fix(verify): arm64 stack save, adaptive canary and host gating

Assisted-by: GLM 5.3

verify/abi.go

@@ -16,21 +16,51 @@ import (
 // canary below SP, returning both the argument block (with results) and an
 // ABIReport.
 //
+// The callee is assumed to declare no local frame ($0 in its TEXT
+// directive); use CallCheckedFrame, or Kernel.CallFuncChecked which reads
+// the frame from the image, for a callee with a frame.  A callee whose
+// frame extends past the fixed call margin would otherwise trip the canary
+// with perfectly legal writes.
+//
+// Not safe for concurrent use: only one JIT call may be in flight at a
+// time, the trampolines keep the saved registers in package globals.
+//
 // The architecture-specific parts live in abi_<arch>.s: enterJITChecked
 // plants sentinels in the registers the Go ABI fixes across calls (the
 // frame pointer and the goroutine pointer) before switching to the
 // prepared stack, and the raw return trampoline leaveJITCheckedRaw
 // compares them and records violations in abiResult.
 func CallChecked(fnAddr uintptr, args []byte) ([]byte, ABIReport, error) {
+	return CallCheckedFrame(fnAddr, args, 0)
+}
+
+// CallCheckedFrame is CallChecked with the canary gap sized for a callee
+// that declares a local frame of frame bytes: the canary window is placed
+// below the deepest write a legal kernel may make, its own frame plus the
+// call margin, so only writes that go below the declared frame trip it.
+// Callers that already hold the function layout pass asm.FuncLayout.Frame.
+//
+// Not safe for concurrent use: only one JIT call may be in flight at a
+// time, the trampolines keep the saved registers in package globals.
+func CallCheckedFrame(fnAddr uintptr, args []byte, frame int) ([]byte, ABIReport, error) {
 	report := ABIReport{}
 
 	// Reset the global result.
 	abiResult = 0
 
-	// Prepare the stack: [canary][padding][leaveJITCheckedRaw][args...]
+	// The gap between the canary window and the entry stack pointer must
-	// The canary sits below the initial SP, so the function would have to
+	// cover every write a legal kernel makes.  Two parts:
-	// write below SP to corrupt it.
+	//   - frame: the callee's declared local frame, which the ABI lets it
-	totalSize := redZoneSize + stackPad + 8 + len(args) + 64
+	//     write anywhere in [SP-frame, SP).
+	//   - stackPad (64): the call margin.  A kernel may CALL another
+	//     function, which pushes a return address below the frame and runs
+	//     a small prologue of its own; 64 bytes covers both.  Callees'
+	//     own frames are not accounted: a kernel calling deep into other
+	//     frames can write below this gap without detection.
+	pad := stackPad + frame
+
+	// Prepare the stack: [canary][frame gap][leaveJITCheckedRaw][args...]
+	totalSize := redZoneSize + pad + 8 + len(args) + 64
 	stackMem, err := syscall.Mmap(-1, 0, totalSize,
 		syscall.PROT_READ|syscall.PROT_WRITE, syscall.MAP_PRIVATE|syscall.MAP_ANON)
 	if err != nil {
@@ -43,8 +73,8 @@ func CallChecked(fnAddr uintptr, args []byte) ([]byte, ABIReport, error) {
 		stackMem[i] = redZoneFill
 	}
 
-	// Return address and args after the canary and padding.
+	// Return address and args after the canary and the frame gap.
-	retOff := redZoneSize + stackPad
+	retOff := redZoneSize + pad
 	binary.LittleEndian.PutUint64(stackMem[retOff:retOff+8], uint64(leaveCheckedPtr))
 	copy(stackMem[retOff+8:], args)
 

verify/abi_arm64.s

@@ -36,7 +36,8 @@ DATA ·leaveCheckedPtr(SB)/8, $·leaveJITCheckedRaw(SB)
 TEXT ·enterJITChecked(SB), NOSPLIT, $0-16
 	MOVD fn+0(FP), R0 // target (before SP switch)
 	MOVD R30, savedLR(SB) // save link register
-	MOVD R3, savedSP(SB) // save Go stack pointer
+	MOVD RSP, R3 // save Go stack pointer: R3 relays the value, the
+	MOVD R3, savedSP(SB) // arm64 assembler cannot store RSP to memory
 	MOVD R29, savedFP(SB) // save frame pointer (vet requires save before clobber)
 	MOVD g, savedG(SB) // save g
 	MOVD $SENTINEL_FP, R29 // sentinel in the frame pointer
@@ -85,8 +86,9 @@ restore:
 
 // Package-level storage for the saved frame pointer.  Like savedSP and
 // savedLR in trampoline_arm64.s, this is assembly-side state: the amd64
-// checked trampoline saves the caller's frame pointer for vet's sake and
+// checked trampoline saves and restores the caller's frame pointer too
-// never restores it, and this file mirrors that.
+// (leaveJITCheckedRaw restores BP before returning), and this file mirrors
+// that.
 GLOBL savedFP(SB), NOPTR, $8
 
 GLOBL savedG(SB), NOPTR, $8

verify/abi_loong64.s

@@ -64,3 +64,14 @@ g_ok:
 	JMP  (R1) // return to Go caller
 
 GLOBL savedG(SB), NOPTR, $8
+
+// leaveRawAddr holds the raw .abi0 address of leaveJIT, read by
+// call_loong64.go in preference to reflect.ValueOf(leaveJIT), which returns
+// the address of the ABIInternal wrapper the linker interposes: the
+// wrapper's prologue clobbers the saved-register window the JIT call
+// depends on.  It lives here rather than beside leaveJIT in
+// trampoline_loong64.s because that file is assembled by gasm itself in
+// the ground-truth byte tests, and gasm cannot yet encode a symbol-valued
+// DATA word.
+GLOBL ·leaveRawAddr(SB), NOPTR, $8
+DATA ·leaveRawAddr(SB)/8, $·leaveJIT(SB)

verify/abi_other.go

@@ -11,3 +11,8 @@ import "fmt"
 func CallChecked(fnAddr uintptr, args []byte) ([]byte, ABIReport, error) {
 	return nil, ABIReport{}, fmt.Errorf("verify: ABI checks are not supported on this architecture")
 }
+
+// CallCheckedFrame is unavailable on unsupported architectures.
+func CallCheckedFrame(fnAddr uintptr, args []byte, frame int) ([]byte, ABIReport, error) {
+	return nil, ABIReport{}, fmt.Errorf("verify: ABI checks are not supported on this architecture")
+}

verify/abi_report.go

@@ -47,7 +47,9 @@ func (r ABIReport) String() string {
 
 // redZoneSize is the canary window below the prepared stack pointer.  On
 // amd64 it is the System V red zone; on every architecture a Go function
-// must not write below SP, so any corruption there is a bug.
+// must not write below its own declared frame, so the canary sits below
+// the frame plus the call margin (see CallCheckedFrame) and any corruption
+// there is a bug.
 const redZoneSize = 128
 
