all.
 //
-// Prologue (autosize > 0), byte-identical to the toolchain:
+// Prologue (autosize > 0, small), byte-identical to the toolchain:
 //
 //	MOVV R1, -autosize(R3)   // save LR below the new SP (traceback-safe)
 //	ADDV $-autosize, R3      // open the frame
 //	MOVV R1, 0(R3)           // save LR again at SP (signal-safety)
 //
+// Large frames (autosize past the 12-bit offset or immediate ranges) expand
+// the store and the adjust through REGTMP (R30) exactly as the toolchain's
+// assembler does: the store via the rounding LU12IW split, the adjust via
+// the floor LU12IW/ORI split.
+//
 // Epilogue: MOVV 0(R3), R1; ADDV $autosize, R3 (non-leaf only for the LR
-// restore); the RET's jirl r0, r1, 0 follows.
+// restore; the adjust materialised when the immediate does not fit); the
+// RET's jirl r0, r1, 0 follows.
 
 // loong64FrameInfo holds the frame layout derived from a TEXT directive.
 type loong64FrameInfo struct {
@@ -84,62 +90,108 @@ func loong64ComputeFrame(t *ast.Text) loong64FrameInfo {
 
 // 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.
+// constants through R30; each materialisation shrinks by one word when the
+// constant's low 12 bits are zero.
 func loong64GuardLen(fi loong64FrameInfo) int {
 	if !fi.needSplit {
 		return 0
 	}
+	off := int64(fi.autosize - stackSmall)
 	switch fi.splitClass {
 	case 0:
 		return 12
 	case 1:
-		return 16
+		if off <= 2048 {
+			return 16 // ADDV $-off fits the signed 12-bit immediate
+		}
+		return 24 // MOVV + LU12IW + ORI + ADDV + SGTU + BEQ
 	default:
-		return 40 // MOVV + [LU12IW+ORI] + SGTU + BNE + [LU12IW+ORI] + ADDV + SGTU + BEQ
+		// MOVV + [mat] + SGTU + BNE + [mat] + ADDV + SGTU + BEQ
+		return (6 + loong64MatLen(off) + loong64MatLen(-off)) * 4
 	}
 }
 
-// loong64Lu12iOri materialises the 32-bit constant v in rd with the
+// loong64MatLen reports the word count of materialising v in R30: a value
-// toolchain's LU12IW/ORI pair (the ORI reads and writes rd itself).
+// with a zero high part needs only the ORI (the toolchain's MOVW $v, R30),
-func loong64Lu12iOri(rd int, v int64) []uint32 {
+// one with a zero low part only the LU12IW.
-	hi := int32(v >> 12)
+func loong64MatLen(v int64) int {
-	lo := int32(v & 0xFFF)
+	if v>>12 == 0 || v&0xFFF == 0 {
-	return []uint32{
+		return 1
-		0x0a<<25 | uint32(hi&0xFFFFF)<<5 | uint32(rd),
-		0x0e<<22 | uint32(lo)<<10 | uint32(rd)<<5 | uint32(rd),
 	}
+	return 2
 }
 
