Re: [RFA] New pass for sign/zero extension elimination

[Joern Rennecke <joern.rennecke@embecosm.com>]

On Tue, 28 Nov 2023 at 13:36, Joern Rennecke
<joern.rennecke@embecosm.com> wrote:
 > For the saturating truncation operations, we have the high-to-low
propagation,
> but no low-to-high propagation, so that would be something separate to model.

P.S.:
For unsigned saturating truncation, the propagation from higher to
lower bits only
happens for bits that are truncated off.
e.g. if we truncate a 64 bit value to a 32 bit value, and only the
lower 16 bit of the
result are live, we got an output live mask
0x000000000000ffff     implying an input live mask:
0xffffffff0000ffff

For signed saturating truncation, we got an extra corner case.  For
the same data widths
as above, the value
0xffffffff80000000
truncates to:
0x80000000
but
0x0000000080000000
truncates to:
0x7fffffff

so the top bit that is included in the truncated mode propagates to
all the lower bits
(irrespective if it itself is live in the output), so it is live in
the input if any bit is live in
the output - just like all the truncated-off bits.