Bug 1861 – (D1 only) .sort fails if opCmp takes a ref param

------ module opcmpref; import std.stdio; struct Pair { int a, b; // Sort by b first then a // Sorting fails if rhs is a 'ref' param. Ok if ref removed. int opCmp(ref Pair rhs) { if (b!=rhs.b) return b-rhs.b; return a-rhs.a; } string toString() { return std.string.format("(%s,%s)",a,b); } } void main() { // Single comparisons are all fine assert( !(Pair(0,0)<Pair(0,0)) ); assert( !(Pair(1,0)<Pair(0,0)) ); assert( !(Pair(0,1)<Pair(0,0)) ); assert( !(Pair(1,1)<Pair(0,0)) ); assert( (Pair(0,0)<Pair(1,0)) ); assert( !(Pair(1,0)<Pair(1,0)) ); assert( !(Pair(0,1)<Pair(1,0)) ); assert( !(Pair(1,1)<Pair(1,0)) ); assert( (Pair(0,0)<Pair(0,1)) ); assert( (Pair(1,0)<Pair(0,1)) ); assert( !(Pair(0,1)<Pair(0,1)) ); assert( !(Pair(1,1)<Pair(0,1)) ); assert( (Pair(0,0)<Pair(1,1)) ); assert( (Pair(1,0)<Pair(1,1)) ); assert( (Pair(0,1)<Pair(1,1)) ); assert( !(Pair(1,1)<Pair(1,1)) ); // But sorting fails when opCmp takes 'ref' param auto p = [Pair(0,0), Pair(3,0), Pair(2,1), Pair(1,4)]; p.sort; writefln("p=",p); assert(p[0] == Pair(0,0)); assert(p[1] == Pair(3,0)); assert(p[2] == Pair(2,1)); assert(p[3] == Pair(1,4)); } ------------ This may be something in the low-level lib rather than the compiler per-se. I'm not sure about that. I have tested and found that the error happens with both Phobos and Tango.

According to http://www.digitalmars.com/d/arrays.html, the parameter to a struct's opCmp() must be S or S*, not ref S.

Also, sort in std.algorithm should work with predicates that take references.

Ok, the documentation says either use Type or Type* for opCmp. The .sort property does in fact work with either. The problem with just using Type* is that it isn't called for regular struct comparisons like Type a,b; a<b The problem with using just Type is that it makes sorting unnecessarily slow, since value arguments will have to be copied for every call of the comparison function. It makes no sense not to allow ref. This is exactly the kind of thing ref was meant for, where you want value semantics, but only want to pass a pointer around. Interestingly, if you provide BOTH a Type* and a ref Type version of opCmp, then regular comparisons will use the ref version, and .sort will use the pointer version. But it would be nicer if ref just worked. I find it very surprising that it doesn't, and until it does I would suggest that the documentation be changed to say in very big bold letters that EVEN THOUGH REF MAY APPEAR TO WORK FOR SMALL TEST CASES IT IS NOT SUPPORTED AND WILL FAIL IF YOU TRY TO .sort.

Marking this as D1 only, as the built-in, AA.sort method is going to be deprecated. With a slight modification, your test example passes and prints the correct result with the std.algorithm.sort function (D2 only). ------ module opcmpref; import std.stdio, std.algorithm; struct Pair { int a, b; // Sort by b first then a // Sorting fails if rhs is a 'ref' param. Ok if ref removed. int opCmp(ref Pair rhs) { if (b!=rhs.b) return b-rhs.b; return a-rhs.a; } @property string toString() { return std.string.format("(%s,%s)",a,b); } } void main() { // Single comparisons are all fine assert( !(Pair(0,0)<Pair(0,0)) ); assert( !(Pair(1,0)<Pair(0,0)) ); assert( !(Pair(0,1)<Pair(0,0)) ); assert( !(Pair(1,1)<Pair(0,0)) ); assert( (Pair(0,0)<Pair(1,0)) ); assert( !(Pair(1,0)<Pair(1,0)) ); assert( !(Pair(0,1)<Pair(1,0)) ); assert( !(Pair(1,1)<Pair(1,0)) ); assert( (Pair(0,0)<Pair(0,1)) ); assert( (Pair(1,0)<Pair(0,1)) ); assert( !(Pair(0,1)<Pair(0,1)) ); assert( !(Pair(1,1)<Pair(0,1)) ); assert( (Pair(0,0)<Pair(1,1)) ); assert( (Pair(1,0)<Pair(1,1)) ); assert( (Pair(0,1)<Pair(1,1)) ); assert( !(Pair(1,1)<Pair(1,1)) ); auto p = [Pair(0,0), Pair(3,0), Pair(2,1), Pair(1,4)]; std.algorithm.sort(p); // prints 'p=[(0,0), (3,0), (2,1), (1,4)]' writefln("p=%s",p); assert(p[0] == Pair(0,0)); assert(p[1] == Pair(3,0)); assert(p[2] == Pair(2,1)); assert(p[3] == Pair(1,4)); } ------

D1 is not supported anymore, and in D2 built-in sort property will be deprecated.

D2 builtin sort is gone, closing.