feat(verify): add the JIT execution substrate and gasm verify subcommand

Assisted-by: Qwen 3.8 Max Preview

cmd/gasm/main.go

@@ -24,11 +24,12 @@ import (
 	"sourcedock.dev/petrbalvin/gasm-devkit/lint"
 	"sourcedock.dev/petrbalvin/gasm-devkit/lsp"
 	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+	"sourcedock.dev/petrbalvin/gasm-devkit/verify"
 )
 
 // version is the release version, stamped at build time via
 // -ldflags "-X main.version=…" (defaulting to the current release).
-var version = "0.16.0"
+var version = "0.17.0"
 
 func main() {
 	if len(os.Args) < 2 {
@@ -46,6 +47,8 @@ func main() {
 		os.Exit(cmdLint(os.Args[2:]))
 	case "asm":
 		os.Exit(cmdAsm(os.Args[2:]))
+	case "verify":
+		os.Exit(cmdVerify(os.Args[2:]))
 	case "lsp":
 		os.Exit(cmdLSP(os.Args[2:]))
 	case "version", "--version", "-V":
@@ -80,6 +83,7 @@ Commands:
   fmt       canonicalise formatting (gofmt for assembly)
   lint      run static checks
   asm       assemble .s files to machine code (amd64)
+  verify    JIT-assemble and run dynamic checks (amd64)
   lsp       run the language server over stdio
   version   print the version (same as --version)
 
@@ -466,3 +470,57 @@ requires -p, the package path, and the installed Go toolchain).
 	}
 	return 0
 }
+
+func cmdVerify(args []string) int {
+	fs := newCommand("verify", "gasm verify <file.s>", `
+Assemble FILE (amd64), map it into executable memory and report the available
+functions.  This confirms the assembled image is self-consistent (no
+unresolved external symbols) and executable — the prerequisite for dynamic
+testing.
+
+With -smoke, each NOSPLIT function is called with a zeroed argument block to
+confirm the JIT trampoline works end-to-end.  This is safe only for functions
+that tolerate nil pointers and zero lengths in their arguments.
+`)
+	smoke := fs.Bool("smoke", false, "call each NOSPLIT function with zeroed args")
+	fs.Parse(args)
+	if fs.NArg() != 1 {
+		fmt.Fprintln(os.Stderr, "usage: gasm verify [-smoke] <file.s>")
+		return 2
+	}
+	path := fs.Arg(0)
+	if arch.FromFilename(path) != arch.AMD64 {
+		fmt.Fprintln(os.Stderr, "gasm verify: only amd64 is supported")
+		return 1
+	}
+
+	k, err := verify.Load(path)
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "gasm verify: %v\n", err)
+		return 1
+	}
+	defer k.Close()
+
+	names := k.FuncNames()
+	fmt.Printf("%s: %d functions JIT-loaded\n", path, len(names))
+	rc := 0
+	for _, name := range names {
+		fl, _ := k.Func(name)
+		flags := ""
+		if fl.NoSplit {
+			flags = " NOSPLIT"
+		}
+		fmt.Printf("  %s: %d bytes, args=%d, frame=%d%s\n", name, fl.Size, fl.Args, fl.Frame, flags)
+		if *smoke && fl.NoSplit {
+			args := make([]byte, fl.Args)
+			_, err := k.CallFunc(name, args)
+			if err != nil {
+				fmt.Printf("    smoke: FAIL — %v\n", err)
+				rc = 1
+			} else {
+				fmt.Printf("    smoke: OK\n")
+			}
+		}
+	}
+	return rc
+}

docs/ARCHITECTURE.md

@@ -284,6 +284,31 @@ references and the implicit funcdata/DWARF symbols remain future work (the
 linker fills the latter's defaults); the rest of Phase 2 is those, the
 remaining EVEX forms and the other architectures.
 