 // redZoneFill is the byte pattern used to detect writes below SP.

verify/abi_test.go

@@ -4,11 +4,13 @@
 package verify
 
 import (
+	"os"
 	"testing"
 )
 
 func loadABIKernel(t *testing.T) *Kernel {
 	t.Helper()
+	requireHost(t, "amd64")
 	k, err := Load("../testdata/verify/abi_amd64.s")
 	if err != nil {
 		t.Fatalf("Load: %v", err)
@@ -93,3 +95,85 @@ func TestCallFuncCheckedErrors(t *testing.T) {
 		t.Fatal("expected error for too-small arg block")
 	}
 }
+
+// loadCanaryKernel writes an inline kernel pair that exercises the canary
+// geometry against declared frames: frameLocal owns a $96 local frame and
+// writes its lowest local (96 bytes below the entry stack pointer, well
+// past the 64-byte call margin a frame-0 kernel gets); belowFrame writes
+// 136 bytes below its own frame, deep into the canary window.
+func loadCanaryKernel(t *testing.T) *Kernel {
+	t.Helper()
+	requireHost(t, "amd64")
+	src := `#include "textflag.h"
+
+// func frameLocal(x int64) int64
+TEXT ·frameLocal(SB), NOSPLIT, $96-16
+	MOVQ  x+0(FP), AX
+	MOVQ  AX, l-96(SP)
+	MOVQ  l-96(SP), AX
+	MOVQ  AX, ret+8(FP)
+	RET
+
+// func belowFrame(x int64) int64
+TEXT ·belowFrame(SB), NOSPLIT, $96-16
+	MOVQ  $1, -136(SP)
+	MOVQ  x+0(FP), AX
+	MOVQ  AX, ret+8(FP)
+	RET
+`
+	file := t.TempDir() + "/canary_amd64.s"
+	if err := os.WriteFile(file, []byte(src), 0o644); err != nil {
+		t.Fatalf("write kernel: %v", err)
+	}
+	k, err := Load(file)
+	if err != nil {
+		t.Fatalf("Load: %v", err)
+	}
+	t.Cleanup(k.Close)
+	return k
+}
+
+// TestCallCheckedFrameLegal checks that a kernel whose declared frame
+// extends below the fixed 64-byte margin does not trip the canary: the
+// protected gap must adapt to the frame the TEXT directive declares.
+func TestCallCheckedFrameLegal(t *testing.T) {
+	k := loadCanaryKernel(t)
+
+	args := make([]byte, 16)
+	PutUint64(args, 0, 42)
+
+	out, report, err := k.CallFuncChecked("frameLocal", args)
+	if err != nil {
+		t.Fatalf("CallFuncChecked(frameLocal): %v", err)
+	}
+	if got := int64(GetUint64(out, 8)); got != 42 {
+		t.Errorf("frameLocal(42) = %d, want 42", got)
+	}
+	if report.RedZoneHit {
+		t.Error("frameLocal: writing its own $96 frame must not count as a red-zone hit")
+	}
+	if !report.OK() {
+		t.Errorf("frameLocal: %s", report)
+	}
+}
+
+// TestCallCheckedBelowFrameCaught checks the other side of the adaptive
+// gap: a kernel that writes below its own frame by more than the call
+// margin must still be reported.
+func TestCallCheckedBelowFrameCaught(t *testing.T) {
+	k := loadCanaryKernel(t)
+
+	args := make([]byte, 16)
+	PutUint64(args, 0, 42)
+
+	out, report, err := k.CallFuncChecked("belowFrame", args)
+	if err != nil {
+		t.Fatalf("CallFuncChecked(belowFrame): %v", err)
+	}
+	if got := int64(GetUint64(out, 8)); got != 42 {
+		t.Errorf("belowFrame(42) = %d, want 42", got)
+	}
+	if !report.RedZoneHit {
+		t.Error("belowFrame: expected RedZoneHit for a write below the declared frame")
+	}
+}

verify/buffers.go

@@ -20,7 +20,11 @@ type BufSpec struct {
 }
 
 // ParseBufSpec parses a "name:size:pattern[,name:size:pattern]" spec string
-// into individual buffer specs.  Empty input yields an empty slice.
+// into individual buffer specs.  Empty input yields an empty slice.  The
+// size must be a plain decimal number over the whole field and the pattern
+// must be a known name or a valid hex blob, so typos fail here with the
+// offending spec in the message rather than silently allocating a zeroed
+// buffer.
 func ParseBufSpec(spec string) ([]BufSpec, error) {
 	if spec == "" {
 		return nil, nil
@@ -31,15 +35,38 @@ func ParseBufSpec(spec string) ([]BufSpec, error) {
 		if len(fields) != 3 {
 			return nil, fmt.Errorf("verify: invalid buffer spec %q (expected name:size:pattern)", part)
 		}
-		var size int
+		// strconv.Atoi parses the whole field, unlike fmt.Sscanf which
-		if _, err := fmt.Sscanf(fields[1], "%d", &size); err != nil || size <= 0 {
+		// accepts trailing garbage ("1024abc" parsed as 1024).
+		size, err := strconv.Atoi(fields[1])
+		if err != nil || size <= 0 {
 			return nil, fmt.Errorf("verify: invalid buffer size %q in %q", fields[1], part)
 		}
+		if err := validPattern(fields[2]); err != nil {
+			return nil, fmt.Errorf("verify: invalid pattern %q in %q: %v", fields[2], part, err)
+		}
 		out = append(out, BufSpec{Name: fields[0], Size: size, Pattern: fields[2]})
 	}
 	return out, nil
 }
 
+// validPattern checks one buffer pattern: a known name, or a non-empty hex
+// blob.  fillBuffer enforces the same rule again at fill time for specs
+// that were not built by ParseBufSpec.
+func validPattern(pattern string) error {
+	switch pattern {
+	case "zero", "ones", "seq":
+		return nil
+	}
+	data, err := hex.DecodeString(pattern)
+	if err != nil {
+		return fmt.Errorf("not a known pattern (zero, ones, seq) and not hex: %w", err)
+	}
+	if len(data) == 0 {
+		return fmt.Errorf("not a known pattern (zero, ones, seq) and the hex blob is empty")
+	}
+	return nil
+}
+
 // allocatedBuf is one live buffer in a pool.
 type allocatedBuf struct {
 	spec BufSpec
@@ -60,11 +87,15 @@ type BufPool struct {
 }
 
