Issue 241349
Summary [ValueTracking] Make `computeKnownFPClass` handle fast math flags correctly afn arcp contract assoc nsz
Labels floating-point
Assignees ZERICO2005
Reporter ZERICO2005
    `computeKnownFPClass` currently ignores `afn`, `arcp`, `contract`, `assoc`. However, some of these are "rewrite" flags, which means that the analysis would be invalidated if the _expression_ is rewritten. Maybe the analysis says that an _expression_ will never produce zero, but then a rewrite makes it possible for the _expression_ to produce zero.

`!fpmath` is also not handled currently.

We also don't currently handle `nsz`, but `nsz` is more tame in that it widens +0.0 or -0.0 to +-0.0, and can be modeled similarly to input denormal flushing.

`nnan` and `ninf` are perfectly safe to ignore as they narrow the possible inputs/outputs instead of widening them.

iirc Alive2 only models `nnan`, `ninf`, `nsz`, and maybe a few others in limited circumstances. Ideally we should make `computeKnownFPClass` handle fast math flags the same way that Alive2 does.

One rewritable _expression_ we do perform analysis on is `fmuladd`. It is not too difficult to handle because we can just compute the union of the two possibilities: `fma(x, y, z)` and `fadd(fmul(x, y), z)`. 

***

Overall, handling the rewrite flags would probably look like taking the union of all possible transformations. 

Some options to handle each fast math flag:
- `nsz`:
  - Model it similarly to input denormal flushing
- `arcp`:
  - Bail out
  - Maybe walk the _expression_ tree to see what transformations may be possible under `arcp`. This is only feasible if the number of possible transformations won't grow or change over time.
- `contract`:
  - Bail out
  - Walk the _expression_ tree for possible transformations.
- `assoc`:
  - Bail out
  - Walk the _expression_ tree for possible transformations.
- `afn`:
  - Bailing out is probably the only feasible solution.
  - Does `afn` guarantee that `fmul(pnorm, pnorm) = non-negative`?
- `!fpmath`:
  - A naive solution would add +-ulp to the set of possible results. A result that is `fcPosSubnormal` would expand to `fcZero | fcSubnormal | fcNormal` under 2ulp.
  - Does 1ulp on `fcPosInf` make `FLT_MAX` a possible result? I feel like it could depend on if the infinity is an exact infinity or an overflow.
  - Does 1ulp on `fcPosNormal` widen it to `fcZero` if output denormals flush?
  - How many ulp is needed to turn `fcPosNormal` into `fcNegNormal`?
  - If the ulp is 2^22, should we trust it, or bail out if it is some large value greater than say 1000 ulp.
 - Maybe a temporary solution would be to bail out when the ulp>5 so fdiv/sqrt analysis results are still performed for GPU targets.

Propagation of non-sNaN is probably the only information that is safe to propagate under all possible transformations. Then again, if you're using fast math flags you probably also have `nnan` anyways.

I wonder if it would be a good idea to distinguish between `ulp`, relative error, and absolute error for `!fpmath`.
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs

Reply via email to