+### `verify`
+
+The dynamic-analysis substrate (Phase 3).  It JIT-loads assembled images into
+executable memory and invokes them directly, enabling differential testing,
+runtime ABI checks and coverage profiling.
+
+The execution model is pure Go (stdlib only).  `Map` copies machine code into
+an anonymous `syscall.Mmap` mapping and enforces W^X (write the bytes, then
+`mprotect` to read-execute).  `Call` prepares a stack whose first word is the
+address of an assembly trampoline (`leaveJIT`), lays the ABI0 argument
+block after it, switches to that stack via `enterJIT` (which saves the Go
+stack pointer in a package global and jumps to the target), and recovers
+control when the function RETs into `leaveJIT` (which restores the Go stack
+and returns).  A 64-byte pad below the return address accommodates the
+ABIInternal wrapper that the Go runtime interposes on assembly functions.
+
+`Load` / `LoadSource` / `LoadAST` parse, assemble and map a `.s` file in one
+step, returning a `Kernel` whose `CallFunc` method marshals the argument block
+by name.  The image must be self-contained (no external relocations); the
+assembler’s `Image.Bytes()` provides the code-and-data concatenation.
+
+The `gasm verify` CLI subcommand exposes this: it loads a file, reports the
+available functions and (with `-smoke`) calls each NOSPLIT function with zeroed
+arguments to confirm the trampoline round-trips.
+
 ## Extension points
 
 - **New architecture:** add an entry to the generator in `_gen`, run

justfile

@@ -3,7 +3,7 @@
 
 # gasm-devkit — developer tooling for Go's Plan 9 assembler (GAsm).
 
-version := "0.16.0"
+version := "0.17.0"
 
 default:
     @just --list

testdata/verify/basic_amd64.s

@@ -0,0 +1,67 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+#include "textflag.h"
+
+// func add(a, b int64) int64
+TEXT ·add(SB), NOSPLIT, $0-24
+	MOVQ a+0(FP), AX
+	ADDQ b+8(FP), AX
+	MOVQ AX, ret+16(FP)
+	RET
+
+// func sum(data []int64) int64
+// Sums all elements of the slice.
+TEXT ·sum(SB), NOSPLIT, $0-32
+	MOVQ data_base+0(FP), SI
+	MOVQ data_len+8(FP), CX
+	XORQ AX, AX
+	TESTQ CX, CX
+	JZ   sum_done
+
+sum_loop:
+	ADDQ (SI), AX
+	ADDQ $8, SI
+	DECQ CX
+	JNZ  sum_loop
+
+sum_done:
+	MOVQ AX, ret+24(FP)
+	RET
+
+// func wideCopy(dst, src []byte)
+// Non-overlapping copy of min(len(dst), len(src)) bytes using 32-byte moves.
+TEXT ·wideCopy(SB), NOSPLIT, $0-48
+	MOVQ dst_base+0(FP), DI
+	MOVQ dst_len+8(FP), BX
+	MOVQ src_base+24(FP), SI
+	MOVQ src_len+32(FP), R8
+	CMPQ BX, R8
+	JLE  wc_have_n
+	MOVQ R8, BX
+
+wc_have_n:
+	CMPQ BX, $32
+	JB   wc_small
+
+	VMOVDQU (SI), Y0
+	VMOVDQU Y0, (DI)
+	VMOVDQU -32(SI)(BX*1), Y0
+	VMOVDQU Y0, -32(DI)(BX*1)
+	VZEROUPPER
+	RET
+
+wc_small:
+	TESTQ BX, BX
+	JZ    wc_done
+
+wc_byte:
+	MOVB (SI), R8B
+	MOVB R8B, (DI)
+	INCQ SI
+	INCQ DI
+	DECQ BX
+	JNZ  wc_byte
+
+wc_done:
+	RET

