numbers test failure on 32bit

[yuyichao]

I got the following test failure on a 32bit machine.....

     * numbers             exception on 1: ERROR: LoadError: test failed: 9.007199f15 == 9.0072f15
 in expression: T(i,r) == T(BigFloat(i),r)
 in error at ./error.jl:21
 in default_handler at ./test.jl:29
 in do_test at ./test.jl:52
 in anonymous at ./no file:977
 in include at ./boot.jl:253
 in runtests at /build/julia-git/src/julia/test/testdefs.jl:197
 in anonymous at ./multi.jl:644
 in run_work_thunk at ./multi.jl:605
 in remotecall_fetch at ./multi.jl:678
 in remotecall_fetch at ./multi.jl:693
 in anonymous at ./task.jl:1422
while loading /build/julia-git/src/julia/test/numbers.jl, in expression starting on line 972
ERROR: LoadError: LoadError: test failed: 9.007199f15 == 9.0072f15
 in expression: T(i,r) == T(BigFloat(i),r)
 in error at ./error.jl:21
 in default_handler at ./test.jl:29
 in do_test at ./test.jl:52
 in anonymous at ./no file:977
 in include at ./boot.jl:253
 in runtests at /build/julia-git/src/julia/test/testdefs.jl:197
 in anonymous at ./multi.jl:644
 in run_work_thunk at ./multi.jl:605
 in remotecall_fetch at ./multi.jl:678
 in remotecall_fetch at ./multi.jl:693
 in anonymous at ./task.jl:1422
while loading /build/julia-git/src/julia/test/numbers.jl, in expression starting on line 972
while loading /build/julia-git/src/julia/test/runtests.jl, in expression starting on line 5

Failing case

julia> const T = Float32
Float32

julia> i = 9007199254740993
9007199254740993

julia> r = RoundUp
Base.Rounding.RoundingMode{:Up}()

julia> b = BigFloat(i)
9.007199254740993000000000000000000000000000000000000000000000000000000000000000e+15

julia> t1 = T(i, r)
9.007199f15

julia> t2 = T(b, r)
9.0072f15

julia> t1 == t2
false

So it seems that Float32(9007199254740993, RoundUp) is not rounding up on 32bit?

Version Info is the same with the other 32bit issue I posted

julia> versioninfo()
Julia Version 0.4.0-dev+5287
Commit c7e2b33* (2015-06-09 22:19 UTC)
Platform Info:
  System: Linux (i686-pc-linux-gnu)
  CPU: Intel(R) Xeon(R) CPU E5-2680 v2 @ 2.80GHz
  WORD_SIZE: 32
  BLAS: libopenblas (DYNAMIC_ARCH NO_AFFINITY Nehalem)
  LAPACK: libopenblas
  LIBM: libm
  LLVM: libLLVM-3.6.1

[yuyichao]

Here's the LLVM and asm output

julia> @code_llvm Base.Rounding._convert_rounding(Float32, 9007199254740993, RoundUp)

define float @julia__convert_rounding_21203(%jl_value_t*, i64) {
top:
  %2 = sitofp i64 %1 to float
  %3 = fcmp olt float %2, 0x43E0000000000000
  %4 = fptosi float %2 to i64
  %5 = icmp slt i64 %4, %1
  %6 = and i1 %3, %5
  br i1 %6, label %if, label %L

if:                                               ; preds = %top
  %7 = call float @julia_nextfloat_20965(float %2, i32 1)
  ret float %7

L:                                                ; preds = %top
  ret float %2
}

julia> @code_native Base.Rounding._convert_rounding(Float32, 9007199254740993, RoundUp)
        .text
Filename: rounding.jl
Source line: 0
        pushl   %ebp
        movl    %esp, %ebp
        subl    $40, %esp
        movl    12(%ebp), %edx
        movl    16(%ebp), %ecx
Source line: 76
        movl    %ecx, -4(%ebp)
        movl    %edx, -8(%ebp)
        fildll  -8(%ebp)
Source line: 77
        fnstcw  -18(%ebp)
        movw    -18(%ebp), %ax
        movw    $3199, -18(%ebp)        # imm = 0xC7F
        fldcw   -18(%ebp)
        movw    %ax, -18(%ebp)
        fld     %st(0)
        fistpll -16(%ebp)
        fldcw   -18(%ebp)
        flds    -836567040
        fucomp  %st(1)
        fnstsw  %ax
        sahf
        jbe     L98
        cmpl    %edx, -16(%ebp)
        setae   %al
        cmpl    %ecx, -12(%ebp)
        setge   %cl
        je      L78
        movb    %cl, %al
L78:    testb   %al, %al
        jne     L98
        fstps   (%esp)
        movl    $1, 4(%esp)
        calll   0x13c0a3
L98:    addl    $40, %esp
        popl    %ebp
        retl

[simonbyrne]

What's your code_llvm(<,(Float32,Int64))?

[simonbyrne]

I'll need to have another look at it, but the code_llvm looks correct (it has optimised out some more operations than LLVM3.3). Is this an x87 excess precision issue?

[yuyichao]

julia> code_llvm(<, Tuple{Float32, Int64})

define i1 @"julia_<_20846"(float, i64) {
top:
  %2 = sitofp i64 %1 to float
  %3 = fcmp ogt float %2, %0
  %4 = fcmp oeq float %2, %0
  %5 = fcmp olt float %2, 0x43E0000000000000
  %6 = and i1 %4, %5
  %7 = fptosi float %2 to i64
  %8 = icmp slt i64 %7, %1
  %9 = and i1 %6, %8
  %10 = or i1 %3, %9
  ret i1 %10
}

[yuyichao]

BTW, there seems to be a new other failure in this container as well. Should I report them individually?....

[nalimilan]

Just to be sure: you didn't pass MARCH=i386 or i686 when building?

[yuyichao]

@nalimilan Build flags are here MARCH is i686.

IIRC, the build fails (or sth similar) if MARCH is not set....

[yuyichao]

openblas is compiled separately in case it looks confusing in versioninfo()

[nalimilan]

OK, so that's the same issue as #7185 (among others). You need to build Julia with at least MARCH=pentium4 at the moment.

It would make sense to raise an error from the Makefile when i386 or i686 is passed, to avoid wasting people's time.

[yuyichao]

I see.

[yuyichao]

@nalimilan So is #11642 also the same issue?

[nalimilan]

I have no idea, but it's really worth retrying with MARCH=pentium4.

[yuyichao]

@nalimilan Doing that

[yuyichao]

@nalimilan It appears to be the same issue.

[nalimilan]

Nice to hear!

[nalimilan]

This PR should make this less of a trap: #11656