[Bug rtl-optimization/94863] New: Failure to use blendps over mov when possible

[Bug rtl-optimization/94863] New: Failure to use blendps over mov when possible

[gabravier at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94863

            Bug ID: 94863
           Summary: Failure to use blendps over mov when possible
           Product: gcc
           Version: 10.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: rtl-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: gabravier at gmail dot com
  Target Milestone: ---

typedef double v2df __attribute__((vector_size(16)));

v2df move_sd(v2df a, v2df b)
{
    v2df result = a;
    result[0] = b[0];
    return result;
}

LLVM -O3 compiles this as such :

move_sd(double __vector(2), double __vector(2)): # @move_sd(double __vector(2),
double __vector(2))
  blendps xmm0, xmm1, 3 # xmm0 = xmm1[0,1],xmm0[2,3]
  ret

GCC gives this : 

move_sd(double __vector(2), double __vector(2)):
  movsd xmm0, xmm1
  ret

Using `blendps` here should be a worthy tradeoff. Here is a table of
throughputs for various CPU architectures formatted as "arch-name:
blendps-throughput, movsd-throughput" :

Wolfdale: 1, 0.33
Nehalem: 1, 1
Westmere: 1, 1
Sandy Bridge: 0.5, 1
Ivy Bridge: 0.5, 1
Haswell: 0.33, 1
Broadwell: 0.33, 1
Skylake: 0.33, 1
Skylake-X: 0.33, 1
Kaby Lake: 0.33, 1
Coffee Lake: 0.33, 1
Cannon Lake: 0.33, 0.33
Ice Lake: 0.33, 0.33
Zen+: 0.5, 0.25
Zen 2: 0.33, 0.25

Unless there is an important factor other than thoughput that could affect
this, this should improve performance or keep it identical on every
architecture except Zen+

[Bug rtl-optimization/94863] Failure to use blendps over mov when possible

[gabravier at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94863

--- Comment #1 from Gabriel Ravier <gabravier at gmail dot com> ---
Note: The given outputs for LLVM and GCC are when compiling with `-O3 -msse4.1`

[Bug target/94863] Failure to use blendps over mov when possible

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94863

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
throughputs put aside - how's port allocation and latency figures?  That said,
GCC usually sides on the smaller insn encoding variant when latency isn't
different - we're usually not looking at throughput since throughput figures
are quite useless if you look at single insns (and our decision is per
individual instruction).  Somehow modeling alternatives during scheduling
might make more sense but that's not implemented.

[Bug target/94863] Failure to use blendps over mov when possible

[gabravier at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94863

--- Comment #3 from Gabriel Ravier <gabravier at gmail dot com> ---
For binary size, the `movsd` takes 4 bytes and the `blendps` takes 6 bytes

The port allocations for the instructions are as such (same formatting as for
the throughputs) : 

Wolfdale: p5, p015
Nehalem: p5, p5
Westmere: p5, p5
Sandy Bridge: p05, p5
Ivy Bridge: p05, p5
Haswell: p015, p5
Broadwell: p015, p5
Skylake: p015, p5
Skylake-X: p015, p5
Kaby Lake: p015, p5
Coffee Lake: p015, p5
Cannon Lake: p015, p015
Ice Lake: p015, p015
Zen+: fp01, fp0123
Zen 2: fp013, fp0123

Something like "p015" meaning that the instruction can be executed on port 0, 1
or 5. Also, all architectures have both instructions take a single uop.

The latency of `blendps` and `movsd` are 1 on every single architecture I could
test

Final note : The numbers are specifically for the `blendps xmm, xmm, imm8` and
the `movsd xmm, xmm` forms of those instructions

[Bug target/94863] Failure to use blendps over mov when possible

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94863

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Severity|normal                      |enhancement

[Bug target/94863] Failure to use blendps over mov when possible

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94863

--- Comment #4 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Gabriel Ravier from comment #1)
> Note: The given outputs for LLVM and GCC are when compiling with `-O3
> -msse4.1`

I think you have the oppsite meaning with respect to `-msse4.1` here. They are
the same at -O3 but LLVM produces blendps with `-msse4.1` .