Bug 3984 – Segfault(interpret.c): CTFE using struct constructor on a local static variable

In the following code snipped, for global variables or enums, this.s is not correctly written in the constructor. The direct way to assign a value without the 'dst' temporary does not work due to http://d.puremagic.com/issues/show_bug.cgi?id=3801. --- struct S { float[1] s; this(float x){ float[] dst = this.s; dst[0] = x; } } S S_zero = S(0); // initialization fails enum S S_zero2 = S(0); // initialization fails void main() { //enum S z = S(0); // ICE //static S z = S(0); // ICE //S z = S(0); // works, initialized with 0 //S z = S_zero; // contains NaN S z = S_zero2; // contains NaN assert(z.s[0] == 0); } ---

Most of the issues described here are duplicates of bug 1330, or bug 3801. The segfault is the only new bug here. Here's a reduced test case: struct Segfault3984 { int a; this(int x){ a = x; } } void bug3984(){ static assert(Segfault3984(3).a == 3); } Root cause: Struct constructors can result in a situation where a CTFE variable is created outside of a CTFE function. They need a context to put the variable into. Effectively, the comma expression acts a very simple function. PATCH: Interpret.c line 2641 Expression *CommaExp::interpret(InterState *istate) { #if LOG printf("CommaExp::interpret() %s\n", toChars()); #endif // If the comma returns a temporary variable, it needs to be an lvalue // (this is particularly important for struct constructors) if (e1->op == TOKdeclaration && e2->op == TOKvar && ((DeclarationExp *)e1)->declaration == ((VarExp*)e2)->var) { + // If there's no context for the variable to be created in, + // we need to create one now. + InterState istateComma; + if (!istate) + istate = &istateComma; VarExp* ve = (VarExp *)e2; VarDeclaration *v = ve->var->isVarDeclaration(); if (!v->init && !v->value) v->value = v->type->defaultInitLiteral();

changeset 423

Fixed DMD1.058 and 2.043.