www.digitalmars.com         C & C++   DMDScript  

digitalmars.D - Function attribute deduction depends on compile invocation

reply Johan Engelen <j j.nl> writes:
The frontend is automatically deducing things like  nogc and 
nothrow, but it is flaky and results in linker errors for a 
complex codebase (Weka's).

What happens in that in one kind of compilation,  nogc is deduced 
for a function. But in another compilation, it isn't. Thus 
references to the function will have the wrong mangling, and 
linker errors happen:
https://issues.dlang.org/show_bug.cgi?id=17541

This is blocking upgrading to a newer compiler with a newer 
frontend (2.071 --> 2.073).
(I already have a large amount of work arounds... Also non-user 
functions have this problem, e.g. there is a nothrow deduction 
problem for __fieldPostblit.)

Thanks for looking into this bug.

- Johan
Jul 14
parent reply Walter Bright <newshound2 digitalmars.com> writes:
On 7/14/2017 9:53 AM, Johan Engelen wrote:
 What happens in that in one kind of compilation,  nogc is deduced for a 
 function. But in another compilation, it isn't. Thus references to the
function 
 will have the wrong mangling, and linker errors happen:
 https://issues.dlang.org/show_bug.cgi?id=17541
At the moment you can work around it by explicitly adding the annotation.
Jul 14
next sibling parent reply Johan Engelen <j j.nl> writes:
On Friday, 14 July 2017 at 23:51:24 UTC, Walter Bright wrote:
 On 7/14/2017 9:53 AM, Johan Engelen wrote:
 What happens in that in one kind of compilation,  nogc is 
 deduced for a function. But in another compilation, it isn't. 
 Thus references to the function will have the wrong mangling, 
 and linker errors happen:
 https://issues.dlang.org/show_bug.cgi?id=17541
At the moment you can work around it by explicitly adding the annotation.
Cannot add something to a compiler internal function (e.g. __fieldPostblit), nogc and pure are infectious, the templates are instantiated with different types where some instantiations require deducing nogc/pure/..., others require gc/impure/..., annotations cannot be applied conditionally (at least I don't know how), cannot add gc, cannot add impure, etc. My current workarounds force deduction one way or another, but it's definitely not nice and it is very time consuming to figure out what tricks the compiler into doing the right thing. Each error needs it's own special work around. For example, working around the __fieldPostblit deduction problem involved searching all member fields (and their members, etc.) to try to find which this(this) should be annotated to fix things. Not so nice to do for a somewhat complex struct in a 9MB D-code base, sprinkled with conditional compilation, templates, CTFE... Each trial build (incremental) takes approx a minute. Fun times. Fortunately, it looks like I've "fixed" all issues and we can now link; kicked of the testing, let's hope there are not too many regression bugs at runtime. -Johan
Jul 15
parent Nicholas Wilson <iamthewilsonator hotmail.com> writes:
On Saturday, 15 July 2017 at 12:29:09 UTC, Johan Engelen wrote:
 On Friday, 14 July 2017 at 23:51:24 UTC, Walter Bright wrote:
 On 7/14/2017 9:53 AM, Johan Engelen wrote:
 What happens in that in one kind of compilation,  nogc is 
 deduced for a function. But in another compilation, it isn't. 
 Thus references to the function will have the wrong mangling, 
 and linker errors happen:
 https://issues.dlang.org/show_bug.cgi?id=17541
At the moment you can work around it by explicitly adding the annotation.
Cannot add something to a compiler internal function (e.g. __fieldPostblit), nogc and pure are infectious, the templates are instantiated with different types where some instantiations require deducing nogc/pure/..., others require gc/impure/..., annotations cannot be applied conditionally (at least I don't know how), cannot add gc, cannot add impure, etc.
Not that it helps you at the moment, but thus is addressed by my attributes DIP. I suppose you could abuse allocSize to force gc, but thats very ugly.
Jul 15
prev sibling parent Walter Bright <newshound2 digitalmars.com> writes:
On 7/14/2017 4:51 PM, Walter Bright wrote:
 On 7/14/2017 9:53 AM, Johan Engelen wrote:
 What happens in that in one kind of compilation,  nogc is deduced for a 
 function. But in another compilation, it isn't. Thus references to the 
 function will have the wrong mangling, and linker errors happen:
 https://issues.dlang.org/show_bug.cgi?id=17541
At the moment you can work around it by explicitly adding the annotation.
https://github.com/dlang/dmd/pull/6995
Jul 15