Repository navigation
Improve parser diagnostics and report unknown type names - #162
Merged
Merged
Conversation
Replace the code-hint block with fixes attached to help and note lines. A fix is built with Fix.Remove, Fix.Insert or Fix.Replace and drawn as one source line: the line as it will read with added text marked, or the current line with removed text marked. Fixes that cannot be shown faithfully on single lines are not drawn. Attach fixes to unneeded mut, missing separators and closing brackets, redundant separators, trailing and match-arm commas, redundant optional markers and leading zeros. Other missing tokens keep their label. Keep the margin unbroken before help lines and end every diagnostic with one blank line. Known gap: a missing-token fix follows the parser's guess, so it can show a wrong line when more code follows on the same line.
Name the token that is missing and show the repaired line where the repair is certain: ':' or '=' in let and const, ':' after a parameter name, '->' before a return type, '=>' in a match arm, ',' in calls, parameters, function types, type parameters, attribute arguments and braced lists, and a closing bracket at the end of a line or before another closer. Several missing commas in one list are one error. Report a token typed for another with a replacement: ';' for ',' in lists, ':' for '=' in a struct literal, '=' for '==' in a condition. Where the repair cannot be told, show both: `let a b;` offers ':' and '='. One word after `mut` is reported without guessing whether it is the name or the type, a repeated `mut` is reported once, and a parameter name without a type keeps the rest of the list. An unclosed call no longer takes the next statement as an argument or drops it. A ';' in a value list closes the list only when its own '}' is missing. A trailing comma before ')' or '>' is the same note as in literals instead of an error. Rename Diagnostic.WithHelp and WithNote to Help and Note, and add HelpWithChoices for repairs that depend on what the author meant.
A type name that is neither declared nor built in used to become a silent placeholder, so `let x: Nope;` compiled and other uses failed with unrelated messages. Report "unknown type" where the name is written, with a similar declared name when one is close. The unknown name becomes the invalid type and keeps its text for display. Checks that meet a type with an invalid part stay silent: the size check, array literal typing, and ownership tracking, which also stops the module-binding error after an unknown constant value. A type query treats a type whose only fault is an unknown name as showable. Reject type parameters on functions with one message; the language does not have generic functions yet. The entry-point test no longer covers that case because it is rejected earlier. Qualified names that do not resolve keep their placeholder.
Cover a shift count equal to the operand width, valid narrow and odd width shift counts, the generated shift guard for each count width, and a left operand read before a call on the right changes it.
Where a separator or a closing bracket is expected but inserting it is not a reliable repair, mark the token that was found and propose nothing. The old label told the reader to add the token at a place that was often wrong. When a closing bracket is missing and its bracket was opened on an earlier line, also mark where it was opened.
The printer drew every label as its own block: the line above it, the labelled lines, then the underline. A second label on the next line repeated that line as context, two labels on one line printed the line twice, and labels over the same lines printed all of them again. Lay the labels out by line instead. Each source line is printed once with a row for every label on it, the line above is shown only when it is not on screen yet, and `...` appears only where lines are really left out. The location header names the primary label instead of the first label in file order. A diagnostic with one label prints as before.
A source file with CRLF line ends kept the `\r` on every line read for diagnostics, so printed source lines and fixed lines ended with a stray carriage return. Trim it in the two places that split source text: the diagnostic source cache and GetSourceLines.
