[Bug target/95566] New: x86 instruction selection --- some REX prefixes unnecessary

["zero at smallinteger dot com" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95566

            Bug ID: 95566
           Summary: x86 instruction selection --- some REX prefixes
                    unnecessary
           Product: gcc
           Version: 10.1.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: zero at smallinteger dot com
  Target Milestone: ---

Created attachment 48696
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=48696&action=edit
sample code

Consider the code attached, compiled with

gcc -O3 sample.c -o sample

Gcc produces unrolled loop code that follows the pattern below.

        movzx   ecx, WORD PTR [rsp-62]
        cmp     rdx, rcx

Here, rdx has the value of k >> 48.  The top 32 bits of rdx are zero after the
shift, so the entirety of k >> 48 is in edx.  Thus, the cmp instructions could
be

        cmp     edx, ecx

instead.  This difference avoids the REX prefix, and thus the instructions are
shorter.  After sufficient unrolling (or with e.g. more complex comparisons
that depend on k >> 48), shorter instructions without the REX prefix will be
better even accounting for the partial register dependency (or an instruction
to break the dependency).  The Intel optimization manual says shorter
instructions are better.

The attachment is the entirety of sample.c.  I did not include other files
because this attachment appears to qualify for that exemption due to excuse
(ii): the attached test case is small and does not include any other file.

I originally found this behavior looking at the disassembly of gcc (Gentoo
9.2.0-r2 p3) 9.2.0.  I verified the same behavior with gcc 10.1 and gcc trunk
at godbolt.