Bug 5079 – ICE(glue.c, 1103) on undefined symbol during CTFE

//import std.conv; // the forgotten import string nthLabel(int n) { return "__Global_"~ to!string(n); } template getGlobal(string file = __FILE__, int line = __LINE__, int N = 0) { static if( !__traits(compiles, mixin(nthLabel(N)) ) ) int getGlobal = N; else int getGlobal = getGlobal!(file,line, N+1 ); } enum Foo = getGlobal!(__FILE__, __LINE__) ; Assertion failure: '0' on line 1103 in file 'glue.c' when std.conv not imported Comment out enum Foo and the correct error message appears: template instance template 'to' is not defined, did you mean n?

This ICE/assertion error is also tripping up std.variant, specifically using DMD 2.049 the following: import std.variant; void main(string[] args) { Algebraic!(This[string]) ice; } Results in an [ICE] Assertion failure: '0' on line 1103 in file 'glue.c' Tracing this down, this is the offending piece of code. template maxSize(T...) { enum size_t maxSize = T[0].sizeof; } struct This; enum foobar = maxSize!(This[string]); In this case, the correct error is Error: struct This is forward referenced This has been blocking Phobos development, both in std.variant and with regard to improving std.json by using Algebraic, so I've upgraded the importance.

P.S. The ICE naturally also affects __traits(compiles, T[0].sizeof ), which make escaping the underlying bugs very difficult.

Having just debugged/patched/filed Issue 5155, I've found that A) that this test case is a different error and B) they both result from trying to calling toIR on a statement containing an expression with an __error type (current ty == 37). In the case of 5155, while debugging I found that detecting the erroneous ReturnStatement inside the CompoundStatement and eliding it produced the correct error message (though this might have been as much luck, as anything else). Looking at the source code, although if(exp) gets tested a fair amount, if(exp->type->ty == terror) never does, so I'm assuming that the code that checks for __error inside expressions lives elsewhere. Also, here is the stack track from DMD 2.050 for the bug in this report (Generated using cv2pdb and MSVC). dmd.exe!halt() Line 231 + 0x2 bytes C++ dmd.exe!Type::totym()() Line 1103 C++ dmd.exe!IntegerExp::toElem(IRState*)(IRState * irs) Line 1224 + 0x14 bytes C++ dmd.exe!eval_Darray(IRState * irs, Expression * e) Line 521 + 0xe bytes C++ dmd.exe!CatExp::toElem(IRState*)(IRState * irs) Line 2111 + 0xe bytes C++ dmd.exe!ReturnStatement::toIR(IRState*)(IRState * irs) Line 1292 + 0x11 bytes C++ dmd.exe!CompoundStatement::toIR(IRState*)(IRState * irs) Line 1344 C++ dmd.exe!FuncDeclaration::toObjFile(int )(int multiobj) Line 914 C++ dmd.exe!Module::genobjfile(int )(int multiobj) Line 280 C++ dmd.exe!main(int argc, char * * argv) Line 1295 C++ dmd.exe!_mainCRTStartup() + 0xa9 bytes C++

Another duplicate of the most important regression. *** This issue has been marked as a duplicate of issue 4269 ***