On a loop back edge both the cached instance and the cached memory
start have to get restored for the next loop iteration. In the original
CL we did not consider the case that by restoring the instance we may
overwrite the currently cached memory start.

Original description:

WebAssembly functions often have subsequent memory accesses, and each of
these memory accesses need the start address of the memory in a register.
With this CL the register with the memory start address is cached, so
only the first memory access has to load the memory start address into a
register, subsequent memory accesses can just reuse the register.

In first measurements with the epic benchmark this reduces the size of
the generated Liftoff code by a bit more than 5%.

R=clemensb@chromium.org

Bug: v8:11862
Change-Id: I884c0da24be8bc6b10f2c6bf5437b9a279819538
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2960220Reviewed-by: Clemens Backes <clemensb@chromium.org>
Commit-Queue: Andreas Haas <ahaas@chromium.org>
Cr-Commit-Position: refs/heads/master@{#75183}