+// loong64MatWords appends the words that materialise v in R30, splitting it
+// as v>>12 plus the zero-extended low 12 bits.
+func loong64MatWords(ws []uint32, v int64) []uint32 {
+	hi := v >> 12
+	lo := v & 0xFFF
+	if hi == 0 {
+		return append(ws, l64irr(l64OriOp, int(v), 0, 30))
+	}
+	ws = append(ws, l64ir(l64Lu12iwOp, int(hi), 30))
+	if lo != 0 {
+		ws = append(ws, l64irr(l64OriOp, int(lo), 30, 30))
+	}
+	return ws
+}
+
+// The LU12IW and ORI opcode bases (2RI20 and 2RI12 formats); the ORI reads
+// and writes rd itself.
+const (
+	l64Lu12iwOp = 0x0a << 25
+	l64OriOp    = 0x0e << 22
+)
+
+// loong64Imm12 reports whether v fits a signed 12-bit immediate.
+func loong64Imm12(v int64) bool { return v >= -2048 && v <= 2047 }
+
 // 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.
+// branch displacements are in instructions and are computed from each
+// branch's own position.
 func loong64GuardBytes(fi loong64FrameInfo, blockStart int) []byte {
 	// MOVV 16(g), R20 (g.stackguard0), g = R22.
 	ws := []uint32{l64irr(l64loadStoreTable["MOVV"].ld, 16, 22, 20)}
+	off := int64(fi.autosize - stackSmall)
+	// beq appends BEQ R20, blockStart from the branch's own position.
+	beq := func() {
+		ws = append(ws, loong64Beqz(20, int32((blockStart-len(ws)*4)>>2)))
+	}
 	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)))
+		beq()
 	case 1:
-		off := int32(fi.autosize - stackSmall)
+		ws = append(ws, loong64MediumWords(off)...)
-		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)))
+		beq()
 	default:
-		off := int64(fi.autosize - stackSmall)
+		// SGTU $off, SP, R24 catches the SP underflow a huge frame would
-		movLen := 8 // LU12IW + ORI
+		// cause; BNE jumps to morestack in that case.
-		ws = append(ws, loong64Lu12iOri(30, off)...)
+		ws = append(ws, loong64MatWords(nil, off)...)
 		ws = append(ws, l64rrr(l64DualTable["SGTU"].rrr, 30, 3, 24))
-		ws = append(ws, loong64Bnez(24, int32((blockStart-(8+movLen))>>2)))
+		ws = append(ws, loong64Bnez(24, int32((blockStart-len(ws)*4)>>2)))
-		ws = append(ws, loong64Lu12iOri(30, -off)...)
+		ws = append(ws, loong64MatWords(nil, -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)))
+		beq()
 	}
 	return l64WordsLE(ws...)
 }
 
+// loong64MediumWords emits the medium-class stack check for offset off: the
+// ADDV immediate when it fits, otherwise the same sequence with the constant
+// materialised in R30.
+func loong64MediumWords(off int64) []uint32 {
+	if off <= 2048 {
+		return []uint32{l64irr(l64DualTable["ADDV"].imm, int(-off), 3, 24)}
+	}
+	ws := loong64MatWords(nil, -off)
+	return append(ws, l64rrr(l64DualTable["ADDV"].rrr, 30, 3, 24))
+}
+
 // 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 {
@@ -150,12 +202,13 @@ func loong64Bnez(rj int, dispInstr int32) uint32 {
 	return l64ir21(l64branch21Table["BNEZ"], int(dispInstr), rj)
 }
 
-// loong64MoreStackBlock emits the trailing block: MOVV R1, R31 (save LR),
+// loong64MoreStackBlock emits the trailing block: MOVV R1, R31 (save LR, the
-// BL runtime.morestack_noctxt, B back to the function entry.
+// toolchain's OR R1, R0, R31 expansion), 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)
+		l64rrr(l64DualTable["OR"].rrr, 0, 1, 31), // MOVV R1, R31 (OR R1, R0, R31)
-		l64bbl(l64jumpTable["BL"], 0),             // BL, patched by the linker
+		l64bbl(l64jumpTable["BL"], 0),            // BL, patched by the linker
 	}
 	disp := (-(blockStart + 8)) >> 2
 	ws = append(ws, l64bbl(l64jumpTable["B"], int(disp)))
@@ -185,17 +238,37 @@ func loong64IsLeaf(t *ast.Text) bool {
 	return true
 }
 
