The previous version could easily generate a large amount of padding
when the constant factor was significantly larger than the width of the
shift data input. This could lead to huge amounts of logic being
generated before then being optimized away at a huge performance and
memory cost.
Additionally and more critically, when the input width was not a
multiple of the constant factor, the input data was padded with 'x bits
to such a multiple before interspersing the 'x padding needed to align
the selectable windows to power-of-two offsets.
Such a final padding would not be correct for shifts besides $shiftx,
and the previous version did attempt to remove that final padding at the
end so that the native zero/sign/x-extension behavior of the shift cell
would be used, but since the last selectable window also got
power-of-two padding appended after the padding the code is trying to
remove got added, it did not actually fully remove it in some cases.
I changed the code to only add 'x padding between selectable windows,
leaving the last selectable window unpadded. This omits the need to add
final padding to a multiple of the constant factor in the first place.
In turn, that means the only 'x bits added are actually impossible to
select. As a side effect no padding is added when the constant factor is
equal to or larger than the width of the shift data input, also solving
the reported performance bug.
This fixes#4056
Remove duplicate %.pmg -> %_pm.h pattern. One of the duplicates overrode
the other, and in some conditions there were build races as to whether
the target directory for the generated header would exist. Instead have
a single rule which is properly generalized.
Generalize what was formerly the unsigned-only architecture to support
both signed and unsigned multiplication, use that as default, and set
aside the special low-power architecture that was formerly used for
signed multipliers.
- moved all selection and filtering logic to the match block
- applied less-verbose code suggestions
- removed constraint on number of bits in shift-amount
- added check for possible wrap-arround in the operation
Add a separate shiftmul pattern to match on left shifts which implement
demuxing. This mirrors the right shift pattern matcher but is probably
best kept separate instead of merging the two into a single matcher.
In any case the diff of the two matchers should be easily readable.
The `opt_expr` pass running before `peepopt` can interfere with the
detection of a shiftmul pattern due to some of the bottom bits of the
shift amount being replaced with constant zero. Extend the detection to
cover those situations as well.
`memory_nordff` has the advantage that it can be called just ahead of
the simulation step no matter whether the clocked read port has been
inferred or was explicitly instantiated in a flow.
In commit fedd12261 ("booth: Move away from explicit `Wire` pointers")
a bug was introduced when checking for vacant slots in arrays holding
some intermediate results. Non-wire SigBit values were taken to imply
a vacant slot, but actually a constant one can make its way into those
results, if the multiplier cell configuration is just right. Fix the
vacancy check to address the bug.
The main speedup comes from swithing from using a SHA1 hash to std::hash<std::string>. There is no need to use an expensive cryptographic hash for fingerprinting in this context.
Jannis Harder
Martin Povišer
N. Engelhardt
phsauter
Lofty
Claire Xen
Marcus Comstedt
Rasmus Munk Larsen