Implement dynamically uniform expression restrictions

[litherum]

Migrated from https://bugs.webkit.org/show_bug.cgi?id=189125:

At 2018-08-29T23:36:49Z, [email protected] wrote:
Implement dynamically uniform expression restrictions

[litherum]

At 2018-09-13T01:50:51Z, [email protected] wrote:
rdar://problem/44403002

[litherum]

At 2018-09-24T17:37:06Z, [email protected] wrote:
Ever value (lvalue and rvalue) will be associated with a "uniformity" attribute, which will be assigned by the compiler. These attributes can even survive being assigned to a variable, if the variable is marked as "uniform" in the program source.

Then, after we know whether or not each value is uniform, we can associate each nested scope with a similar uniformity attribute. This is true iff the condition leading to that scope is uniform.

We apply the restriction that any AssignmentExpression inside a non-uniform scope must never assign to a uniform variable that is declared in the non-innermost scope. This should guarantee that all uniform variables are, indeed, uniform.

Then, we can require any IndexExpression that indexes into an array which contains textures to use a uniform index.

[litherum]

At 2018-09-24T18:14:24Z, [email protected] wrote:
Functions that return uniform variables must only have a single return point

[litherum]

At 2018-09-24T18:18:09Z, [email protected] wrote:
UAVs can't hold uniform values

[litherum]

At 2018-09-24T18:25:41Z, [email protected] wrote:
Conditional continues and breaks should contribute to the uniformity of the blocks they target

[litherum]

At 2018-09-24T18:28:16Z, [email protected] wrote:
(In reply to Myles C. Maxfield from comment #5)

Conditional continues and breaks should contribute to the uniformity of the
blocks they target

This isn't quite right - the continue & break can affect just the remainder of the scope.

Also, uniforms should only exist in thread / constant address space.

[litherum]

At 2018-09-24T18:37:48Z, [email protected] wrote:
(In reply to Myles C. Maxfield from comment #6)

(In reply to Myles C. Maxfield from comment #5)

Conditional continues and breaks should contribute to the uniformity of the
blocks they target

This isn't quite right - the continue & break can affect just the remainder
of the scope.

and "return"

In general, returns should be modeled as an assignment to a variable that's outside the function scope.

[litherum]

At 2018-09-24T18:41:52Z, [email protected] wrote:
(In reply to Myles C. Maxfield from comment #3)

Functions that return uniform variables must only have a single return point

Actually, we can do better - just treat early returns as marking the rest of the scope non-uniform, and treating the return as writing to a value that's outside the scope

[litherum]

At 2018-10-05T02:45:53Z, [email protected] wrote:
*** Bug 189123 has been marked as a duplicate of this bug. ***

[litherum]

At 2018-10-05T02:47:32Z, [email protected] wrote:
Created attachment 351650
WIP

[litherum]

At 2018-10-06T02:35:24Z, [email protected] wrote:
Created attachment 351711
WIP

[litherum]

At 2018-10-09T07:07:01Z, [email protected] wrote:
Created attachment 351866
WIP

[litherum]

At 2018-10-09T07:25:46Z, [email protected] wrote:
Created attachment 351868
WIP

[litherum]

#58