f471c827fb
The new unit test demonstrates load/store reordering is error-prone.
At head we're allowing loads from a given pointer to reorder later than
a store to that same pointer, and boy, that's just not sound. In the
scenario constructed by the test we reorder this swap,
x = load32 X
y = load32 Y
store32 X y
store32 Y x
using schedule() (following Op argument data dependencies) into
y = load32 Y
store32 X y
x = load32 X
store32 Y x
which moves `x = load32 X` illegally past `store X y`.
We write `y` twice instead of swapping `x` and `y`.
It's not impossible to implement that extra reordering constraint: I
think it's easiest to think about by adding implicit use edges in
schedule() from stores to prior loads of the same pointer. But that'd
be a little complicated to implement, and doesn't handle aliasing at
all, so I decided to ponder on other approaches that handle a wider
range of programs or would have a simpler implementation to reason
about. I ended up walking through this rough chain of ideas:
0) reorder using only Op argument data dependencies (HEAD)
1) don't let load(ptr) pass store(ptr) (above)
2) don't let any load pass any store (allows aliasing)
3) don't reorder any Op that touches memory
4) don't reorder any Op, period.
This CL is 4). It's certainly the easiest and cheapest implementation.
It's not clear to me that we need this scheduling, and should we find we
really want it I'll come back and work back through the list until we
find something that meets our needs.
(Hoisting of uniforms is unaffected here.)
Change-Id: I7765b1d16202e0645b11295f7e30c5e09f2b7339
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/350256
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
|
||
|---|---|---|
| .. | ||
| golden | ||
| SampleWithConstantMatrix.rte | ||
| SampleWithExplicitCoord.rte | ||
| SampleWithUniformMatrix.rte | ||
| SampleWithVariableMatrix.rte | ||