verify/call_amd64.go

@@ -0,0 +1,77 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+//go:build amd64
+
+package verify
+
+import (
+	"encoding/binary"
+	"fmt"
+	"reflect"
+	"syscall"
+	"unsafe"
+)
+
+// savedSP holds the Go stack pointer while a JIT call is in flight.
+// Referenced by the assembly trampoline (trampoline_amd64.s).
+var savedSP uintptr
+
+// enterJIT switches to the prepared stack and jumps to fn.
+// It does not return normally; the JIT function's RET transfers control
+// to leaveJIT, which restores the Go stack.
+//
+//go:nosplit
+func enterJIT(fn uintptr, stack uintptr)
+
+// leaveJIT restores the Go stack after a JIT function returns.
+// Its address is placed as the return address on the prepared stack.
+//
+//go:nosplit
+func leaveJIT()
+
+// leaveJITAddr is the machine address of leaveJIT, resolved once at init.
+var leaveJITAddr uintptr
+
+func init() {
+	leaveJITAddr = reflect.ValueOf(leaveJIT).Pointer()
+}
+
+// stackPad is padding below the return address on the prepared stack.
+// The ABIInternal wrapper that leaveJIT's address resolves to executes
+// PUSHQ BP and CALL before reaching the raw assembly, writing up to 16
+// bytes below the return-address slot.  64 bytes of headroom is ample.
+const stackPad = 64
+
+// Call invokes the assembled function at fnAddr with the given ABI0 argument
+// block (the raw bytes that would appear at FP+0).  It returns the argument
+// block after the call, which contains any results the function wrote back
+// (the ABI0 convention shares the argument area for inputs and outputs).
+//
+// The function must be NOSPLIT (no stack growth) and must not reference
+// external symbols — the image is self-contained.
+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
+	stackMem, err := syscall.Mmap(-1, 0, stackSize,
+		syscall.PROT_READ|syscall.PROT_WRITE, syscall.MAP_PRIVATE|syscall.MAP_ANON)
+	if err != nil {
+		return nil, fmt.Errorf("verify: stack mmap: %w", err)
+	}
+	defer syscall.Munmap(stackMem)
+
+	// The return address sits after the padding; the function's SP will
+	// point here, leaving stackPad bytes below for the wrapper's pushes.
+	retOff := stackPad
+	binary.LittleEndian.PutUint64(stackMem[retOff:retOff+8], uint64(leaveJITAddr))
+	// The ABI0 argument area follows the return address.
+	copy(stackMem[retOff+8:], args)
+
+	stackBase := uintptr(unsafe.Pointer(&stackMem[retOff]))
+	enterJIT(fnAddr, stackBase)
+
+	// Copy out the (possibly modified) argument area.
+	out := make([]byte, len(args))
+	copy(out, stackMem[retOff+8:retOff+8+len(args)])
+	return out, nil
+}

verify/call_other.go

@@ -0,0 +1,13 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+//go:build !amd64
+
+package verify
+
+import "fmt"
+
+// Call is unavailable on non-amd64 architectures.
+func Call(fnAddr uintptr, args []byte) ([]byte, error) {
+	return nil, fmt.Errorf("verify: JIT execution requires amd64")
+}

verify/jit.go

@@ -0,0 +1,88 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+// Package verify provides the dynamic-analysis substrate for gasm: it
+// JIT-assembles Plan 9 amd64 kernels into executable memory and calls them
+// directly, enabling differential testing against portable Go references,
+// runtime ABI checks and basic-block coverage profiling.
+//
+// The execution model is pure Go (stdlib only): machine code is mapped with
+// syscall.Mmap and invoked through an assembly trampoline that switches to a
+// prepared ABI0 stack.  No cgo, no external toolchain.
+package verify
+
+import (
+	"encoding/binary"
+	"fmt"
+	"syscall"
+	"unsafe"
+)
+
+// Executable maps a copy of code into a read-execute memory region suitable
+// for direct invocation.  The mapping is anonymous and private; the original
+// slice is not retained.  Call Unmap to release the region.
+type Executable struct {
+	addr uintptr // base address of the mapping
+	size int
+	mem  []byte // the mmap'd slice (for Unmap)
+}
+
+// Map copies code into a freshly allocated RX region and returns it.
+// The mapping is PROT_READ|PROT_EXEC; writes are not permitted after the
+// copy, matching W^X policy.
+func Map(code []byte) (*Executable, error) {
+	size := len(code)
+	if size == 0 {
+		return nil, fmt.Errorf("verify: cannot map zero-length code")
+	}
+	// Round up to the page size.
+	const pageSize = 4096
+	mapSize := (size + pageSize - 1) &^ (pageSize - 1)
+
+	mem, err := syscall.Mmap(-1, 0, mapSize,
+		syscall.PROT_READ|syscall.PROT_WRITE, syscall.MAP_PRIVATE|syscall.MAP_ANON)
+	if err != nil {
+		return nil, fmt.Errorf("verify: mmap: %w", err)
+	}
+	copy(mem, code)
+
+	// Remove write permission (W^X).
+	if err := syscall.Mprotect(mem, syscall.PROT_READ|syscall.PROT_EXEC); err != nil {
+		syscall.Munmap(mem)
+		return nil, fmt.Errorf("verify: mprotect: %w", err)
+	}
+	return &Executable{
+		addr: uintptr(unsafe.Pointer(&mem[0])),
+		size: size,
+		mem:  mem,
+	}, nil
+}
+
+// Unmap releases the executable region.
+func (e *Executable) Unmap() {
+	if e.mem != nil {
+		syscall.Munmap(e.mem)
+		e.mem = nil
+	}
+}
+
+// FuncAddr returns the absolute address of a function at the given offset
+// within the mapped image.
+func (e *Executable) FuncAddr(offset int) uintptr {
+	return e.addr + uintptr(offset)
+}
+
+// PutUint64 writes v into buf at byte offset off (little-endian).
+func PutUint64(buf []byte, off int, v uint64) {
+	binary.LittleEndian.PutUint64(buf[off:off+8], v)
+}
+
+// GetUint64 reads a little-endian uint64 from buf at byte offset off.
+func GetUint64(buf []byte, off int) uint64 {
+	return binary.LittleEndian.Uint64(buf[off : off+8])
+}
+
+// PutPtr writes a pointer value into buf at byte offset off.
+func PutPtr(buf []byte, off int, p unsafe.Pointer) {
+	binary.LittleEndian.PutUint64(buf[off:off+8], uint64(uintptr(p)))
+}

