Bug 5373 – Regression (2.051) CTFE and std.string.replace() causes "Bad binary function q{a == b}..

Using replace() in a mixin doesn't seem to work if the mixin is placed inside a class. (I have no idea what's going on here.) Isolated example: // - - - - 8< - - - - - - - - - - - - import std.stdio, std.string; version(bug) { class Foo { mixin(anint("a")); this() { a = 9; } } } void main() { version(bug) { auto a = new Foo; writeln("a:", a.a); } else //works { mixin(anint("a")); a = 9; writeln("a:",a); } } string anint(string name) { return replace("int $name;", "$name", name); } // - - - - 8< - - - - - - - - - - - - # dmd -run testmixin.d a:9 # dmd -version=bug -run testmixin.d /Library/Compilers/dmd2/osx/bin/../../src/phobos/std/functional.d(177): Error: static assert "Bad binary function q{a == b}. You need to use a valid D expression using symbols a of type dchar and b of type const(char)[]." /Library/Compilers/dmd2/osx/bin/../../src/phobos/std/functional.d(180): instantiated from here: Body!(dchar,const(char)[]) /Library/Compilers/dmd2/osx/bin/../../src/phobos/std/algorithm.d(2149): instantiated from here: result!(dchar,const(char)[]) # _ // - - - - 8< - - - - - - - - - - - - I'm pretty sure this worked in 2.050, or at least 2.049. (Some program I'm working on started emitting these errors today. Didn't before. Not entirely sure when I last built it. Sorry.) BR /HF

Very interesting that you got code like that as well. I also use replace in a mixin inside a class in cl4d. This error was introduced in 2.051 I think.

Ok I tested different revisions of dmd with phobos etc. of 2.050 Seems like r809 introduced it. Can anybody confirm?

Yeah it was introduced with 2.051. The release date in the changelog is wrong. 2.051 probably contains dmd r810 and the bug was introduced in r809.

*** Issue 5406 has been marked as a duplicate of this issue. ***

This is the patch from Issue 5406. It appear to solve both issue's test cases. template binaryFunImpl(alias fun, string parm1Name, string parm2Name) { static if (is(typeof(fun) : string)) { enum testAsExpression = "{ ElementType1 " ~parm1Name~"; ElementType2 " ~parm2Name~"; return ("~fun~");}()"; typeof(mixin(testAsExpression)) result(ElementType1, ElementType2)(ElementType1 __a, ElementType2 __b) if(__traits(compiles, mixin(testAsExpression))) { mixin("alias __a "~parm1Name~";"); mixin("alias __b "~parm2Name~";"); mixin("return (" ~ fun ~ ");"); } } else { alias fun result; } }

Well the other bug report was purely about phobos. The question is if this is just a workaround for an issue that actually needs to be addressed in dmd.

(In reply to comment #6) > Well the other bug report was purely about phobos. > The question is if this is just a workaround for an issue that actually needs > to be addressed in dmd. _Not_ a Phobos bug? *sigh* Here's the bug in all it's detail: //Reduction 1 import std.string; pragma(msg, replace(int.stringof,"int","real")); //Reduction 2 pragma(msg, indexOf("int","real") ); //Reduction 3 pragma(msg, is(typeof(startsWith!"a == b"("int","real"))) ); //Reduction 4 bool startsWith2(alias pred = "a == b", R, E) (R doesThisStart, E withThis) if (is(typeof(binaryFun!pred(doesThisStart.front, withThis)))) { return true; } pragma(msg, is(typeof(startsWith2!"a == b"("int","real"))) ); //Final Reduction pragma(msg, is(typeof(binaryFun!"a == b"("int".front, "real"))) ); So what's happening here is that DMD is trying to execute replace at compile-time. Replace calls indexOf, which calls startsWith. Now startsWith has three overload sets which differ by their template constraints. So DMD evaluates each constraint in turn. Which then causes in invalid binaryFun call, which triggers a static assert. And static assert stops compilation then and there. As this is per spec (AssignExpression is evaluated at compile time, and converted to a boolean value. If the value is true, the static assert is ignored. If the value is false, an error diagnostic is issued and the compile fails. Unlike AssertExpressions, StaticAsserts are always checked and evaluted by the compiler unless they appear in an unsatisfied conditional.) This isn't a regression in DMD; it's a regression in Phobos due to better conformance of DMD to the spec. Now, there does appear to be an inconsistency between static assert's behavior during CTFE vs Non-CTFE, but that's more of an issue with the Non-CTFE behavior vs the CTFE behavior.

Well, as of 2.052 replace() has moved to std.array and can't be called at compile-time anymore at all cause of Appender using pointers. I've opened a new report: http://d.puremagic.com/issues/show_bug.cgi?id=5632

I think we can close this one now.

replace is broken again in 2.054. I get the "Bad binary function q{a == b}" with my code but this example doesn't even get so far: Error: argument to mixin must be a string, not (['i','n','t',' ','a',';'])

(In reply to comment #10) > replace is broken again in 2.054. > I get the "Bad binary function q{a == b}" with my code but this example doesn't > even get so far: > Error: argument to mixin must be a string, not (['i','n','t',' ','a',';']) That's bug 2156, which is completely different to this one. It has nothing to do with CTFE.

https://github.com/D-Programming-Language/dmd/commit/842db08516f0606ea9b278d179f8f7c51776f805 https://github.com/D-Programming-Language/dmd/commit/ff1965347d6ff2b0c3ebb71ff7c06254aba38c66