 // Alloc allocates and fills the buffers described by specs.  The returned
-// pool must be kept alive until every call using it has returned.
+// pool must be kept alive until every call using it has returned.  An
+// unknown pattern name or invalid hex blob is an error rather than a
+// silently zeroed buffer.
 func (p *BufPool) Alloc(specs []BufSpec) error {
 	for _, s := range specs {
 		data := make([]byte, s.Size+safetyMargin)
-		fillBuffer(data, s.Pattern)
+		if err := fillBuffer(data, s.Pattern); err != nil {
+			return fmt.Errorf("verify: buffer %q: %w", s.Name, err)
+		}
 		p.bufs = append(p.bufs, allocatedBuf{spec: s, data: data})
 	}
 	return nil
@@ -128,8 +159,10 @@ func (p *BufPool) matchBuf(name string) *allocatedBuf {
 }
 
 // fillBuffer fills buf with the named pattern: "zero" (no-op, already zeroed),
-// "ones" (0xFF), "seq" (i mod 256), or a hex blob repeated to fill.
+// "ones" (0xFF), "seq" (i mod 256), or a hex blob repeated to fill.  An
-func fillBuffer(buf []byte, pattern string) {
+// unknown pattern name or undecodable hex returns an error instead of
+// leaving the buffer silently zeroed.
+func fillBuffer(buf []byte, pattern string) error {
 	switch pattern {
 	case "zero":
 		// Already zeroed by make.
@@ -142,12 +175,18 @@ func fillBuffer(buf []byte, pattern string) {
 			buf[i] = byte(i)
 		}
 	default:
-		if data, err := hex.DecodeString(pattern); err == nil && len(data) > 0 {
+		data, err := hex.DecodeString(pattern)
-			for i := range buf {
+		if err != nil {
-				buf[i] = data[i%len(data)]
+			return fmt.Errorf("unknown pattern (want zero, ones, seq or hex): %w", err)
-			}
+		}
+		if len(data) == 0 {
+			return fmt.Errorf("unknown pattern (want zero, ones, seq or hex): empty hex blob")
+		}
+		for i := range buf {
+			buf[i] = data[i%len(data)]
 		}
 	}
+	return nil
 }
 
 // ApplyScalarArgs writes user-supplied scalar argument values into an ABI0

verify/buffers_test.go

@@ -115,6 +115,48 @@ func TestParseBufSpec(t *testing.T) {
 			t.Error("ParseBufSpec(\"dst:0:zero\") should error on zero size")
 		}
 	})
+
+	t.Run("trailing-garbage-size", func(t *testing.T) {
+		// fmt.Sscanf("%d") accepted "1024abc" as 1024; the strict parse
+		// must reject the whole field.
+		if _, err := ParseBufSpec("dst:1024abc:zero"); err == nil {
+			t.Error("ParseBufSpec(\"dst:1024abc:zero\") should reject a size with trailing garbage")
+		}
+	})
+
+	t.Run("hex-size", func(t *testing.T) {
+		if _, err := ParseBufSpec("dst:0x40:zero"); err == nil {
+			t.Error("ParseBufSpec(\"dst:0x40:zero\") should reject a non-decimal size")
+		}
+	})
+
+	t.Run("unknown-pattern", func(t *testing.T) {
+		if _, err := ParseBufSpec("dst:64:wibble"); err == nil {
+			t.Error("ParseBufSpec(\"dst:64:wibble\") should reject an unknown pattern")
+		}
+	})
+
+	t.Run("bad-hex-pattern", func(t *testing.T) {
+		if _, err := ParseBufSpec("dst:64:zz"); err == nil {
+			t.Error("ParseBufSpec(\"dst:64:zz\") should reject undecodable hex")
+		}
+	})
+
+	t.Run("empty-pattern", func(t *testing.T) {
+		if _, err := ParseBufSpec("dst:64:"); err == nil {
+			t.Error("ParseBufSpec(\"dst:64:\") should reject an empty pattern")
+		}
+	})
+
+	t.Run("hex-pattern", func(t *testing.T) {
+		specs, err := ParseBufSpec("dst:6:deadbeef")
+		if err != nil {
+			t.Fatal(err)
+		}
+		if len(specs) != 1 || specs[0].Pattern != "deadbeef" {
+			t.Errorf("ParseBufSpec(\"dst:6:deadbeef\") = %+v", specs)
+		}
+	})
 }
 
 func TestBufPoolBuildArgs(t *testing.T) {

verify/call_amd64.go

@@ -58,6 +58,10 @@ const stackPad = 64
 //
 // The function must be NOSPLIT (no stack growth) and must not reference
 // external symbols, the image is self-contained.
+//
+// Not safe for concurrent use: only one JIT call may be in flight at a
+// time, the trampolines keep the saved registers in package globals
+// (savedSP and friends).
 func Call(fnAddr uintptr, args []byte) ([]byte, error) {
 	// Prepare the stack: [padding][leaveJIT addr][args...]
 	stackSize := stackPad + 8 + len(args) + 64 // padding + ret + args + safety

verify/call_arm64.go

@@ -34,6 +34,9 @@ const stackPad = 64
 // Call invokes the assembled function at fnAddr with the given ABI0 argument
 // block.  On arm64, the ABI0 convention places arguments on the stack starting
 // at [SP+8] (same as amd64 ABI0 in this project).
+//
+// Not safe for concurrent use: only one JIT call may be in flight at a
+// time, the trampolines keep the saved registers in package globals.
 func Call(fnAddr uintptr, args []byte) ([]byte, error) {
 	stackSize := stackPad + 8 + len(args) + 64
 	stackMem, err := syscall.Mmap(-1, 0, stackSize,

verify/call_loong64.go

@@ -8,7 +8,6 @@ package verify
 import (
 	"encoding/binary"
 	"fmt"
-	"reflect"
 	"syscall"
 	"unsafe"
 )
@@ -19,15 +18,20 @@ func enterJIT(fn uintptr, stack uintptr)
 //go:nosplit
 func leaveJIT()
 
-var leaveJITAddr uintptr
+// leaveRawAddr is the raw ABI0 address of leaveJIT, handed over by the
+// GLOBL/DATA in abi_loong64.s (reflect would return the interposed
+// ABIInternal wrapper instead: its prologue clobbers the saved-register
+// window the JIT call depends on).
+var leaveRawAddr uintptr
 
-func init() {
+var leaveJITAddr = leaveRawAddr
-	leaveJITAddr = reflect.ValueOf(leaveJIT).Pointer()
-}
 
 const stackPad = 64
 
 // Call invokes the assembled function at fnAddr with the given ABI0 argument block.
+//
+// Not safe for concurrent use: only one JIT call may be in flight at a
+// time, the trampolines keep the saved registers in package globals.
 func Call(fnAddr uintptr, args []byte) ([]byte, error) {
 	stackSize := stackPad + 8 + len(args) + 64
 	stackMem, err := syscall.Mmap(-1, 0, stackSize,

verify/call_other.go

@@ -7,7 +7,7 @@ package verify
 
 import "fmt"
 
