gasm-sdk/lint/lint.go

// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
// SPDX-License-Identifier: BSD-3-Clause

// Package lint runs static checks over a parsed GAsm file.  The rules are
// deliberately conservative: where a check cannot be certain (for example an
// instruction whose operand count varies), it stays silent rather than emit a
// false positive.  Every diagnostic carries a stable rule code so callers can
// disable individual rules.
package lint

import (
	"fmt"
	"strings"

	"sourcedock.dev/petrbalvin/gasm-devkit/arch"
	"sourcedock.dev/petrbalvin/gasm-devkit/ast"
	"sourcedock.dev/petrbalvin/gasm-devkit/token"
)

// Severity ranks a diagnostic.
type Severity int

// Diagnostic severities, mirroring the language-server protocol ordering.
const (
	Error Severity = iota
	Warning
	Information
	Hint
)

// String returns a lower-case label for the severity.
func (s Severity) String() string {
	switch s {
	case Error:
		return "error"
	case Warning:
		return "warning"
	case Information:
		return "information"
	default:
		return "hint"
	}
}

// Diagnostic is one lint finding.
type Diagnostic struct {
	Pos      token.Position
	End      token.Position
	Severity Severity
	Code     string
	Message  string
}

// Config controls a lint run.
type Config struct {
	// Arch is the target architecture.  When it is arch.Unknown the
	// architecture-specific rules (unknown instruction, operand count) are
	// skipped because no instruction table can be selected.
	Arch arch.Arch
	// Disable lists rule codes to suppress.
	Disable map[string]bool
}

// Rule codes.
const (
	CodeUnknownInstr    = "unknown-instruction"
	CodeOperandCount    = "operand-count"
	CodeUndefinedLabel  = "undefined-label"
	CodeDuplicateLabel  = "duplicate-label"
	CodeMissingRet      = "missing-ret"
	CodeMissingTextflag = "missing-textflag-include"
	CodeUnreachable     = "unreachable-code"
	CodeABIArgSize      = "abi-argsize"
	CodeNosplitFrame    = "nosplit-frame"
	CodeRegisterClobber = "register-clobber"
	CodeFuncdata        = "funcdata-pcdata"
)

// pseudoOps are assembler pseudo-operations that are valid instruction-position
// tokens but are not machine instructions and so absent from the arch tables.
var pseudoOps = map[string]bool{
	"BYTE": true, "WORD": true, "LONG": true, "QUAD": true, "FLOAT": true,
	"PCALIGN": true, "FUNCDATA": true, "PCDATA": true, "GO_ARGS": true,
}

// File lints a parsed file and returns the diagnostics in source order.
func File(f *ast.File, cfg Config) []Diagnostic {
	var out []Diagnostic
	tab := arch.ForArch(cfg.Arch)
	archKnown := cfg.Arch != arch.Unknown

	hasTextflag := false
	usesFlags := false
	var firstFlagPos token.Position

	// Macros (in-file #define, or any #include other than textflag.h, which
	// only defines flag constants) make label resolution unreliable.
	macrosInPlay := len(f.Macros) > 0
	// Preprocessor conditionals (#ifdef …) make control-flow analysis
	// unreliable, since mutually exclusive branches look sequential.
	hasConditionals := false
	for _, d := range f.Decls {
		switch dd := d.(type) {
		case *ast.Include:
			if !strings.Contains(dd.Header.Text, "textflag.h") {
				macrosInPlay = true
			}
		case *ast.Preproc:
			if isConditionalDirective(dd.Raw) {
				hasConditionals = true
			}
		}
	}

	for _, d := range f.Decls {
		switch dd := d.(type) {
		case *ast.Include:
			if strings.Contains(dd.Header.Text, "textflag.h") {
				hasTextflag = true
			}
		case *ast.Text:
			out = append(out, lintText(dd, tab, archKnown, cfg, f.Macros, !macrosInPlay, !hasConditionals)...)
			if len(dd.Flags) > 0 && !firstFlagPos.IsValid() {
				usesFlags = true
				firstFlagPos = dd.Pos()
			}
		case *ast.Globl:
			if len(dd.Flags) > 0 && !firstFlagPos.IsValid() {
				usesFlags = true
				firstFlagPos = dd.Pos()
			}
		}
	}

	if !cfg.Disable[CodeMissingTextflag] && usesFlags && !hasTextflag {
		out = append(out, Diagnostic{
			Pos:      firstFlagPos,
			Severity: Warning,
			Code:     CodeMissingTextflag,
			Message:  "TEXT/GLOBL flags are used but textflag.h is not #included",
		})
	}