-// loong64Prologue returns the prologue bytes for a loong64 function.
+// loong64Prologue returns the prologue bytes for a loong64 function.  When
+// the LR store offset leaves the toolchain's 12-bit store range ([-2046,
+// 2045], BIG_12 = 2046) or the SP adjust immediate its 12-bit immediate
+// range, each switches to the R30 materialisation the assembler expands it
+// to: the store uses the rounding %hi/%lo split (LU12IW of (v+2048)>>12,
+// REGTMP += SP, store at the raw offset), the adjust the floor split
+// (LU12IW, ORI when the low part is non-zero, REGTMP += SP).
 func loong64Prologue(fi loong64FrameInfo) []byte {
 	if fi.autosize == 0 {
 		return nil
 	}
 	addiD := l64DualTable["ADDV"].imm
-	return l64WordsLE(
+	var ws []uint32
-		l64irr(l64loadStoreTable["MOVV"].st, -fi.autosize, 3, 1), // MOVV R1, -autosize(R3)
+	storeBase := 3
-		l64irr(addiD, -fi.autosize, 3, 3),                        // ADDV $-autosize, R3
+	if fi.autosize > 2046 {
-		l64irr(l64loadStoreTable["MOVV"].st, 0, 3, 1),            // MOVV R1, 0(R3)
+		// The store goes through REGTMP: LU12IW of the rounding split,
-	)
+		// REGTMP += SP, then the store at REGTMP with the truncated offset.
+		v := -int64(fi.autosize)
+		ws = append(ws, l64ir(l64Lu12iwOp, int((v+2048)>>12), 30))
+		ws = append(ws, l64rrr(l64DualTable["ADDV"].rrr, 3, 30, 30))
+		storeBase = 30
+	}
+	ws = append(ws, l64irr(l64loadStoreTable["MOVV"].st, -fi.autosize, storeBase, 1)) // MOVV R1, -autosize(base)
+	if loong64Imm12(-int64(fi.autosize)) {
+		ws = append(ws, l64irr(addiD, -fi.autosize, 3, 3)) // ADDV $-autosize, R3
+	} else {
+		ws = append(ws, loong64MatWords(nil, -int64(fi.autosize))...)
+		ws = append(ws, l64rrr(l64DualTable["ADDV"].rrr, 30, 3, 3))
+	}
+	ws = append(ws, l64irr(l64loadStoreTable["MOVV"].st, 0, 3, 1)) // MOVV R1, 0(R3)
+	return l64WordsLE(ws...)
 }
 
 // loong64Return returns the bytes for a RET: the epilogue (restore LR and
@@ -207,14 +280,46 @@ func loong64Return(fi loong64FrameInfo) []byte {
 			// MOVV 0(R3), R1, restore the link register.
 			ws = append(ws, l64irr(l64loadStoreTable["MOVV"].ld, 0, 3, 1))
 		}
-		// ADDV $autosize, R3, close the frame.
+		// ADDV $autosize, R3, close the frame (materialised when the
-		ws = append(ws, l64irr(l64DualTable["ADDV"].imm, fi.autosize, 3, 3))
+		// immediate does not fit).
+		if loong64Imm12(int64(fi.autosize)) {
+			ws = append(ws, l64irr(l64DualTable["ADDV"].imm, fi.autosize, 3, 3))
+		} else {
+			ws = append(ws, loong64MatWords(nil, int64(fi.autosize))...)
+			ws = append(ws, l64rrr(l64DualTable["ADDV"].rrr, 30, 3, 3))
+		}
 	}
 	// jirl r0, r1, 0, return.
 	ws = append(ws, l64irr16(l64branchTable["JIRL"], 0, 1, 0))
 	return l64WordsLE(ws...)
 }
 
+// loong64StoreWords reports the prologue word count of the LR store, and
+// loong64AdjustWords the word count of an SP adjust of v: the immediate
+// forms when they fit, otherwise the R30 materialisation sequences.
+func loong64StoreWords(autosize int) int {
+	if autosize > 2046 {
+		return 3
+	}
+	return 1
+}
+
+func loong64AdjustWords(v int64) int {
+	if loong64Imm12(v) {
+		return 1
+	}
+	return loong64MatLen(v) + 1
+}
+
+// loong64EpilogueWords reports the epilogue word count the RET expands to.
+func loong64EpilogueWords(fi loong64FrameInfo) int {
+	n := loong64AdjustWords(int64(fi.autosize))
+	if !fi.leaf {
+		n++
+	}
+	return n
+}
+
 // loong64ResolvePseudo translates a pseudo-register memory reference into a
 // hardware base register and offset.  x+N(FP) → (N + autosize + 8)(SP);
 // x-N(SP) → (autosize - N)(SP).  Returns base = -1 for an unresolvable