Re: [Tinycc-devel] [BUG] assertion failure when compiling initialization of array of structs on x86_64

[Michael Matz]

Hi,

On Wed, 25 May 2016, grischka wrote:

> Could it be that a compiler not necessarily needs to support this?

I don't see any leeway in the initializers rules of the standard. 
Compilers have to support this: it's a braced initializer list for a 
single element array, the initializer list (correctly) contains one 
element, that initializer itself is type compatible (in fact the same 
type) with the type of the initialized subobject (the struct).

> http://repo.or.cz/tinycc.git/commitdiff/1ca685f887310b5cbdc415cdfc3a578dbc8d82d8

Hmm, I don't think this is complete.  It may avoid the assert but probably 
generates wrong code (try e.g. with the struct containing multiple 
elements).  For instance it doesn't assert on i386-tcc, but generates a 
too short memmove when the struct is larger (for the example it works by 
accident).

I've traced this until I found that the dbt in gen_assign_cast is actually 
int or llong, which is clearly bogus, the destination type is a struct 
itself but code higher up the chain already looks into the structs first 
member and tries to initialize that (though the initializer on the right 
hand side doesn't init the individual members but the struct).

So, for this code:

struct S { int i;}
void f(struct S *s) { struct S y[1] = { *x }; }
void g(struct S *s) { struct S y[1] = { { x->i } }; }

TCC needs to differ between both cases and when parsing the LHS needs to 
dive into the type struct S for function g() but not for function f(), 
i.e. it needs to do that based on the form of the RHS.

At that point I got distracted then :)


Ciao,
Michael.