	sortDiagnostics(out)
	return out
}

// lintText lints one TEXT function body.  doLabelChecks is false for files
// that use macros (an in-file #define or a non-textflag #include): without a
// preprocessor we cannot resolve labels that macros define or reference, so the
// label and RET heuristics are suppressed there to avoid false positives.
func lintText(t *ast.Text, tab *arch.Table, archKnown bool, cfg Config, macros map[string]bool, doLabelChecks bool, doUnreachable bool) []Diagnostic {
	var out []Diagnostic

	defined := map[string]token.Position{}
	referenced := map[string]token.Position{}
	hasRet := false
	lastTerminal := false
	hasMacro := false
	instrCount := 0
	dead := false                       // inside a region unreachable from above
	reportedDead := false               // the current dead region has already been reported
	hasPCRel := referencesPC(t)         // PC-relative jumps defeat reachability analysis
	hasIndirect := hasIndirectBranch(t) // register-indirect branches do too
	// Unreachable-code analysis is only sound in functions whose control flow is
	// fully label-resolvable: no PC-relative jumps, no register-indirect
	// branches, and (file-level) no preprocessor conditionals.
	analyzable := doUnreachable && !hasPCRel && !hasIndirect

	for _, s := range t.Body {
		switch st := s.(type) {
		case *ast.Label:
			name := st.Name.Text
			if prev, dup := defined[name]; dup {
				if !cfg.Disable[CodeDuplicateLabel] {
					out = append(out, Diagnostic{
						Pos:      st.Name.Pos,
						End:      st.Name.End,
						Severity: Error,
						Code:     CodeDuplicateLabel,
						Message:  fmt.Sprintf("label %q already defined at %s", name, prev),
					})
				}
			} else {
				defined[name] = st.Name.Pos
			}
			// A label is a jump target: code after it is reachable again.
			dead = false
			reportedDead = false

		case *ast.Instr:
			instrCount++
			mnem := st.Mnemonic.Text
			upper := strings.ToUpper(mnem)

			// Unreachable code: a real instruction following a RET/UNDEF and
			// before any label, in a function whose control flow is fully
			// resolvable.  Only RET/UNDEF are treated as terminators here — an
			// unconditional jump may be one entry of a hand-arranged branch
			// table (e.g. the generated callback tables), so it is not assumed
			// to make the following code dead.  Pseudo-ops and macro invocations
			// are never flagged.
			if analyzable && dead && !reportedDead && !pseudoOps[upper] && !isMacroInvocation(mnem, macros) &&
				!cfg.Disable[CodeUnreachable] {
				out = append(out, Diagnostic{
					Pos:      st.Mnemonic.Pos,
					End:      st.Mnemonic.End,
					Severity: Warning,
					Code:     CodeUnreachable,
					Message:  "unreachable code after terminating instruction",
				})
				reportedDead = true
			}

			// RET never falls through.  (UNDEF is a trap/marker rather than a
			// control-flow terminator: code placed after it is occasionally
			// deliberate metadata, so it is not treated as making the following
			// code dead.)
			if upper == "RET" {
				dead = true
			}
			// A function need not RET if it ends in an unconditional jump (tail
			// call / loop) or in UNDEF (a deliberate trap that never returns).
			lastTerminal = isUnconditionalJump(cfg.Arch, upper) || upper == "UNDEF"
			if isMacroInvocation(mnem, macros) {
				hasMacro = true
			}

			if upper == "RET" {
				hasRet = true
			}

			if archKnown && !cfg.Disable[CodeUnknownInstr] && !pseudoOps[upper] && !isMacroInvocation(mnem, macros) {
				if _, ok := tab.Lookup(mnem); !ok {
					out = append(out, Diagnostic{
						Pos:      st.Mnemonic.Pos,
						End:      st.Mnemonic.End,
						Severity: Error,
						Code:     CodeUnknownInstr,
						Message:  fmt.Sprintf("unknown %s instruction %q", cfg.Arch, mnem),
					})
				}
			}

			if archKnown && !cfg.Disable[CodeOperandCount] && !isMacroInvocation(mnem, macros) {
				if in, ok := tab.Lookup(mnem); ok && in.MinOps >= 0 {
					n := len(st.Operands)
					if n < in.MinOps || n > in.MaxOps {
						out = append(out, Diagnostic{
							Pos:      st.Mnemonic.Pos,
							End:      st.Mnemonic.End,
							Severity: Warning,
							Code:     CodeOperandCount,
							Message: fmt.Sprintf("%s expects %s, got %d operand(s)",
								mnem, countRange(in.MinOps, in.MaxOps), n),
						})
					}
				}
			}

			if isJump(cfg.Arch, upper) {
				for _, op := range st.Operands {
					if name, pos, ok := localLabelRef(op); ok && !tab.IsRegister(name) && !arch.IsPseudoReg(name) {
						referenced[name] = pos
					}
				}
			}
		}
	}