verify/jit_test.go

@@ -0,0 +1,159 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package verify
+
+import (
+	"bytes"
+	"testing"
+	"unsafe"
+)
+
+func loadBasic(t *testing.T) *Kernel {
+	t.Helper()
+	k, err := Load("../testdata/verify/basic_amd64.s")
+	if err != nil {
+		t.Fatalf("Load: %v", err)
+	}
+	t.Cleanup(k.Close)
+	return k
+}
+
+func TestJITAdd(t *testing.T) {
+	k := loadBasic(t)
+
+	tests := []struct {
+		a, b, want int64
+	}{
+		{0, 0, 0},
+		{1, 2, 3},
+		{-1, 1, 0},
+		{1 << 62, 1 << 62, -9223372036854775808}, // overflow wraps (MinInt64)
+		{-100, -200, -300},
+	}
+	for _, tt := range tests {
+		args := make([]byte, 24)
+		PutUint64(args, 0, uint64(tt.a))
+		PutUint64(args, 8, uint64(tt.b))
+
+		out, err := k.CallFunc("add", args)
+		if err != nil {
+			t.Fatalf("CallFunc(add, %d, %d): %v", tt.a, tt.b, err)
+		}
+		got := int64(GetUint64(out, 16))
+		if got != tt.want {
+			t.Errorf("add(%d, %d) = %d, want %d", tt.a, tt.b, got, tt.want)
+		}
+	}
+}
+
+func TestJITSum(t *testing.T) {
+	k := loadBasic(t)
+
+	tests := []struct {
+		data []int64
+		want int64
+	}{
+		{nil, 0},
+		{[]int64{1}, 1},
+		{[]int64{1, 2, 3, 4, 5}, 15},
+		{[]int64{-10, 20, -30, 40}, 20},
+	}
+	for _, tt := range tests {
+		args := make([]byte, 32)
+		if len(tt.data) > 0 {
+			PutPtr(args, 0, unsafe.Pointer(&tt.data[0]))
+		}
+		PutUint64(args, 8, uint64(len(tt.data)))
+		PutUint64(args, 16, uint64(cap(tt.data)))
+
+		out, err := k.CallFunc("sum", args)
+		if err != nil {
+			t.Fatalf("CallFunc(sum, %v): %v", tt.data, err)
+		}
+		got := int64(GetUint64(out, 24))
+		if got != tt.want {
+			t.Errorf("sum(%v) = %d, want %d", tt.data, got, tt.want)
+		}
+	}
+}
+
+func TestJITWideCopy(t *testing.T) {
+	k := loadBasic(t)
+
+	tests := []struct {
+		name string
+		n    int
+	}{
+		{"empty", 0},
+		{"tiny", 7},
+		{"exact32", 32},
+		{"overlap_range", 48},
+		{"exact64", 64},
+		{"unaligned", 45},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			src := make([]byte, tt.n)
+			for i := range src {
+				src[i] = byte(i * 7)
+			}
+			dst := make([]byte, tt.n)
+
+			args := make([]byte, 48)
+			if tt.n > 0 {
+				PutPtr(args, 0, unsafe.Pointer(&dst[0]))
+				PutPtr(args, 24, unsafe.Pointer(&src[0]))
+			}
+			PutUint64(args, 8, uint64(tt.n))  // dst_len
+			PutUint64(args, 16, uint64(tt.n)) // dst_cap
+			PutUint64(args, 32, uint64(tt.n)) // src_len
+			PutUint64(args, 40, uint64(tt.n)) // src_cap
+
+			_, err := k.CallFunc("wideCopy", args)
+			if err != nil {
+				t.Fatalf("CallFunc(wideCopy): %v", err)
+			}
+			if !bytes.Equal(dst, src) {
+				t.Errorf("wideCopy: dst ≠ src\n  got  %x\n  want %x", dst, src)
+			}
+		})
+	}
+}
+
+func TestKernelFuncNames(t *testing.T) {
+	k := loadBasic(t)
+	names := k.FuncNames()
+	want := []string{"add", "sum", "wideCopy"}
+	if len(names) != len(want) {
+		t.Fatalf("FuncNames() = %v, want %v", names, want)
+	}
+	for i, n := range names {
+		if n != want[i] {
+			t.Errorf("FuncNames()[%d] = %q, want %q", i, n, want[i])
+		}
+	}
+}
+
+func TestKernelFuncNotFound(t *testing.T) {
+	k := loadBasic(t)
+	_, err := k.CallFunc("nonexistent", make([]byte, 8))
+	if err == nil {
+		t.Fatal("expected error for nonexistent function")
+	}
+}
+
+func TestKernelArgTooSmall(t *testing.T) {
+	k := loadBasic(t)
+	_, err := k.CallFunc("add", make([]byte, 8)) // needs 24
+	if err == nil {
+		t.Fatal("expected error for too-small arg block")
+	}
+}
+
+func TestMapZeroLength(t *testing.T) {
+	_, err := Map(nil)
+	if err == nil {
+		t.Fatal("expected error for zero-length code")
+	}
+}