The printer placed a marker one column per character, so it sat too far left after East Asian wide characters and emoji, which take two terminal cells, and too far right after combining marks, which take none. Count cells for marker position, underline length and tab stops. A label that runs to the end of a line counted bytes; count characters. A name with letters outside ASCII was one "illegal character" error per character and broke the statement around it. Report the whole word once and read it as a name so parsing continues. Text outside ASCII in strings, character literals and comments stays allowed.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
:/=inletandconst,:after a parameter name,->,=>,,in lists, closing brackets). Reports a token typed for another with a replacement (;for,,:for=in a struct literal,=for==in a condition). Where the repair cannot be told it shows both (let a b;) or points at the token found and proposes nothing.let x: Nope;compiled.WithHelp/WithNoterenamed toHelp/Note; newHelpWithChoices; fixed lines are drawn beneath help lines; the printer prints each source line once when a diagnostic has several labels.internal/frontend/parser,internal/diagnostics,internal/semantics/typeinfo,typeresolution,typechecker,binder,analysis(ownership).pkg::Missing) are unchanged. No redesign of the diagnostic format.Production ownership
Diagnostic.HelpWithChoices,Choice:Parser.parseBindingFields.Fix.Replace(existing, first callers):parseBracedItemList,parseCompositeLiteral,parseIfStmt.listContinues,reportMissingCommas,reportTrailingComma,startsArgument,startsValue,closerAhead,isSemicolonForComma,isBraceClosedBeforeArrow,parseReturnType,parseParamModifiers,startsNamedParam,withOpener,insertionError,isMissingTokenLikely,hasBeforeStatementEnd,isLoneNameAfterBindingName,isTypeWithoutColon,isOnLineOfPrev,isTypeStart: all called from the list, binding, parameter, match and close-bracket parsers.typeinfo.SyntaxUnknownName,InvalidType.Name:resolveTypeName, reported bytyperesolution.reportSyntaxIssues.typeinfo.ContainsMalformed:Resolver.Query.declaredTypeNames:reportSyntaxIssues.snippetFiles,printSnippetLine,contextLine:Emitter.Emit.peeper check|build|run→ parser → binder/typechecker (type syntax resolution) →DiagnosticBag.EmitAll→Emitter.Emit.SyntaxIssuelist,InvalidType,ContainsInvalid,diagnostics.NearestName, and the existing fix drawing.rejectUnsizedType,typeArrayLitandownershipTrackedTypenow stay silent on a type with an invalid part.WithHelp/WithNoteare renamed with every caller, no alias kept.printPeeperSnippetBlockandprintPrevNonEmptyLineare replaced and deleted.Design and behavior
Xhere" at a place that was wrong, or cascaded (for example.{ x = 1; y = 2 }gave 7 errors after an earlier draft of this work, 1–2 before). An unclosed call could take the next statement as an argument or drop it. An unknown type name became a silent placeholder. A diagnostic with two labels printed shared lines twice and could put the header on the secondary label.)or>in calls, parameters, function types, type parameters and attribute arguments is now the same note as in literals (S0001) instead of an error, so such code compiles.)is still ahead.Validation
Run on the final tree (
0fadf11), locally on Linux amd64.negative_missing_token_at_known_spot,negative_missing_token_unreliable_guess,negative_missing_comma_in_lists,negative_semicolon_in_value_list,negative_parameter_without_type,negative_single_repair_spots,negative_let_missing_separator,negative_missing_closers,negative_bang_after_argument,negative_found_token_and_open_bracket,negative_unknown_names,negative_unknown_type_name,negative_unknown_const_value,positive_multiline_lists_and_continuations,llvm_shift_guard_exact_bound,llvm_shift_guard_narrow_valid,runtime_mir_rvalue_eval_order_binary_capture. Updated:negative_struct_literal_value_prefix(now "unknown typepoint").6ff091e, and every diagnostic was read. More errors than before appear only where a real follow-on error or an unknown type is now reported.Diff sanity
expectClosecallsisMissingTokenLikelyaftermissingTokenErrorhas called it with the same arguments. The typechecker's two parameter loops each skip a parameter with no written type.Risks and follow-up
fn Add(i32 i32)still suggests:;Id(a !b)gets no repair; an unclosed bracket before{(if f(a {) still cascades.const c: [2]Nope = [2]Nope{1, 2};reports only the firstNope. "not all control paths return a value" draws the same span as a primary and a secondary label. Inside a function with type parameters,let x: Talso says "unknown typeT".