FMA regression on windows on master (CPU:9600KF)

[N5N3]

The detection seems incorrect? (I think 9600KF has native FMA support?)
On 1.7.0

julia> versioninfo()
Julia Version 1.7.0
Commit 3bf9d17731 (2021-11-30 12:12 UTC)
Platform Info:
  OS: Windows (x86_64-w64-mingw32)
  CPU: Intel(R) Core(TM) i5-9600KF CPU @ 3.70GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-12.0.1 (ORCJIT, skylake)
Environment:
  JULIA_CUDA_VERSION = 11.4.0

julia> @code_native fma(1.0,1.0,1.0)
        .text
; ┌ @ floatfuncs.jl:347 within `fma`
        pushq   %rbp
        movq    %rsp, %rbp
; │┌ @ floatfuncs.jl:336 within `fma_llvm`
        vfmadd213sd     %xmm2, %xmm1, %xmm0     # xmm0 = (xmm1 * xmm0) + xmm2
; │└
        popq    %rbp
        retq
        nopl    (%rax,%rax)

on master

julia> versioninfo()
Julia Version 1.8.0-DEV.1090
Commit d16f4806e9 (2021-12-01 10:54 UTC)
Platform Info:
  OS: Windows (x86_64-w64-mingw32)
  CPU: Intel(R) Core(TM) i5-9600KF CPU @ 3.70GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-12.0.1 (ORCJIT, skylake)
Environment:
  JULIA_CUDA_VERSION = 11.4.0

julia> @code_native fma(1.0,1.0,1.0)
        .text
        .file   "fma"
        .globl  julia_fma_552                   # -- Begin function julia_fma_552
        .p2align        4, 0x90
        .type   julia_fma_552,@function
julia_fma_552:                          # @julia_fma_552
; ┌ @ floatfuncs.jl:416 within `fma`
        .cfi_startproc
# %bb.0:                                # %L5
        pushq   %rbp
        .cfi_def_cfa_offset 16
        .cfi_offset %rbp, -16
        movq    %rsp, %rbp
        .cfi_def_cfa_register %rbp
        subq    $32, %rsp
        movabsq $j_fma_emulated_554, %rax
        callq   *%rax
        addq    $32, %rsp
        popq    %rbp
        retq
.Lfunc_end0:
        .size   julia_fma_552, .Lfunc_end0-julia_fma_552
        .cfi_endproc
; └
                                        # -- End function
        .section        ".note.GNU-stack","",@progbits

[DilumAluthge]

Duplicate of #43088, maybe?

cc: @oscardssmith

[oscardssmith]

Your CPU has FMA, but your operating system has a broken version of it that causes problems.

[oscardssmith]

We could consider letting windows users keep their broken version given the performance difference and the fact that it's not our fault that their OS is broken.

[N5N3]

So the problem is that fma on windows isnot hardware based even if the cpu has that instruction?

[oscardssmith]

The problem is that LLVM will constant fold using the software version sometimes since it assumes that the libm version (provided by your OS) is correct. This leads to really weird bugs like the same fma call in a function producing different results depending on whether LLVM is able to constant fold based on context.

[KristofferC]

To me, completely disabling it on Windows feels way too heavy handed.

[yuyichao]

It should at least be safe to enable it if the compiling machine also has fma.

[oscardssmith]

I haven't tried it with windows built from source. Can someone with access to windows do so?

[N5N3]

I built master with 543ff67 using Cygwin64.
Unfortunately, test code in #43088 is still broken.

[KristofferC]

Completely turning off fma on Windows is not an acceptable solution.

[vtjnash]

There is no functional or safe option until mingw64 is fixed