Bug 3379 – [tdpl] Parameter names not visible in the if clause of a template

This code does not compile: T1[] find(T1, T2)(T1[] longer, T2[] shorter) if (is(typeof(longer[0 .. 1] == shorter) : bool)) { while (longer.length >= shorter.length) { if (longer[0 .. shorter.length] == shorter) break; longer = longer[1 .. $]; } return longer; } unittest { double[] d1 = [ 6.0, 1.5, 2.4, 3 ]; double[] d2 = [ 1.5, 2.4 ]; assert(find(d1, d2) == d1[1 .. $]); } (I believe I'd submitted this bug already. Please don't mark as duplicate so all TDPL-related errors stay together.)

I've convinced myself that this isn't too difficult, but it contains a lot of subtlties. I don't have a complete patch, but I've got basic cases working. Basically in deduceFunctionTemplateMatch() you need to create variables for each of the parameters (by analogy to func.c line 904). for (i = 0; i < nfparams; i++) { Argument *fparam = Argument::getNth(fparameters, i); if (!fparam->ident) continue; // don't add it, if it has no name Type *vtype = // get the type, not sure how to do this properly ... deal with pure, etc. VarDeclaration *v = new VarDeclaration(loc, vtype, fparam->ident, NULL); ... set the storage class v->semantic(paramscope); if (!paramscope->insert(v)) error("parameter %s.%s is already defined", toChars(), v->toChars()); } Not sure how to deal with deal with the different flavours of variadics, though. I originally thought that in TemplateDeclaration::matchWithInstance() you'd be able avoid checking the constraint again if it's a function template, but you can't because of cases like: T1[] find(T1, T2)(T1[] longer, T2[] shorter) if (false) { assert(0); return longer; } void main() { double[] d1 = [ 6.0, 1.5, 2.4, 3 ]; double[] d2 = [ 1.5, 2.4 ]; assert(find!(double, double)(d1, d2) == d1[1 .. $]); // THIS MUST NOT COMPILE. } So it needs to have the same thing, wrapped in a FuncDeclaration *fd = onemember->toAlias()->isFuncDeclaration(); if (fd) { ... }

Created attachment 474 Patch against 2.035 svn 215

Since the patch contains two nearly-identical parts they probably should be factored out into a function. But it works. Test cases for the patch: ------ / bug3379 T1[] find(T1, T2)(T1[] longer, T2[] shorter) if (is(typeof(longer[0 .. 1] == shorter) )) { while (longer.length >= shorter.length) { if (longer[0 .. shorter.length] == shorter) break; longer = longer[1 .. $]; } return longer; } // bug2983 auto max(T...)(T a) if (T.length == 2 && is(typeof(a[1] > a[0] ? a[1] : a[0])) || T.length > 2 && is(typeof(max(max(a[0], a[1]), a[2 .. $])))) { static if (T.length == 2) { return a[1] > a[0] ? a[1] : a[0]; } else { return max(max(a[0], a[1]), a[2 .. $]); } } void main() { assert(max(4, 5) == 5); assert(max(3, 4, 5) == 5); double[] d1 = [ 6.0, 1.5, 2.4, 3 ]; double[] d2 = [ 1.5, 2.4 ]; assert(find(d1, d2) == d1[1 .. $]); } void bad(T)(T x) if (is(T : idouble)) { } static assert(!is(typeof(bad(17)))); // must be rejected static assert(!is(typeof(bad!(int)(2)))); // reject even if types specified static assert(is(typeof(bad(3.0i)))); void fun(T)(char a, T[] b...) if (is(typeof(b[0]>b[1]))) {} static assert(is(typeof(fun('x', 3.0, 4.0, 5.0))));

*** Issue 2983 has been marked as a duplicate of this issue. ***

Created attachment 480 Revised patch against 2.035 svn 215. My patch had a nasty segfault regression. Trivial though. The line: FuncDeclaration *fd = onemember->toAlias()->isFuncDeclaration(); if (fd) causes segfaults if either onemember or toAlias() returns NULL. fd should be NULL in either of those cases.

Unfortunately my patch still isn't complete. The way I'm setting the storage class of the parameters sometimes interferes with the later code. This surpises me since I thought that the paramscope->pop() would undo all changes. Clearly, there's something about the compiler I don't yet understand.

Created attachment 493 Patch against svn 241 I've moved the implementation to a separate function. I fixed a couple of bugs; it now compiles the Phobos test suite. Currently, it only works with free functions; I've disabled it for templated member functions (enabling it causes internal compiler errors). There's probably something I need to do with the variable declarations to make it work.

Created attachment 494 Test cases Some test cases I've used. Includes tests which previously caused segfaults.

Created attachment 495 Patch against svn 246. I think I've finally nailed it. Now works for member functions. I was inserting the symbols into the wrong scope (they need to be added into the function scope, not the struct scope).

SVN commit: http://www.dsource.org/projects/dmd/changeset/263

Fixed dmd 2.037