verify/lz4_test.go

@@ -0,0 +1,160 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package verify
+
+import (
+	"bytes"
+	"os"
+	"testing"
+	"unsafe"
+)
+
+// lz4KernelPath is the sibling repository's AVX2 kernel, used for
+// integration testing.  The test is skipped when the file is absent
+// (e.g. in CI without the sibling checkout).
+const lz4KernelPath = "../../go-libraries/go-lz4/avx2_amd64.s"
+
+func loadLZ4Kernel(t *testing.T) *Kernel {
+	t.Helper()
+	if _, err := os.Stat(lz4KernelPath); err != nil {
+		t.Skipf("sibling kernel not available: %v", err)
+	}
+	k, err := Load(lz4KernelPath)
+	if err != nil {
+		t.Fatalf("Load(%s): %v", lz4KernelPath, err)
+	}
+	t.Cleanup(k.Close)
+	return k
+}
+
+// callDecodeBlockAVX2 invokes the JIT-assembled decodeBlockAVX2 with the
+// given src and dst buffers, returning (n, code).
+func callDecodeBlockAVX2(t *testing.T, k *Kernel, src, dst []byte) (int, int) {
+	t.Helper()
+	args := make([]byte, 64)
+	if len(src) > 0 {
+		PutPtr(args, 0, unsafe.Pointer(&src[0]))
+	}
+	PutUint64(args, 8, uint64(len(src)))
+	PutUint64(args, 16, uint64(cap(src)))
+	if len(dst) > 0 {
+		PutPtr(args, 24, unsafe.Pointer(&dst[0]))
+	}
+	PutUint64(args, 32, uint64(len(dst)))
+	PutUint64(args, 40, uint64(cap(dst)))
+
+	out, err := k.CallFunc("decodeBlockAVX2", args)
+	if err != nil {
+		t.Fatalf("CallFunc(decodeBlockAVX2): %v", err)
+	}
+	return int(GetUint64(out, 48)), int(GetUint64(out, 56))
+}
+
+func TestLZ4DecodeKnownAnswers(t *testing.T) {
+	k := loadLZ4Kernel(t)
+
+	tests := []struct {
+		name     string
+		src      []byte
+		dstSize  int
+		wantDst  []byte
+		wantN    int
+		wantCode int
+	}{
+		{
+			name:     "literals_only",
+			src:      []byte{0x50, 'H', 'e', 'l', 'l', 'o'},
+			dstSize:  16,
+			wantDst:  []byte("Hello"),
+			wantN:    5,
+			wantCode: 0,
+		},
+		{
+			name:     "literals_and_match",
+			src:      []byte{0x54, 'A', 'A', 'A', 'A', 'A', 0x05, 0x00, 0x30, 'B', 'B', 'B'},
+			dstSize:  32,
+			wantDst:  []byte("AAAAAAAAAAAAABBB"),
+			wantN:    16,
+			wantCode: 0,
+		},
+		{
+			name: "overlapping_match",
+			// 1 literal 'X', then match offset=1 length=4+4=8 → "XXXXXXXXX",
+			// then final 1 literal 'Y'.
+			src:      []byte{0x14, 'X', 0x01, 0x00, 0x10, 'Y'},
+			dstSize:  16,
+			wantDst:  []byte("XXXXXXXXXY"),
+			wantN:    10,
+			wantCode: 0,
+		},
+		{
+			name:     "malformed_truncated",
+			src:      []byte{0x50, 'H', 'e'}, // claims 5 literals, has 2
+			dstSize:  16,
+			wantN:    0,
+			wantCode: 1,
+		},
+		{
+			name:     "zero_offset",
+			src:      []byte{0x14, 'X', 0x00, 0x00},
+			dstSize:  16,
+			wantN:    0,
+			wantCode: 2,
+		},
+		{