	// Undefined labels.
	if doLabelChecks && !cfg.Disable[CodeUndefinedLabel] {
		for name, pos := range referenced {
			if _, ok := defined[name]; !ok {
				out = append(out, Diagnostic{
					Pos:      pos,
					Severity: Error,
					Code:     CodeUndefinedLabel,
					Message:  fmt.Sprintf("jump to undefined label %q", name),
				})
			}
		}
	}

	// Missing RET heuristic.  Functions that invoke a macro are skipped: the
	// macro body (opaque to us) may supply the RET.
	if doLabelChecks && !cfg.Disable[CodeMissingRet] && instrCount > 0 && !hasRet && !lastTerminal && !hasMacro {
		out = append(out, Diagnostic{
			Pos:      t.Keyword.Pos,
			Severity: Warning,
			Code:     CodeMissingRet,
			Message:  fmt.Sprintf("function %q has no RET", t.Name.Name),
		})
	}

	// ABI conformance: the argument area declared in the TEXT directive should
	// match the size computed from the // func signature in the doc comment.
	// Only applies to stack-argument (ABI0) functions, which reference their
	// arguments through FP; register-ABI functions declare a zero arg area.  Also
	// skipped when there is no parseable signature or it uses an unknown type.
	if !cfg.Disable[CodeABIArgSize] {
		got := int64(0)
		if t.Args != nil && t.Args.Imm.HasVal {
			got = t.Args.Imm.Val
		}
		// Only meaningful for stack-argument (ABI0) functions: a non-zero
		// declared arg area that is actually addressed through FP.
		if got > 0 && usesFPArgs(t) {
			if want, ok := abiExpectedArgSize(t.Doc); ok {
				if want != got {
					out = append(out, Diagnostic{
						Pos:      t.Keyword.Pos,
						Severity: Warning,
						Code:     CodeABIArgSize,
						Message:  fmt.Sprintf("TEXT declares arg size %d but the // func signature implies %d", got, want),
					})
				}
			}
		}
	}

	// Register liveness: a callee-saved register that is written but never
	// saved and restored is clobbered across the call.  The check runs over the
	// control-flow graph and is skipped for macro-using files, where an opaque
	// macro may perform the save/restore.
	if doLabelChecks && archKnown && !cfg.Disable[CodeRegisterClobber] {
		live := analyzeLiveness(t, cfg.Arch)
		if clobbered := clobberedCalleeSaved(live, cfg.Arch); len(clobbered) > 0 {
			out = append(out, Diagnostic{
				Pos:      t.Keyword.Pos,
				Severity: Warning,
				Code:     CodeRegisterClobber,
				Message:  fmt.Sprintf("callee-saved register(s) %s written but never saved/restored", strings.Join(clobbered, ", ")),
			})
		}
	}

	// FUNCDATA / PCDATA structural validation.
	out = append(out, checkFuncdata(t, cfg)...)

	return out
}

// usesFPArgs reports whether a function references its arguments through the FP
// pseudo-register — i.e. it uses the stack-based ABI0 layout, where the
// declared argument size must match the signature.
func usesFPArgs(t *ast.Text) bool {
	for _, s := range t.Body {
		in, ok := s.(*ast.Instr)
		if !ok {
			continue
		}
		for _, op := range in.Operands {
			if op.Addr.Sym != nil && op.Addr.Sym.Pseudo == "FP" {
				return true
			}
		}
	}
	return false
}

// referencesPC reports whether a function uses a PC-relative operand (e.g.
// `JMP 2(PC)`).  Such jumps target a computed offset rather than a label, so
// reachability cannot be determined statically and the unreachable-code check
// is suppressed for the whole function.
func referencesPC(t *ast.Text) bool {
	for _, s := range t.Body {
		in, ok := s.(*ast.Instr)
		if !ok {
			continue
		}
		for _, op := range in.Operands {
			if strings.Contains(strings.ReplaceAll(op.Raw, " ", ""), "(PC)") {
				return true
			}
		}
	}
	return false
}

// hasIndirectBranch reports whether a function transfers control through a
// register (JALR/JR/JIRL/BR/BLR).  Such targets are computed at runtime, so
// reachability cannot be determined statically and the unreachable-code check is
// suppressed for the whole function.
func hasIndirectBranch(t *ast.Text) bool {
	for _, s := range t.Body {
		in, ok := s.(*ast.Instr)
		if !ok {
			continue
		}
		switch strings.ToUpper(in.Mnemonic.Text) {
		case "JALR", "JR", "JIRL", "BR", "BLR":
			return true
		}
	}
	return false
}

