struct S{ byte[10] m_a; //byte[9] does not ice } S var; const(S) getVar() //returning just S does not crash { return var; } void main() { const(S) foo = getVar(); } crashes dmd with: Internal error: ..\ztc\cod4.c 357
This is probably related to or a duplicate of #2560.
*** Issue 3203 has been marked as a duplicate of this issue. ***
(In reply to comment #0) here is (probably) another case of this problem. struct A {} struct B { const A a; } B f() { return B(A()); } Internal error: ..\ztc\cod4.c 354 @ DMD2.031 "const(struct)" is the common component for the issues related to ICE(cod4.c) 35#. Here is a list of problematic code: In the function "cdeq" which generates code for an assignment sz = tysize[tyml]; assert((int)sz > 0); <- failed this assertion. where tyml is a type of lvalue, sz represents # of bytes to transfer. This issue is apparently due to bypass of type-size{tysize} registration for const(struct), besides registration for struct itself is done.
(In reply to comment #3) > "const(struct)" is the common component for the issues related to ICE(cod4.c) > 35#. > Here is a list of problematic code: > In the function "cdeq" which generates code for an assignment > > sz = tysize[tyml]; > assert((int)sz > 0); <- failed this assertion. > > where tyml is a type of lvalue, sz represents # of bytes to transfer. > > This issue is apparently due to bypass of type-size{tysize} registration for > const(struct), besides registration for struct itself is done. A variation of that test case is interesting: struct A {} struct B { const A a; } void f() { //A a; // ---- if you uncomment this, it doesn't ICE! B b = B(A()); } Comparing the intermediate code using elem_print(e); shows that there's a difference between the two cases by the start of codelem(). But there's no difference in the code generated by DeclarationExp::toElem() in e2ir.c. Somewhere between the two, an optimisation/rewriting step is performed in the correct case, but not in the ICE case. I haven't yet worked out where it happens.
Like 2560, this is a regression between 2.022 and 2.023. It's clearly the same bug.
Fixed dmd 2.032