-// Call is unavailable on non-amd64 architectures.
+// Call is unavailable off the four supported architectures.
 func Call(fnAddr uintptr, args []byte) ([]byte, error) {
-	return nil, fmt.Errorf("verify: JIT execution requires amd64")
+	return nil, fmt.Errorf("verify: JIT execution requires amd64, arm64, riscv64 or loong64")
 }

verify/call_riscv64.go

@@ -28,6 +28,9 @@ func init() {
 const stackPad = 64
 
 // Call invokes the assembled function at fnAddr with the given ABI0 argument block.
+//
+// Not safe for concurrent use: only one JIT call may be in flight at a
+// time, the trampolines keep the saved registers in package globals.
 func Call(fnAddr uintptr, args []byte) ([]byte, error) {
 	stackSize := stackPad + 8 + len(args) + 64
 	stackMem, err := syscall.Mmap(-1, 0, stackSize,

verify/corpus_test.go

@@ -4,8 +4,8 @@
 package verify
 
 import (
+	"encoding/hex"
 	"encoding/json"
-	"math/rand"
 	"os"
 	"runtime"
 	"strconv"
@@ -14,6 +14,7 @@ import (
 
 func loadBasicKernel(t *testing.T) *Kernel {
 	t.Helper()
+	requireHost(t, "amd64")
 	k, err := Load("../testdata/verify/basic_amd64.s")
 	if err != nil {
 		t.Fatalf("Load: %v", err)
@@ -71,7 +72,7 @@ func TestGenDualArgsEntryReplayable(t *testing.T) {
 	if !ok {
 		t.Fatal("parseFuncSig failed")
 	}
-	_, _, bufs, entry := genDualArgs(rand.New(rand.NewSource(1)), sig, 24)
+	_, _, bufs, entry := genDualArgs(newRNG(1), sig, 24)
 	if len(entry.Args) != 2 || entry.Args[0].Kind != "int" {
 		t.Fatalf("unexpected entry: %+v", entry)
 	}
@@ -96,11 +97,88 @@ func entryInt(t *testing.T, a CorpusArg) int64 {
 	return int64(v)
 }
 
+// strProbeSrc is a kernel that consumes an ABI0 string header: it
+// dereferences the data pointer (proving it points at live memory) and
+// returns len(s) + s[0] when the string is non-empty, len(s) otherwise.
+const strProbeSrc = `#include "textflag.h"
+
+// func strProbe(s string) int64
+TEXT ·strProbe(SB), NOSPLIT, $0-24
+	MOVQ  s_base+0(FP), SI
+	MOVQ  s_len+8(FP), CX
+	XORQ  AX, AX
+	TESTQ CX, CX
+	JZ    probe_done
+	MOVB  (SI), AL
+	ADDQ  AX, CX
+
+probe_done:
+	MOVQ  CX, ret+16(FP)
+	RET
+`
+
+func loadStrProbeKernel(t *testing.T) *Kernel {
+	t.Helper()
+	requireHost(t, "amd64")
+	file := t.TempDir() + "/strprobe_amd64.s"
+	if err := os.WriteFile(file, []byte(strProbeSrc), 0o644); err != nil {
+		t.Fatalf("write kernel: %v", err)
+	}
+	k, err := Load(file)
+	if err != nil {
+		t.Fatalf("Load: %v", err)
+	}
+	t.Cleanup(k.Close)
+	return k
+}
+
+// TestStringParamRoundTrip pins the ABI0 string marshalling end to end:
+// genDualArgs lays a string parameter out as a two-word header pointing at
+// a live buffer, the corpus entry records it, and ReplayEntry rebuilds an
+// equivalent header.
+func TestStringParamRoundTrip(t *testing.T) {
+	k := loadStrProbeKernel(t)
+
+	sig, ok := parseFuncSig("// func strProbe(s string) int64")
+	if !ok {
+		t.Fatal("parseFuncSig failed")
+	}
+	args, _, bufs, entry := genDualArgs(newRNG(7), sig, 24)
+	if len(entry.Args) != 1 || entry.Args[0].Kind != "string" {
+		t.Fatalf("unexpected entry: %+v", entry)
+	}
+
+	out, err := k.CallFunc("strProbe", args)
+	if err != nil {
+		t.Fatalf("CallFunc: %v", err)
+	}
+	live := int64(GetUint64(out, 16))
+	want := int64(entry.Args[0].Len)
+	if d, err := hex.DecodeString(entry.Args[0].Data); err != nil {
+		t.Fatalf("entry data: %v", err)
+	} else if len(d) > 0 {
+		want += int64(d[0])
+	}
+	if live != want {
+		t.Errorf("live call = %d, want %d (len + first byte)", live, want)
+	}
+
+	replayed, err := k.ReplayEntry("strProbe", entry)
+	if err != nil {
+		t.Fatalf("ReplayEntry: %v", err)
+	}
+	if got := int64(GetUint64(replayed, 16)); got != want {
+		t.Errorf("replayed call = %d, want %d", got, want)
+	}
+	runtime.KeepAlive(bufs)
+}
+
 // TestFuzzHookSavesFailures fuzzes add against the go-tool-asm build of a
 // sub kernel with the same signature, so every iteration mismatches (safely:
 // both kernels read only their own arguments) and the hook must record
 // replayable entries.
 func TestFuzzHookSavesFailures(t *testing.T) {
+	requireHost(t, "amd64")
 	src := `#include "textflag.h"
 
 // func add(a, b int) int

verify/coverage.go

@@ -5,7 +5,7 @@ package verify
 
 import (
 	"fmt"
-	"sort"
+	"slices"
 )
 
 // Block describes one basic block within a function: a maximal sequence of
@@ -13,7 +13,7 @@ import (
 // a single exit (a jump, conditional jump or RET).
 type Block struct {
 	Offset int    // byte offset within the function
-	Label  string // label name ("" for the entry block)
+	Label  string // label name ("(entry)" for the function entry block)
 }
 
 // Blocks identifies the basic blocks of a function from its local labels.
@@ -40,7 +40,7 @@ func (k *Kernel) Blocks(name string) ([]Block, error) {
 	for off := range offToLabel {
 		offsets = append(offsets, off)
 	}
-	sort.Ints(offsets)
+	slices.Sort(offsets)
 	for _, off := range offsets {
 		blocks = append(blocks, Block{Offset: off, Label: offToLabel[off]})
 	}

verify/fuzz.go

@@ -6,7 +6,7 @@ package verify
 import (
 	"encoding/hex"
 	"fmt"
-	"math/rand"
+	"math/rand/v2"
 	"regexp"
 	"runtime"
 	"strconv"
@@ -14,6 +14,28 @@ import (
 	"unsafe"
 )
 
