feat(asm): assemble the extended instruction layer on arm64

Assisted-by: GLM 5.3 Flash

asm/arm64_assemble.go

@@ -9,6 +9,7 @@ import (
 	"strconv"
 	"strings"
 
+	"sourcedock.dev/petrbalvin/gasm-sdk/arch"
 	"sourcedock.dev/petrbalvin/gasm-sdk/ast"
 )
 
@@ -269,6 +270,13 @@ func arm64InstrSize(instr *ast.Instr, fi arm64FrameInfo, pos int) int {
 	case "NO_LOCAL_POINTERS", "GO_ARGS", "GO_RESULTS_INITIALIZED", "END", "FUNCDATA", "PCDATA":
 		return 0
 	}
+	// The extended-instruction layer is one instruction word in every form
+	// the registry takes: pass 1 must size a pinned statement at the 4 bytes
+	// the encoder will lay down, ahead of the scalar immediate expansions
+	// below, whose sizes would misread a Z destination (asm/arm64_ext.go).
+	if _, pinned, _ := arm64ExtStatement(mnem, ops); pinned {
+		return 4
+	}
 	switch mnem {
 	case "MOV", "MOVD", "MOVW", "MOVWU", "MOVH", "MOVHU", "MOVB", "MOVBU",
 		"FMOVS", "FMOVD":
@@ -307,6 +315,23 @@ func encodeARM64Instr(instr *ast.Instr, pc int, offsets map[string]int, fi arm64
 	mnem := strings.ToUpper(instr.Mnemonic.Text)
 	ops := instr.Operands
 
+	// The extended-instruction layer: a statement whose mnemonic is
+	// registered in the extension registry and whose operands carry a
+	// scalable vector or predicate register encodes through the registry,
+	// before any scalar route can misread those operands.  Scalar, NEON and
+	// FP operand lists never pin, so everything below runs exactly as it
+	// did (asm/arm64_ext.go).
+	if extops, pinned, convErr := arm64ExtStatement(mnem, ops); pinned {
+		if convErr != nil {
+			return nil, convErr
+		}
+		code, encErr := EncodeExtension(arch.ARM64, mnem, extops...)
+		if encErr != nil {
+			return nil, encErr
+		}
+		return code, nil
+	}
+
 	// Pseudo-instructions and special cases first.
 	switch mnem {
 	case "RET":