These compute the average of two given values, rounding up and down,
respectively, if the result isn't exact. When possible, these are
mapped to target-specific intrinsics (PADD[BW] on IA and VH[R]ADD[US]
on NEON.)
A subsequent commit will add pattern-matching to generate calls to
these intrinsincs when the corresponding patterns are detected in the
IR.)
On a target with a 16-bit mask (for example), we would choose the type
of an integer literal "1024" to be an int16. Previously, we used an int32,
which is a worse fit and leads to less efficient code than an int16
on a 16-bit mask target. (However, we'd still give an integer literal
1000000 the type int32, even in a 16-bit target.)
Updated the tests to still pass with 8 and 16-bit targets, given this
change.
There were a number of places throughout the system that assumed that the
execution mask would only have either 32-bit or 1-bit elements. This
commit makes it possible to have a target with an 8- or 16-bit mask.
For varying int8/16/32 types, divides by small constants can be
implemented efficiently through multiplies and shifts with integer
types of twice the bit-width; this commit adds this optimization.
(Implementation is based on Halide.)
The stdilb code just calls the signed int{32,64} functions,
which gives the right result for the unsigned case anyway.
The various targets didn't consistently define the unsigned
variants in any case.
In particular, this gives us desired behavior for NaNs (all compares
involving a NaN evaluate to true). This in turn allows writing the
canonical isnan() function as "v != v".
Added isnan() to the standard library as well.
Previously, we were trying to take a uniform seed and then shuffle that
around to initialize the state for each of the program instances. This
was becoming increasingly untenable and brittle.
Now a varying seed is expected and used.
This was unnecessary overhead to impose on all callers; the user
should handle these as needed on their own.
Also added some explanatory text to the documentation that highlights
that memory_barrier() is only needed across HW threads/cores, not
across program instances in a gang.
Rewrite things to be able to do a float MINPS, for slightly
better code on SSE2 (which has that but not an signed int
min). SSE2 code now 23 instructions (vs 21 intrinsics).