// isMacroInvocation reports whether a mnemonic is a macro invocation rather
// than a machine instruction.  No Plan 9 mnemonic contains an underscore, so an
// underscore is a reliable macro marker (the runtime headers define macros such
// as get_tls and NO_LOCAL_POINTERS).  Names introduced by an in-file #define
// are recognised too (CALLFN, DISPATCH, …).  Full macro expansion is out of
// scope; this only keeps the linter quiet on invocations it cannot expand.
func isMacroInvocation(mnem string, macros map[string]bool) bool {
	return strings.Contains(mnem, "_") || macros[mnem]
}

// isConditionalDirective reports whether a preprocessor directive (the text
// after '#') is a conditional-compilation directive whose branches the parser
// cannot resolve.
func isConditionalDirective(raw string) bool {
	fields := strings.Fields(raw)
	if len(fields) == 0 {
		return false
	}
	switch fields[0] {
	case "if", "ifdef", "ifndef", "else", "elif", "endif":
		return true
	}
	return false
}

// localLabelRef returns the name and position of a bare local-label reference
// operand (no pseudo-register, no memory base), if op is one.
func localLabelRef(op *ast.Operand) (string, token.Position, bool) {
	if op == nil || op.Kind != ast.OpAddr || op.Addr.Sym == nil {
		return "", token.Position{}, false
	}
	sym := op.Addr.Sym
	if sym.Pseudo != "" || op.Addr.Base != "" || sym.Name == "" {
		return "", token.Position{}, false
	}
	return sym.Name, op.Pos, true
}

// riscvBranches and loong64Branches are the conditional-branch mnemonics; they
// are listed explicitly rather than matched by a "B" prefix so that bit-manip
// instructions (BCLR, BSET, …) are never mistaken for branches.
var riscvBranches = map[string]bool{
	"BEQ": true, "BNE": true, "BLT": true, "BGE": true, "BLTU": true, "BGEU": true,
	"BEQZ": true, "BNEZ": true, "BLEZ": true, "BGEZ": true, "BLTZ": true, "BGTZ": true,
}

var loong64Branches = map[string]bool{
	"BEQ": true, "BNE": true, "BLT": true, "BGE": true, "BLTU": true, "BGEU": true,
	"BLEZ": true, "BLTZ": true, "BGEZ": true, "BGTZ": true,
}

// isJump reports whether the mnemonic is any branch.
func isJump(a arch.Arch, upper string) bool {
	switch a {
	case arch.ARM64:
		return upper == "CALL" || upper == "BR" || upper == "BLR" || upper == "JMP" ||
			strings.HasPrefix(upper, "B") ||
			strings.HasPrefix(upper, "CBZ") || strings.HasPrefix(upper, "CBNZ") ||
			strings.HasPrefix(upper, "TBZ") || strings.HasPrefix(upper, "TBNZ")
	case arch.RISCV:
		return upper == "CALL" || riscvBranches[upper] ||
			upper == "JMP" || upper == "J" || upper == "JAL" || upper == "JALR" ||
			upper == "JR" || upper == "BR"
	case arch.LOONG64:
		return upper == "CALL" || loong64Branches[upper] ||
			upper == "JIRL" || upper == "JMP" || upper == "BR"
	default: // amd64
		return upper == "CALL" || strings.HasPrefix(upper, "J")
	}
}

// isUnconditionalJump reports whether the mnemonic is an unconditional branch
// (used to suppress the missing-RET heuristic for tail calls and loops).
func isUnconditionalJump(a arch.Arch, upper string) bool {
	switch a {
	case arch.ARM64:
		return upper == "B" || upper == "BR" || upper == "JMP"
	case arch.RISCV:
		return upper == "JMP" || upper == "J" || upper == "JAL" ||
			upper == "JALR" || upper == "JR" || upper == "BR"
	case arch.LOONG64:
		return upper == "JMP" || upper == "JIRL" || upper == "BR"
	default:
		return upper == "JMP"
	}
}

func countRange(min, max int) string {
	if min == max {
		return fmt.Sprintf("%d operand(s)", min)
	}
	return fmt.Sprintf("%d–%d operands", min, max)
}

// sortDiagnostics orders diagnostics by line, then column, then code.
func sortDiagnostics(d []Diagnostic) {
	for i := 1; i < len(d); i++ {
		for j := i; j > 0 && lessDiag(d[j], d[j-1]); j-- {
			d[j], d[j-1] = d[j-1], d[j]
		}
	}
}

func lessDiag(a, b Diagnostic) bool {
	if a.Pos.Line != b.Pos.Line {
		return a.Pos.Line < b.Pos.Line
	}
	if a.Pos.Column != b.Pos.Column {
		return a.Pos.Column < b.Pos.Column
	}
	return a.Code < b.Code
}