<html>
<head>
<meta http-equiv="Content-Type" content="text/html; charset=utf-8">
</head>
<body text="#000000" bgcolor="#FFFFFF">
<p>This would be helpful. Within Coverity, we hacked together our own error recovery mechanism (mostly involving demoting errors to warnings and discarding function bodies and variable initializers with errors), but our mechanism sometimes leads to other down
stream errors (I investigated a spurious "enumerator 'x' does not exist in instantiation of 'y' just this week). The ability to explicitly indicate errors in the AST would be helpful.</p>
<p>Ideally, I'd like to have ErrorDecl, ErrorStmt, ErrorExpr, and ErrorType nodes. But there is no reason they would all have to be introduced at the same time.<br>
</p>
<p>Tom.<br>
</p>
<div class="moz-cite-prefix">On 5/9/2019 7:31 AM, Sam McCall via cfe-dev wrote:<br>
</div>
<blockquote type="cite" cite="mid:CA+sYUsJn8bM=BAhM1HgOKg9xLsD1KLmi9h3PRp6mbNauc1kCVw@mail.gmail.com">
<div dir="ltr">In the presence of broken code, the AST clang generates is often very limited.
<div>Often Sema diagnoses errors and doesn't create AST nodes. For example, if the arguments to a function call are wrong, we get no CallExpr, and all subexpressions are discarded.<br>
</div>
<div>IDE tools like clangd see a lot of broken code, and rely on the resulting AST, so I'd like to improve this situation.<br>
</div>
<div><br>
</div>
<div>If the user writes the following erroneous code:</div>
<div> a(b, c).d() // error: a only takes one parameter</div>
<div>then no AST is emitted for this expression, and:</div>
<div>1) clangd's "go to definition" doesn't work on a, even though there's no overloading</div>
<div>2) "go to definition" also doesn't work on b and c or their subexpressions, even though these expressions are valid</div>
<div>3) code completion doesn't work properly for d(), even though we know what type a() returns. (go to definition doesn't either).</div>
<div>There are lots of similar cases: (foo->bar when bar doesn't exist, etc).<br>
</div>
<div><br>
</div>
<div>My proposal is to add a fairly generic RecoveryExpr that the AST can emit in these situations. It would capture:</div>
<div> - the approximate StmtClass we were trying to create (e.g. CallExpr)</div>
<div> - the subexpressions - (e.g. an overloadexpr for the callee, and the arguments)</div>
<div> - the type, if we can make a reasonable guess (e.g. all overloads have the same type)</div>
<div><br>
</div>
<div>I also considered other approaches, and think they don't work so well:</div>
<div> - just generate the e.g. CallExpr etc anyway, and maybe add an "invalid" flag. This means weakening many invariants in many different node types. It's subtle and error-prone, and adds conceptual and implementation complexity to many nodes that now have
to be able to represent various error cases. Even within clang it's hard to find all the analyses that need to be updated. (<a href="https://urldefense.proofpoint.com/v2/url?u=http-3A__lists.llvm.org_pipermail_cfe-2Ddev_2019-2DApril_062151.html&d=DwMFaQ&c=DPL6_X_6JkXFx7AXWqB0tg&r=oOJsFtgkFCCcDlfLWKmfrVDtpXLihXJTBGJfSapXCV0&m=k3m3ocpqnJDQ1DBLtX6DNqMbG51U4RJ9VyH8qj2SfSc&s=jddDcS9V5Twt16Gxg9n1WTXzsxqOYx-KQIOIvXEovEc&e=" moz-do-not-send="true">http://lists.llvm.org/pipermail/cfe-dev/2019-April/062151.html</a>)</div>
<div> - extract the extra information we need from Sema, rather than the AST. e.g. Sema should invoke a callback whenever an identifier is resolved to a Decl. This solves go-to-definition pretty well, but isn't a very flexible approach. The AST is rich, and
provides context, can be dumped etc, a Sema callback would be an ad-hoc solution that is less convenient to use and understand.</div>
<div> - add a hierarchy of new nodes to represent errors (RecoveryCallExpr etc). It seems like a lot of work to model all the different ways code can be invalid, and I don't think the extra structure would be useful enough to justify it.</div>
<div><br>
</div>
<div>I have a very rough prototype of this idea: <a href="https://urldefense.proofpoint.com/v2/url?u=https-3A__reviews.llvm.org_D61722&d=DwMFaQ&c=DPL6_X_6JkXFx7AXWqB0tg&r=oOJsFtgkFCCcDlfLWKmfrVDtpXLihXJTBGJfSapXCV0&m=k3m3ocpqnJDQ1DBLtX6DNqMbG51U4RJ9VyH8qj2SfSc&s=x4VyJNvrS0sFh72jDS_z6ox6SL4zxZLE4lIP3i7tnLc&e=" target="_blank" moz-do-not-send="true">https://reviews.llvm.org/D61722</a> (emits
RecoveryExpr for failed overload resolution).</div>
<div>It does create a form of recovery that affects diagnostics. But the changes exposed by tests are few and seem mostly good (especially compared to emitting CallExpr).</div>
<div>Some open questions:</div>
<div> - what's the type of the expression when we can't guess? (IntTy in the patch)</div>
<div> - will there be surprises if RecoveryExpr is emitted in more places and they start to interact?</div>
<div> - should RecursiveASTVisitor, ASTMatchers etc skip these nodes by default, for compatibility with external code?</div>
<div> - what are the important use cases that I haven't thought about?</div>
<div><br>
</div>
</div>
<br>
<fieldset class="mimeAttachmentHeader"></fieldset>
<pre class="moz-quote-pre" wrap="">_______________________________________________
cfe-dev mailing list
<a class="moz-txt-link-abbreviated" href="mailto:cfe-dev@lists.llvm.org">cfe-dev@lists.llvm.org</a>
<a class="moz-txt-link-freetext" href="https://urldefense.proofpoint.com/v2/url?u=https-3A__lists.llvm.org_cgi-2Dbin_mailman_listinfo_cfe-2Ddev&d=DwIGaQ&c=DPL6_X_6JkXFx7AXWqB0tg&r=oOJsFtgkFCCcDlfLWKmfrVDtpXLihXJTBGJfSapXCV0&m=k3m3ocpqnJDQ1DBLtX6DNqMbG51U4RJ9VyH8qj2SfSc&s=pBwBCrRoYbCV6AZvGdWTEDd5gKfDDajyTQ8VRoQ38LY&e=">https://urldefense.proofpoint.com/v2/url?u=https-3A__lists.llvm.org_cgi-2Dbin_mailman_listinfo_cfe-2Ddev&d=DwIGaQ&c=DPL6_X_6JkXFx7AXWqB0tg&r=oOJsFtgkFCCcDlfLWKmfrVDtpXLihXJTBGJfSapXCV0&m=k3m3ocpqnJDQ1DBLtX6DNqMbG51U4RJ9VyH8qj2SfSc&s=pBwBCrRoYbCV6AZvGdWTEDd5gKfDDajyTQ8VRoQ38LY&e=</a>
</pre>
</blockquote>
</body>
</html>