+// newRNG builds the deterministic generator for a fuzz seed.  PCG seeds
+// with two 64-bit words; deriving the second from the first keeps one seed
+// one stream and rules out the all-zero seed.  The sequences differ from
+// the retired math/rand ones for the same seed, but remain reproducible
+// run to run, which is the property the fuzzers rely on.
+func newRNG(seed int64) *rand.Rand {
+	lo := uint64(seed)
+	return rand.New(rand.NewPCG(lo, ^lo))
+}
+
+// fillRandom fills buf from rng, eight bytes per draw.  math/rand/v2
+// dropped Read from *rand.Rand, and this loop keeps the byte sequence a
+// pure function of the generator state.
+func fillRandom(rng *rand.Rand, buf []byte) {
+	for off := 0; off < len(buf); off += 8 {
+		v := rng.Uint64()
+		for j := 0; j < 8 && off+j < len(buf); j++ {
+			buf[off+j] = byte(v >> (8 * uint(j)))
+		}
+	}
+}
+
 // FuzzResult reports the outcome of a differential fuzz campaign for one
 // function.
 type FuzzResult struct {
@@ -43,9 +65,9 @@ func (r FuzzResult) String() string {
 
 // CorpusArg is one replayable argument of a corpus entry.
 type CorpusArg struct {
-	Kind  string `json:"kind"`            // "slice", "ptr", "int", "scalar"
+	Kind  string `json:"kind"`            // "slice", "string", "ptr", "int", "scalar"
-	Len   int    `json:"len,omitempty"`   // slice: declared length in elements
+	Len   int    `json:"len,omitempty"`   // slice: declared length in elements; string: length in bytes
-	Data  string `json:"data,omitempty"`  // slice/ptr: hex-encoded buffer content
+	Data  string `json:"data,omitempty"`  // slice/string/ptr: hex-encoded buffer content
 	Value string `json:"value,omitempty"` // int/scalar: decimal value
 }
 
@@ -175,6 +197,9 @@ func ExtractSignatures(src string) map[string]funcSig {
 //
 // The signature comment must appear immediately above the TEXT directive
 // in the source (the conventional Go assembly layout).
+//
+// Not safe for concurrent use: only one JIT call may be in flight at a
+// time, the trampolines keep the saved registers in package globals.
 func (k *Kernel) FuzzFunc(name string, sig funcSig, goCode []byte, iterations int, seed int64) FuzzResult {
 	return k.FuzzFuncHook(name, sig, goCode, iterations, seed, nil)
 }
@@ -182,10 +207,13 @@ func (k *Kernel) FuzzFunc(name string, sig funcSig, goCode []byte, iterations in
 // FuzzFuncHook is FuzzFunc with a hook invoked for every failing input (a
 // crash or a mismatch), receiving a replayable corpus entry.  A nil hook
 // behaves exactly like FuzzFunc.
+//
+// Not safe for concurrent use: only one JIT call may be in flight at a
+// time, the trampolines keep the saved registers in package globals.
 func (k *Kernel) FuzzFuncHook(name string, sig funcSig, goCode []byte, iterations int, seed int64, onSave func(CorpusEntry)) FuzzResult {
 	result := FuzzResult{Func: name, Iterations: iterations}
 
-	rng := rand.New(rand.NewSource(seed))
+	rng := newRNG(seed)
 
 	// Map the Go-assembled code into a second executable region.
 	goExec, err := Map(goCode)
@@ -203,6 +231,18 @@ func (k *Kernel) FuzzFuncHook(name string, sig funcSig, goCode []byte, iteration
 		return result
 	}
 
+	// The result comparison below slices the parameter area off the
+	// argument block; a // func comment declaring more parameter bytes
+	// than the TEXT frame carries would slice past its end and panic.
+	// Fail the whole campaign with a clear message instead.
+	if ps := paramsSize(sig); ps > fl.Args {
+		result.Mismatches = iterations
+		result.FirstFail = fmt.Sprintf(
+			"signature declares %d parameter bytes, but the TEXT frame of %s carries %d argument bytes",
+			ps, name, fl.Args)
+		return result
+	}
+
 	for i := range iterations {
 		// Generate inputs and build TWO independent arg blocks (one per
 		// version) so that functions which write to their arguments
@@ -276,7 +316,7 @@ func genDualArgs(rng *rand.Rand, sig funcSig, argSize int) (gasmArgs, goArgs []b
 		switch {
 		case strings.HasPrefix(p.typ, "[]"):
 			elemSize := elemSizeFor(p.typ)
-			n := 1 + rng.Intn(127)
+			n := 1 + rng.IntN(127)
 			var declaredLen int
 			if sliceIdx == 0 {
 				declaredLen = n
@@ -290,7 +330,7 @@ func genDualArgs(rng *rand.Rand, sig funcSig, argSize int) (gasmArgs, goArgs []b
 			// Two independent buffers with identical random content.
 			buf1 := make([]byte, bufBytes)
 			buf2 := make([]byte, bufBytes)
-			rng.Read(buf1[:n*elemSize])
+			fillRandom(rng, buf1[:n*elemSize])
 			copy(buf2, buf1)
 			bufs = append(bufs, buf1, buf2)
 			putPtr(gasmArgs, off, unsafe.Pointer(&buf1[0]))
@@ -309,13 +349,35 @@ func genDualArgs(rng *rand.Rand, sig funcSig, argSize int) (gasmArgs, goArgs []b
 				Data: hex.EncodeToString(buf1[:n*elemSize]),
 			})
 
