Bug 10750 – Strict aliasing semantics

From a discussion on dm.D (http://forum.dlang.org/post/[email protected]): ––– On Saturday, 27 July 2013 at 08:59:54 UTC, Walter Bright wrote: > On 7/27/2013 1:57 AM, David Nadlinger wrote: >> On Saturday, 27 July 2013 at 06:58:04 UTC, Walter Bright wrote: >>> Although it isn't in the spec, D should be "strict aliasing". >>> This is because: >>> >>> 1. it enables better code generation >>> >>> 2. there are ways, such as unions, to get the other aliasing >>> that doesn't >>> break strict aliasing >> >> We need to carefully formalize this then, and quickly. The >> problem GCC, Clang >> and others are facing is that (as you are probably aware) 2. >> isn't guaranteed to >> work for type-casting pointers either by the specs, but people >> want to be able >> to do this nonetheless. >> >> Thus, they both accept pointer aliasing through union types, >> trying to optimize >> as much as possible while avoiding to break people's >> expectations and existing >> code. This is a very unfortunate situation for both compiler >> developers and >> users; just search for something like "gcc strict aliasing" on >> StackOverflow for >> examples. >> >> There is already quite a lot of D code out there that violates >> the C-style >> strict aliasing rules. > > I agree. Want to do an enhancement request on bugzilla for it? –––

(In reply to comment #0) > (http://forum.dlang.org/post/[email protected]): My comments was: Is it good to add to Phobos a small template (named like "PointerCast" or something similar) that uses a union internally to perform pointer type conversions? Is then the compiler going to warn the programmer when the pointer type aliasing rule is violated? I mean when the D code uses cast() between different pointer types (beside constness). An alternative design is to even deprecate (and later turn those into errors, where the error message suggests to use PointerCast).

Such a PointerCast is not safe in all cases when compiling with GDC as even unions are not an exception to strict aliasing rules: http://stackoverflow.com/questions/2906365/gcc-strict-aliasing-and-casting-through-a-union I'm not sure if it's possible to change this in the GDC frontend.

@bearophile: To further expand on this: http://gcc.gnu.org/onlinedocs/gcc-4.8.2/gcc/Optimize-Options.html says: "type-punning is allowed, provided the memory is accessed through the union type. [...] access by taking the address, casting the resulting pointer and dereferencing the result has undefined behavior, even if the cast uses a union type, e.g.: " @David What would a safe cast with strict pointer aliasing look like? First some background information on how aliasing is implemented in gcc (alias.c): Every type is assigned an alias set. The alias set is only a unique id + a flattened list of the uids of all 'member types'. For example, this struct: ---------------------- struct B { char member; } struct A { int member1, float member2; B member3; } ---------------------- will generate this alias set: uid=1, children={2(int),3(float),4(char)} Then for code like this: ---------------------- A instance; instance.member1 = 0; A copy = a; ---------------------- The compiler now inspects the line instance.member1 = 0; and assigns alias set 2(int) to it. Line 3 has alias set 1(B). When gcc now schedules instructions it checks if set 2 conflicts with set 1 by checking: (set1 == set2 || set1 in set2.children || set2 in set1.children). If they don't conflict gcc reorders instructions. This explains the problems with type punning: ---------------------- int a = 3; //alias set 0(int), children = {} int b = a; //alias set 0(int), children = {} *(cast(float*)&a) = 3.0f; //alias set 1(float), children = {} ---------------------- as you can see these types don't conflict and gcc may reorder line 2 and 3. Access through unions now solves this problem as the alias set for a union would include both {float, int} as children. But as for as I understand these strict alising rules make it impossible to safely cast from one pointer type to another. Only _access_ through unions will work. As an example: ---------------------- T* safeCast(T, U)(U* input) { union wrap { U inp; T outp; } return &(cast(wrap*)input).outp; } void withFloat(float* f) { *f = 0.1f; } int b; void withInt(int* i) { b = *i; } void main() { int x = 0; auto asFloat = (safeCast!float(&x)); withFloat(asFloat) withInt(&i); } ---------------------- now with optimizations (inlining) ------------------------------------ union wrap { int inp; float outp; } int b; void main() { int x = 0; //alias set: int auto asFloat = (&(cast(wrap*)x).outp) //alias set: wrap (but noop) *asFloat = 0.1f; //alias set: float b = x; //alias set: int } ------------------------------------ I know from unfortunate experienc, that gcc may even completely discard the "auto asFloat" line. But even if it didn't, "*asFloat = 0.1f;" and "b = x;" can be reordered according to strict aliasing rules. If "auto asFloat" is discarded, even "int x = 0;" and "*asFloat = 0.1f;" may be reordered. So to summarize this: I don't know how you could make a safe cast from T* to U* assuming strict aliasing rules. Unions are only safe if all access goes through unions, but that is not possible when dealing with 3rd party functions. (Assume you can't change withFloat, withInt). We had problems with this in GDC right now on ARM (std.algorithm.find uses cast(ubyte[])string which internally translates to invalid pointer aliasing) and as a result we'll now have to disable strict aliasing in the GCC backend. I think type based aliasing, even if it may provide some optimization benefits, is in general a horrible idea.

(In reply to comment #3) > We had problems with this in GDC right now on ARM (std.algorithm.find uses > cast(ubyte[])string which internally translates to invalid pointer aliasing) > and as a result we'll now have to disable strict aliasing in the GCC backend. > Which is a shame, because dynamic arrays are perhaps the one type in D that should instead benefit from strict aliasing rules...

(In reply to comment #4) > (In reply to comment #3) > > We had problems with this in GDC right now on ARM (std.algorithm.find uses > > cast(ubyte[])string which internally translates to invalid pointer aliasing) > > and as a result we'll now have to disable strict aliasing in the GCC backend. > > > > Which is a shame, because dynamic arrays are perhaps the one type in D that > should instead benefit from strict aliasing rules... Alternatively, we can just define better aliasing rules that better suit D. ie: - Permit type-punning when accessing through a union. - Determine aliasing rules of dynamic arrays from the elem type, instead of treating it as aliasing the overall structure. This might actually be the better solution for us - shall I send you a patch? :o)

Ping, did anything ever happen with this? I cannot find anything in D's spec about aliasing/type punning and this seems to be the only relevant bug report.

THIS ISSUE HAS BEEN MOVED TO GITHUB https://github.com/dlang/dmd/issues/17602 DO NOT COMMENT HERE ANYMORE, NOBODY WILL SEE IT, THIS ISSUE HAS BEEN MOVED TO GITHUB