Seen when compiling self-hosted D compiler. release/gcc-9 compiles: isBaseOf (struct TypeClass * const this, struct Type * t, int * poffset) { if (t != 0B && t->ty == 7) { { struct ClassDeclaration * cd; cd = ((struct TypeClass *) t)->sym; if (MEM[(bool (*<T1704>) (struct ClassDeclaration *, struct ClassDeclaration *, int *))this->sym->__vptr + 704B] (this->sym, cd, poffset)) { return <retval> = 1; } } } return <retval> = 0; } release/gcc-10 compiles: isBaseOf (struct TypeClass * const this, struct Type * t, int * poffset) { if (t != 0B && t->ty == 7) { { struct ClassDeclaration * cd; cd = ((struct TypeClass *) t)->sym; if (MEM[(bool (*<T171b>) (struct ClassDeclaration *, struct ClassDeclaration *, int *))this->sym->__vptr + 736B] (this->sym, cd, poffset)) { return <retval> = 1; } } } return <retval> = 0; } Applied all changes to gcc/d to the gcc-9 branch, and the problem gets resolved, so will have to comb through the diff to find out what missing change is causing gdc-9 to miscompile.
Looks like fix was in r10-7280, taken from the upstream backport in https://github.com/dlang/dmd/pull/10913
The releases/gcc-9 branch has been updated by Iain Buclaw <ibuclaw@gcc.gnu.org>: https://gcc.gnu.org/g:7d505b0ed8565b8c120ddd2b0b4630c93eecdec5 commit r9-8599-g7d505b0ed8565b8c120ddd2b0b4630c93eecdec5 Author: Iain Buclaw <ibuclaw@gdcproject.org> Date: Sat May 16 23:33:15 2020 +0200 d: Fix wrong vtable offset in virtual function call The Semantic (pass 1) analysis for classes is handled by ClassDeclaration::semantic. For a given class, this method may be ran multiple times in order to resolve forward references. The method incrementally tries to resolve the types referred to by the members of the class. The subsequent calls to this method are short-circuited if the class members have been fully analyzed. For this the code tests that it is not the first/main call to the method (semanticRun == PASS.init else branch), scx is not set, and that the this->symtab is already set. If all these conditions are met, the method returns. But before returning, the method was setting this->semanticRun to PASSsemanticdone. It should not set semanticRun since the class has not been fully analyzed yet. The base class analysis for this class could be pending and as a result vtable may not have been fully created. This fake setting of semanticRun results in the semantic analyzer to believe that the class has been fully analyzed. As exposed by the issues in upstream, it may result in compile time errors when a derived type class is getting analyzed and because of this fake semanticdone on the base class, the semantic analysis construes that an overriden method is not defined in the base class. PR95155 exposes anoter scenario where a buggy vtable may be created and a call to a class method may result in execution of some adhoc code. gcc/d/ChangeLog: PR d/95155 * dmd/dclass.c (ClassDeclaration::semantic): Don't prematurely set done on semantic analysis. gcc/testsuite/ChangeLog: PR d/95155 * gdc.test/compilable/imports/pr9471a.d: New test. * gdc.test/compilable/imports/pr9471b.d: New test. * gdc.test/compilable/imports/pr9471c.d: New test. * gdc.test/compilable/imports/pr9471d.d: New test. * gdc.test/compilable/pr9471.d: New test.
Fixed.