+		case p.typ == "string":
+			// An ABI0 string is a two-word header (data pointer +
+			// length); a random pointer would fault kernels that read
+			// the string, so the header points at a real buffer with
+			// the same safety margin slices get.
+			n := 1 + rng.IntN(127)
+			buf1 := make([]byte, n+8192)
+			buf2 := make([]byte, n+8192)
+			fillRandom(rng, buf1[:n])
+			copy(buf2, buf1)
+			bufs = append(bufs, buf1, buf2)
+			putPtr(gasmArgs, off, unsafe.Pointer(&buf1[0]))
+			putPtr(goArgs, off, unsafe.Pointer(&buf2[0]))
+			putU64(gasmArgs, off+8, uint64(n))
+			putU64(goArgs, off+8, uint64(n))
+			off += 16
+			entry.Args = append(entry.Args, CorpusArg{
+				Kind: "string",
+				Len:  n,
+				Data: hex.EncodeToString(buf1[:n]),
+			})
+
 		case strings.HasPrefix(p.typ, "*["):
 			nElem := arrayLen(p.typ)
 			elem := elemSizeFor("[]" + p.typ[strings.Index(p.typ, "]")+1:])
 			size := max(nElem*elem, 8)
 			buf1 := make([]byte, size)
 			buf2 := make([]byte, size)
-			rng.Read(buf1)
+			fillRandom(rng, buf1)
 			copy(buf2, buf1)
 			bufs = append(bufs, buf1, buf2)
 			putPtr(gasmArgs, off, unsafe.Pointer(&buf1[0]))
@@ -327,12 +389,24 @@ func genDualArgs(rng *rand.Rand, sig funcSig, argSize int) (gasmArgs, goArgs []b
 			})
 
 		case p.typ == "int" || p.typ == "uint" || p.typ == "int64" || p.typ == "uint64":
-			v := uint64(rng.Intn(256))
+			v := uint64(rng.IntN(256))
 			putU64(gasmArgs, off, v)
 			putU64(goArgs, off, v)
 			off += 8
 			entry.Args = append(entry.Args, CorpusArg{Kind: "int", Value: strconv.FormatUint(v, 10)})
 
+		case p.typ == "complex64", p.typ == "complex128":
+			// complex64 is two float32s (8 bytes), complex128 two
+			// float64s (16): plain data words to the marshaller, one
+			// corpus scalar per word so replay rebuilds them exactly.
+			for range paramSize(p.typ) / 8 {
+				v := rng.Uint64()
+				putU64(gasmArgs, off, v)
+				putU64(goArgs, off, v)
+				off += 8
+				entry.Args = append(entry.Args, CorpusArg{Kind: "scalar", Value: strconv.FormatUint(v, 10)})
+			}
+
 		default:
 			v := rng.Uint64()
 			putU64(gasmArgs, off, v)
@@ -346,8 +420,12 @@ func genDualArgs(rng *rand.Rand, sig funcSig, argSize int) (gasmArgs, goArgs []b
 
 // ReplayEntry rebuilds the argument block of a corpus entry and invokes the
 // named function once, returning the argument block after the call.  Slice
-// buffers get the same safety padding the fuzzer uses, so over-reads that
+// and string buffers get the same safety padding the fuzzer uses, so
-// were harmless during the original run stay harmless on replay.
+// over-reads that were harmless during the original run stay harmless on
+// replay.
+//
+// Not safe for concurrent use: only one JIT call may be in flight at a
+// time, the trampolines keep the saved registers in package globals.
 func (k *Kernel) ReplayEntry(name string, e CorpusEntry) ([]byte, error) {
 	fl, err := k.Func(name)
 	if err != nil {
@@ -374,6 +452,21 @@ func (k *Kernel) ReplayEntry(name string, e CorpusEntry) ([]byte, error) {
 			putU64(args, off+16, uint64(a.Len))
 			off += 24
 
+		case "string":
+			data, err := hex.DecodeString(a.Data)
+			if err != nil {
+				return nil, fmt.Errorf("corpus: string data: %w", err)
+			}
+			buf := make([]byte, len(data)+8192)
+			copy(buf, data)
+			bufs = append(bufs, buf)
+			if off+16 > len(args) {
+				return nil, fmt.Errorf("corpus: entry does not fit the argument block of %s", name)
+			}
+			putPtr(args, off, unsafe.Pointer(&buf[0]))
+			putU64(args, off+8, uint64(a.Len))
+			off += 16
+
 		case "ptr":
 			data, err := hex.DecodeString(a.Data)
 			if err != nil {
@@ -431,6 +524,12 @@ func paramsSize(sig funcSig) int {
 			size += 8 // pointer
 		case p.typ == "bool":
 			size += 1
+		case p.typ == "string":
+			size += 16 // data pointer + length
+		case p.typ == "complex64":
+			size += 8 // two float32s
+		case p.typ == "complex128":
+			size += 16 // two float64s
 		default:
 			size += 8 // int, uint, etc.
 		}

verify/fuzz_extra_test.go

@@ -90,6 +90,16 @@ func TestParamsSize(t *testing.T) {
 	if got := paramsSize(sig); got != 32 {
 		t.Errorf("paramsSize = %d, want 32 (24 for slice + 8 for int)", got)
 	}
+
+	// ABI0 strings are ptr+len (16), complex64 packs two float32s (8),
+	// complex128 two float64s (16).
+	sig = funcSig{
+		name:   "hdr",
+		params: []param{{name: "s", typ: "string"}, {name: "c64", typ: "complex64"}, {name: "c128", typ: "complex128"}},
+	}
+	if got := paramsSize(sig); got != 40 {
+		t.Errorf("paramsSize = %d, want 40 (16 string + 8 complex64 + 16 complex128)", got)
+	}
 }
 
 func TestBlockCount(t *testing.T) {
@@ -103,11 +113,34 @@ func TestBlockCount(t *testing.T) {
 	}
 }
 
