When an opcode generates a known high bit state (typically, a sub-word
operation that zeros the high bits), we can remove any subsequent
extension operation that would be a no-op.
x = (OP ...)
y = (ZeroExt32to64 x)
If OP zeros the high 32 bits, then we can replace y with x, as the
zero extension doesn't do anything.
However, x in this situation normally has a sub-word-sized type. The
semantics of values in registers is typically that the high bits
beyond the value's type size are junk. So although the opcode
generating x *currently* zeros the high bits, after x is rewritten to
another opcode it may not - rewrites of sub-word-sized values can
trash the high bits.
To fix, move the extension-removing rules to late lower. That ensures
that their arguments won't be rewritten to change their high bits.
I am also worried about spilling and restoring. Spilling and restoring
doesn't preserve the high bits, but instead sets them to a known value
(often 0, but in some cases it could be sign-extended). I am unable
to come up with a case that would cause a problem here, so leaving for
another time.
Fixes#66066
Change-Id: I3b5c091b3b3278ccbb7f11beda8b56f4b6d3fde7
Reviewed-on: https://go-review.googlesource.com/c/go/+/568616
Reviewed-by: Keith Randall <khr@google.com>
Reviewed-by: Cherry Mui <cherryyz@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