+			name:     "empty_token",
+			src:      []byte{0x00}, // 0 literals, end of block
+			dstSize:  16,
+			wantDst:  nil,
+			wantN:    0,
+			wantCode: 0,
+		},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			dst := make([]byte, tt.dstSize)
+			n, code := callDecodeBlockAVX2(t, k, tt.src, dst)
+			if n != tt.wantN || code != tt.wantCode {
+				t.Fatalf("decodeBlockAVX2: got (n=%d, code=%d), want (n=%d, code=%d)",
+					n, code, tt.wantN, tt.wantCode)
+			}
+			if tt.wantCode == 0 && tt.wantDst != nil {
+				if !bytes.Equal(dst[:n], tt.wantDst) {
+					t.Errorf("output mismatch:\n  got  %q\n  want %q", dst[:n], tt.wantDst)
+				}
+			}
+		})
+	}
+}
+
+func TestLZ4WideCopyAVX2(t *testing.T) {
+	k := loadLZ4Kernel(t)
+
+	sizes := []int{0, 1, 15, 16, 31, 32, 33, 63, 64, 100, 256, 1024}
+	for _, n := range sizes {
+		src := make([]byte, n)
+		for i := range src {
+			src[i] = byte(i*13 + 7)
+		}
+		dst := make([]byte, n)
+
+		args := make([]byte, 48)
+		if n > 0 {
+			PutPtr(args, 0, unsafe.Pointer(&dst[0]))
+			PutPtr(args, 24, unsafe.Pointer(&src[0]))
+		}
+		PutUint64(args, 8, uint64(n))
+		PutUint64(args, 16, uint64(n))
+		PutUint64(args, 32, uint64(n))
+		PutUint64(args, 40, uint64(n))
+
+		_, err := k.CallFunc("wideCopyAVX2", args)
+		if err != nil {
+			t.Fatalf("wideCopyAVX2(n=%d): %v", n, err)
+		}
+		if !bytes.Equal(dst, src) {
+			t.Errorf("wideCopyAVX2(n=%d): output mismatch", n)
+		}
+	}
+}

verify/trampoline_amd64.s

@@ -0,0 +1,30 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+#include "textflag.h"
+
+// ABI0 JIT trampoline.  enterJIT switches from the Go stack to a prepared
+// stack and jumps to the assembled function; when the function RETs, control
+// lands in leaveJIT, which restores the Go stack and returns to the Go caller.
+//
+// The prepared stack must begin with the address of leaveJIT (the return
+// address the JIT function will pop), followed by the function's ABI0
+// argument area.
+//
+// Single-threaded: savedSP is a package global, so only one JIT call may be
+// in flight at a time.  gasm verify runs sequentially.
+
+// func enterJIT(fn uintptr, stack uintptr)
+// Switches to the prepared stack and jumps to fn.  Does not return normally;
+// the JIT function's RET transfers control to leaveJIT.
+TEXT ·enterJIT(SB), NOSPLIT, $0-16
+	MOVQ fn+0(FP), AX      // target function address (before SP switch)
+	MOVQ SP, ·savedSP(SB)  // preserve the Go stack pointer
+	MOVQ stack+8(FP), SP   // switch to the prepared stack
+	JMP  AX
+
+// func leaveJIT()
+// Restores the Go stack pointer and returns to enterJIT's caller.
+TEXT ·leaveJIT(SB), NOSPLIT, $0-0
+	MOVQ ·savedSP(SB), SP
+	RET