+// TestNewRNGDeterministic pins the reproducibility contract of the fuzz
+// seeds: the same seed must rebuild the same PCG stream, including the
+// sub-word tail of fillRandom.
+func TestNewRNGDeterministic(t *testing.T) {
+	draw := func() []byte {
+		rng := newRNG(7)
+		out := make([]byte, 20) // 8+8+4: exercises a full word and a tail
+		fillRandom(rng, out)
+		return out
+	}
+	a, b := draw(), draw()
+	for i := range a {
+		if a[i] != b[i] {
+			t.Fatalf("seed 7 produced different bytes at %d: %02x vs %02x", i, a[i], b[i])
+		}
+	}
+	if newRNG(0) == nil {
+		t.Fatal("newRNG(0) must build a generator")
+	}
+}
+
 func TestFillBuffer(t *testing.T) {
 	t.Run("zero", func(t *testing.T) {
 		// fillBuffer("zero") is a no-op; relies on make already zeroing.
 		buf := make([]byte, 16)
-		fillBuffer(buf, "zero")
+		if err := fillBuffer(buf, "zero"); err != nil {
+			t.Fatalf("fillBuffer(zero): %v", err)
+		}
 		for _, b := range buf {
 			if b != 0 {
 				t.Error("zero pattern: make should produce zeroed buffer")
@@ -117,7 +150,9 @@ func TestFillBuffer(t *testing.T) {
 	})
 	t.Run("ones", func(t *testing.T) {
 		buf := make([]byte, 16)
-		fillBuffer(buf, "ones")
+		if err := fillBuffer(buf, "ones"); err != nil {
+			t.Fatalf("fillBuffer(ones): %v", err)
+		}
 		for _, b := range buf {
 			if b != 0xFF {
 				t.Error("ones pattern should fill with 0xFF")
@@ -127,7 +162,9 @@ func TestFillBuffer(t *testing.T) {
 	})
 	t.Run("seq", func(t *testing.T) {
 		buf := make([]byte, 256)
-		fillBuffer(buf, "seq")
+		if err := fillBuffer(buf, "seq"); err != nil {
+			t.Fatalf("fillBuffer(seq): %v", err)
+		}
 		for i, b := range buf {
 			if b != byte(i) {
 				t.Errorf("seq[%d] = %d, want %d", i, b, i)
@@ -137,7 +174,9 @@ func TestFillBuffer(t *testing.T) {
 	})
 	t.Run("hex", func(t *testing.T) {
 		buf := make([]byte, 6)
-		fillBuffer(buf, "deadbeef")
+		if err := fillBuffer(buf, "deadbeef"); err != nil {
+			t.Fatalf("fillBuffer(deadbeef): %v", err)
+		}
 		want := []byte{0xDE, 0xAD, 0xBE, 0xEF, 0xDE, 0xAD}
 		for i, b := range buf {
 			if b != want[i] {
@@ -146,6 +185,34 @@ func TestFillBuffer(t *testing.T) {
 			}
 		}
 	})
+	t.Run("unknown-pattern", func(t *testing.T) {
+		buf := make([]byte, 8)
+		if err := fillBuffer(buf, "wibble"); err == nil {
+			t.Error("fillBuffer(wibble): expected an error for an unknown pattern name")
+		}
+	})
+	t.Run("bad-hex", func(t *testing.T) {
+		buf := make([]byte, 8)
+		if err := fillBuffer(buf, "zz"); err == nil {
+			t.Error("fillBuffer(zz): expected an error for undecodable hex")
+		}
+	})
+	t.Run("empty-hex", func(t *testing.T) {
+		buf := make([]byte, 8)
+		if err := fillBuffer(buf, ""); err == nil {
+			t.Error("fillBuffer(\"\"): expected an error for an empty pattern")
+		}
+	})
+}
+
+// TestBufPoolAllocBadPattern checks that Alloc surfaces fill errors under
+// the buffer's name, so hand-built specs fail as loudly as parsed ones.
+func TestBufPoolAllocBadPattern(t *testing.T) {
+	var pool BufPool
+	defer pool.Close()
+	if err := pool.Alloc([]BufSpec{{Name: "dst", Size: 16, Pattern: "nope"}}); err == nil {
+		t.Fatal("Alloc with an unknown pattern must fail")
+	}
 }
 
 func TestFuzzFuncChecked(t *testing.T) {

verify/fuzz_test.go

@@ -4,6 +4,8 @@
 package verify
 
 import (
+	"os"
+	"strings"
 	"testing"
 )
 
@@ -65,6 +67,72 @@ func TestFuzzWideCopy(t *testing.T) {
 	}
 }
 
+// TestFuzzFuncSigWiderThanFrame pins the guard against a // func comment
+// that declares more parameter bytes than the TEXT frame carries: before
+// the guard, slicing the result area at paramsSize(sig) past the end of
+// the argument block panicked the whole test binary.
+func TestFuzzFuncSigWiderThanFrame(t *testing.T) {
+	k := loadBasic(t)
+
+	gt, err := GroundTruth("../testdata/verify/basic_amd64.s")
+	if err != nil {
+		t.Skipf("go tool asm unavailable: %v", err)
+	}
+	goCode, ok := gt["add"]
+	if !ok {
+		t.Skip("add not in ground truth")
+	}
+
+	// add carries $0-24; four int parameters declare 32 bytes.
+	sig := funcSig{
+		name: "add",
+		params: []param{
+			{name: "a", typ: "int"},
+			{name: "b", typ: "int"},
+			{name: "c", typ: "int"},
+			{name: "d", typ: "int"},
+		},
+	}
+
+	res := k.FuzzFunc("add", sig, goCode, 3, 42)
+	if res.OK() {
+		t.Fatal("expected the over-wide signature to fail the campaign")
+	}
+	if !strings.Contains(res.FirstFail, "parameter bytes") || !strings.Contains(res.FirstFail, "argument bytes") {
+		t.Errorf("FirstFail = %q, want a clear signature-versus-frame message", res.FirstFail)
+	}
+}
+
+// TestFuzzStringParam runs the differential fuzzer over a kernel whose
+// only parameter is a string: the marshaller lays out a real two-word
+// header (data pointer + length) and the comparison slices at
+// paramsSize(sig) = 16, so the length word is input, not result.
+func TestFuzzStringParam(t *testing.T) {
+	k := loadStrProbeKernel(t)
+
+	file := t.TempDir() + "/strprobe_amd64.s"
+	if err := os.WriteFile(file, []byte(strProbeSrc), 0o644); err != nil {
+		t.Fatalf("write kernel: %v", err)
+	}
+	gt, err := GroundTruth(file)
+	if err != nil {
+		t.Skipf("go tool asm unavailable: %v", err)
+	}
+	goCode, ok := gt["strProbe"]
+	if !ok {
+		t.Skip("strProbe not in ground truth")
+	}
+
+	sig, ok := parseFuncSig("// func strProbe(s string) int64")
+	if !ok {
+		t.Fatal("parseFuncSig failed")
+	}
+	res := k.FuzzFunc("strProbe", sig, goCode, 100, 42)
+	if !res.OK() {
+		t.Errorf("strProbe fuzz: %s", res)
+	}
+}
+
 func TestParseFuncSig(t *testing.T) {
 	tests := []struct {
 		comment string

verify/groundtruth.go

@@ -20,8 +20,8 @@ import (
 // This is the universal oracle: any file that `go tool asm` accepts can
 // be verified, with no hand-written reference.
 //
-// For RISC-V sources the assembler is invoked with GOARCH=riscv64;
+// For the other architectures use the cross-assembly entry points
-// the caller must set the architecture via GroundTruthArch.
+// GroundTruthARM64, GroundTruthRISCV and GroundTruthLOONG64.
 func GroundTruth(path string) (map[string][]byte, error) {
 	return groundTruthArch(path, "")
 }
@@ -82,7 +82,19 @@ func groundTruthArch(path, goarch string) (map[string][]byte, error) {
 
 	cmd := exec.Command(asmBin, "-I", includeDir, "-p", pkg, "-o", objPath, path)
 	if goarch != "" {
-		cmd.Env = append(os.Environ(), "GOARCH="+goarch)
+		// Replace, do not duplicate, any GOARCH the ambient environment
+		// exports.  The Go runtime resolves duplicated keys last-wins,
+		// so the appended entry happens to win today; a single
+		// unambiguous entry keeps the child's target architecture from
+		// depending on that resolution order.
+		environ := os.Environ()
+		env := make([]string, 0, len(environ)+1)
+		for _, e := range environ {
+			if !strings.HasPrefix(e, "GOARCH=") {
+				env = append(env, e)
+			}
+		}
+		cmd.Env = append(env, "GOARCH="+goarch)
 	}
 	if out, err := cmd.CombinedOutput(); err != nil {
 		return nil, fmt.Errorf("verify: go tool asm (%s): %w\n%s", goarch, err, out)

