Lower inlining cost of floating point div #50428
Merged
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Our inlining cost model is extremely primitive, though surprisingly functional given its limitations. The basic idea for it was just that we'd give every intrinsic the approximate cost in cycles, such that for sufficiently large functions (>100 cycles), the cost of the extra call would be dwarfed by the cost of the function. However, there's a few problems with this. For one, the real issue is usually not the extra overhead of the call (which is small and well-predicated), but rather the inhibition of optimizations that inlining might have allowed. Additionally, the relevant cost comparison is not generally latency, but rather the size of the resulting binary. Lastly, the latency metric is misleading on modern superscalar architectures, because the core will perform other tasks while the operation is executing. In fact, somewhat counter-intuitively, this means that it is more important to inline high-latency instructions to allow the compiler to perform better latency hiding by spreading out the high-latency instructions.
We probably need a full-on rethink of the inlining model at some point, but for the time being, this fixes a problem that I ran into in real code by reducing the inlining cost for floating point division to be the same as that of floating point multiplication. The particular case where I saw this was the batched forward AD rule for division, which had 6 calls to div_float. Inlining these provided substantially better performance.