[llvm-bugs] [Bug 32550] New: GlobalISel -O0 for AArch64 moves floating point values through GPRs way too often

[via llvm-bugs]

http://bugs.llvm.org/show_bug.cgi?id=32550

            Bug ID: 32550
           Summary: GlobalISel -O0 for AArch64 moves floating point values
                    through GPRs way too often
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: All
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: GlobalISel
          Assignee: unassignedbugs at nondot.org
          Reporter: kristof.beyls at gmail.com
                CC: llvm-bugs at lists.llvm.org

Investigating the biggest performance and code size regressions I reported in
http://lists.llvm.org/pipermail/llvm-dev/2017-April/111713.html, it seems like
most boil down to floating point values being moved through general purpose
registers.

A very simple example is:
$ cat t3.c
double f(double r) {
  return r*r;
}
$ clang -target aarch64  -O0 -o - -S t3.c
...
f:                                      // @f
// BB#0:                                // %entry
        sub     sp, sp, #16             // =16
        str     d0, [sp, #8]
        ldr     d0, [sp, #8]
        ldr     d1, [sp, #8]
        fmul    d0, d0, d1
        add     sp, sp, #16             // =16
        ret
...
$ clang -target aarch64 -mllvm -global-isel -O0 -o - -S t3.c
...
f:                                      // @f
// BB#0:                                // %entry
        sub     sp, sp, #16             // =16
        fmov    x8, d0
        str     x8, [sp, #8]
        ldr     x8, [sp, #8]
        ldr     x9, [sp, #8]
        fmov    d0, x8
        fmov    d1, x9
        fmul    d0, d0, d1
        add     sp, sp, #16             // =16
        ret
...

It might be that this happens "just" every time a floating point variable is
loaded/stored to the stack, or it may happen more generally.
I think this needs to be fixed first to get rid of the source of the most
severe performance and code size regressions when enabling global-isel at -O0.
After this is fixed, another round of investigation can be done to find the
remaining biggest code size and performance regressions.

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20170406/70e9781f/attachment.html>