verify/groundtruth_test.go

@@ -82,6 +82,22 @@ func TestGroundTruthBadFile(t *testing.T) {
 	}
 }
 
+// TestGroundTruthCrossEnvGOARCH checks that a GOARCH exported in the
+// ambient environment does not override cross-assembly: the child
+// assembler must honour the requested GOARCH even when the environment
+// already carries a GOARCH entry (os.Getenv reads the first entry, so a
+// duplicate would win).
+func TestGroundTruthCrossEnvGOARCH(t *testing.T) {
+	t.Setenv("GOARCH", "amd64")
+	gt, err := GroundTruthARM64("../testdata/verify/basic_arm64.s")
+	if err != nil {
+		t.Fatalf("GroundTruthARM64 with ambient GOARCH=amd64: %v", err)
+	}
+	if len(gt) == 0 {
+		t.Fatal("no functions extracted from ground truth")
+	}
+}
+
 func keys(m map[string][]byte) []string {
 	out := make([]string, 0, len(m))
 	for k := range m {

verify/jit_test.go

@@ -12,6 +12,7 @@ import (
 
 func loadBasic(t *testing.T) *Kernel {
 	t.Helper()
+	requireHost(t, "amd64")
 	k, err := Load("../testdata/verify/basic_amd64.s")
 	if err != nil {
 		t.Fatalf("Load: %v", err)

verify/sigparse.go

@@ -153,6 +153,13 @@ func paramSize(typ string) int {
 		return 2
 	case typ == "int8", typ == "uint8", typ == "byte", typ == "bool":
 		return 1
+	case typ == "string":
+		// ABI0 strings are a two-word header: data pointer + length.
+		return 16
+	case typ == "complex64":
+		return 8
+	case typ == "complex128":
+		return 16
 	default:
 		// Default to 8 bytes for unknown types.
 		return 8

verify/sigparse_test.go

@@ -143,7 +143,10 @@ func TestParamSize(t *testing.T) {
 		{"uint8", 1},
 		{"byte", 1},
 		{"bool", 1},
-		{"string", 8}, // unknown type defaults to 8
+		{"string", 16}, // ABI0 header: data pointer + length
+		{"complex64", 8},
+		{"complex128", 16},
+		{"error", 8}, // unknown type defaults to 8
 	}
 	for _, tt := range tests {
 		if got := paramSize(tt.typ); got != tt.want {

verify/trampoline_arm64.s

@@ -17,7 +17,8 @@
 TEXT ·enterJIT(SB), NOSPLIT, $0-16
 	MOVD fn+0(FP), R0 // target function address
 	MOVD R30, savedLR(SB) // save link register
-	MOVD R3, savedSP(SB) // save Go stack pointer
+	MOVD RSP, R3 // save Go stack pointer: R3 relays the value, the
+	MOVD R3, savedSP(SB) // arm64 assembler cannot store RSP to memory
 	MOVD stack+8(FP), R3 // load prepared stack pointer
 	MOVD 0(R3), R30 // load leaveJIT address into LR
 	MOVD R3, RSP // switch to the prepared stack: SP stays on

verify/verify.go

@@ -5,7 +5,6 @@ package verify
 
 import (
 	"fmt"
-	"math/rand"
 	"os"
 	"runtime"
 
@@ -101,6 +100,9 @@ func (k *Kernel) FuncNames() []string {
 // The arg block is the raw bytes of the function's argument/result area
 // (as declared by the TEXT $frame-args suffix).  Returns the arg block
 // after the call (with any results written back by the function).
+//
+// Not safe for concurrent use: only one JIT call may be in flight at a
+// time, the trampolines keep the saved registers in package globals.
 func (k *Kernel) CallFunc(name string, args []byte) ([]byte, error) {
 	idx, ok := k.funcs[name]
 	if !ok {
@@ -116,7 +118,12 @@ func (k *Kernel) CallFunc(name string, args []byte) ([]byte, error) {
 
 // CallFuncChecked invokes the named function with ABI sentinels and a
 // red-zone canary, returning the argument block and an ABIReport that
-// records any callee-saved register or red-zone violations.
+// records any callee-saved register or red-zone violations.  The canary
+// gap is sized from the function's declared frame, so legal frame writes
+// do not count as violations.
+//
+// Not safe for concurrent use: only one JIT call may be in flight at a
+// time, the trampolines keep the saved registers in package globals.
 func (k *Kernel) CallFuncChecked(name string, args []byte) ([]byte, ABIReport, error) {
 	idx, ok := k.funcs[name]
 	if !ok {
@@ -127,11 +134,14 @@ func (k *Kernel) CallFuncChecked(name string, args []byte) ([]byte, ABIReport, e
 		return nil, ABIReport{}, fmt.Errorf("verify: %s: arg block too small: got %d, need %d", name, len(args), fl.Args)
 	}
 	fnAddr := k.exec.FuncAddr(fl.Offset)
-	return CallChecked(fnAddr, args)
+	return CallCheckedFrame(fnAddr, args, fl.Frame)
 }
 
 // FuzzFuncCheckedByName is like FuzzFuncChecked but extracts the signature
 // from the source code internally.
+//
+// Not safe for concurrent use: only one JIT call may be in flight at a
+// time, the trampolines keep the saved registers in package globals.
 func (k *Kernel) FuzzFuncCheckedByName(name, src string, iterations int, seed int64) FuzzResult {
 	result := FuzzResult{Func: name, Iterations: iterations}
 	sig, ok := ExtractSignatures(src)[name]
@@ -146,9 +156,12 @@ func (k *Kernel) FuzzFuncCheckedByName(name, src string, iterations int, seed in
 // FuzzFuncChecked combines fuzzing with ABI checks: it generates varied
 // inputs and verifies that callee-saved registers and the red zone are
 // preserved even on deep execution paths (not just early exits).
+//
+// Not safe for concurrent use: only one JIT call may be in flight at a
+// time, the trampolines keep the saved registers in package globals.
 func (k *Kernel) FuzzFuncChecked(name string, sig funcSig, iterations int, seed int64) FuzzResult {
 	result := FuzzResult{Func: name, Iterations: iterations}
-	rng := rand.New(rand.NewSource(seed))
+	rng := newRNG(seed)
 
 	fl, err := k.Func(name)
 	if err != nil {