verify/verify.go

@@ -0,0 +1,106 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+package verify
+
+import (
+	"fmt"
+	"os"
+
+	"sourcedock.dev/petrbalvin/gasm-devkit/asm"
+	"sourcedock.dev/petrbalvin/gasm-devkit/ast"
+	"sourcedock.dev/petrbalvin/gasm-devkit/parser"
+)
+
+// Kernel is a JIT-loaded assembly image ready for direct invocation.
+// It wraps an executable memory mapping and the function layout metadata
+// needed to marshal ABI0 calls.
+type Kernel struct {
+	exec  *Executable
+	img   *asm.Image
+	funcs map[string]int // function name → index into img.Funcs
+}
+
+// Load parses, assembles and maps a .s file into executable memory.
+// The returned Kernel is ready for Call.  The caller must call Close to
+// release the mapping.
+func Load(path string) (*Kernel, error) {
+	src, err := os.ReadFile(path)
+	if err != nil {
+		return nil, fmt.Errorf("verify: %w", err)
+	}
+	return LoadSource(path, string(src))
+}
+
+// LoadSource parses, assembles and maps assembly source into executable memory.
+func LoadSource(filename, src string) (*Kernel, error) {
+	file, errs := parser.Parse(filename, src)
+	if len(errs) > 0 {
+		return nil, fmt.Errorf("verify: parse %s: %v", filename, errs[0])
+	}
+	return LoadAST(file)
+}
+
+// LoadAST assembles a parsed AST file and maps the result into executable
+// memory.
+func LoadAST(file *ast.File) (*Kernel, error) {
+	img, err := asm.AssembleFile(file)
+	if err != nil {
+		return nil, fmt.Errorf("verify: assemble: %w", err)
+	}
+	if len(img.Externals) > 0 {
+		return nil, fmt.Errorf("verify: unresolved external symbols: %v", img.Externals)
+	}
+	code := img.Bytes()
+	exec, err := Map(code)
+	if err != nil {
+		return nil, err
+	}
+	funcs := make(map[string]int, len(img.Funcs))
+	for i, f := range img.Funcs {
+		funcs[f.Name] = i
+	}
+	return &Kernel{exec: exec, img: img, funcs: funcs}, nil
+}
+
+// Func returns the layout metadata for the named function.
+func (k *Kernel) Func(name string) (asm.FuncLayout, error) {
+	idx, ok := k.funcs[name]
+	if !ok {
+		return asm.FuncLayout{}, fmt.Errorf("verify: function %q not found", name)
+	}
+	return k.img.Funcs[idx], nil
+}
+
+// FuncNames returns the names of all functions in the kernel, in source order.
+func (k *Kernel) FuncNames() []string {
+	names := make([]string, len(k.img.Funcs))
+	for i, f := range k.img.Funcs {
+		names[i] = f.Name
+	}
+	return names
+}
+
+// CallFunc invokes the named function with the given ABI0 argument block.
+// 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).
+func (k *Kernel) CallFunc(name string, args []byte) ([]byte, error) {
+	idx, ok := k.funcs[name]
+	if !ok {
+		return nil, fmt.Errorf("verify: function %q not found", name)
+	}
+	fl := k.img.Funcs[idx]
+	if len(args) < fl.Args {
+		return nil, fmt.Errorf("verify: %s: arg block too small: got %d, need %d", name, len(args), fl.Args)
+	}
+	fnAddr := k.exec.FuncAddr(fl.Offset)
+	return Call(fnAddr, args)
+}
+
+// Close releases the executable mapping.
+func (k *Kernel) Close() {
+	if k.exec != nil {
+		k.exec.Unmap()
+	}
+}