linalg/pinv failure on 32bit

[yuyichao]

I got the following error on a 32bit container which I think I haven't seen on the CI before.

[root@yuyichao julia]# ./julia test/runtests.jl linalg/pinv
     * linalg/pinv         exception on 1: ERROR: LoadError: mismatch of non-finite elements: 
  a = Inf
  0.0 = 0.0
 in test_approx_eq at ./test.jl:131
 in anonymous at ./no file:298
 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/linalg/pinv.jl, in expression starting on line 294
ERROR: LoadError: LoadError: mismatch of non-finite elements: 
  a = Inf
  0.0 = 0.0
 in test_approx_eq at ./test.jl:131
 in anonymous at ./no file:298
 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/linalg/pinv.jl, in expression starting on line 294
while loading /build/julia-git/src/julia/test/runtests.jl, in expression starting on line 5

Version info.

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]

A slightly better repro.

julia> using Base.Test

julia> debug = true
true

julia> for eltya in (Float32, Float64)
           debug && println("\n\n<<<<<", eltya, ">>>>>")
           debug && println("\n--- sub-normal constant ---")
           a = pinv(realmin(eltya)/100)
           @test_approx_eq a 0.0
           debug && println("\n\n<<<<<Complex{", eltya, "}>>>>>")
           debug && println("\n--- sub-normal constant ---")
           a = pinv(realmin(eltya)/100*(1+1im))
           @test_approx_eq a 0.0
       end


<<<<<Float32>>>>>

--- sub-normal constant ---
ERROR: mismatch of non-finite elements: 
  a = Inf
  0.0 = 0.0
 in test_approx_eq at ./test.jl:131
 in anonymous at ./no file:5

[yuyichao]

Maybe a llvm 3.6.1 bug?

julia> a = (realmin(Float32) / 100)
1.1755f-40                                                                        

julia> @code_llvm pinv(a)

define float @julia_pinv_21040(float) {                                           
top:                                                                              
  %1 = fdiv float 1.000000e+00, %0                                                
  %2 = fsub float %1, %1                                                          
  store volatile float %2, float* bitcast (i128* @jl_float_temp to float*), align 4                                                                                 
  %3 = load volatile float* bitcast (i128* @jl_float_temp to float*), align 4     
  %4 = fcmp une float %3, 0.000000e+00                                            
  %5 = select i1 %4, float 0.000000e+00, float %1                                 
  ret float %5                                                                    
}                                                                                 

julia> inv(a)
Inf32                                                                             

julia> isfinite(inv(a))
false

julia> @code_lowered pinv(a)
1-element Array{Any,1}:
 :($(Expr(:lambda, Any[:x], Any[Any[Any[:x,:Any,0],Any[:xi,:Any,18]],Any[],0,Any[]], :(begin  # linalg/dense.jl, line 442:
        xi = inv(x) # line 443:
        return ifelse(isfinite(xi),xi,zero(xi))
    end))))

julia> @code_native pinv(a)
        .text
Filename: dense.jl
Source line: 0
        pushl   %ebp
        movl    %esp, %ebp
        fld1
Source line: 442
        fdivs   8(%ebp)
Source line: 443
        fld     %st(0)
        fsub    %st(1)
        fstps   -835670016
        flds    -835670016
        fldz
        fxch    %st(1)
        fucomp  %st(1)
        fnstsw  %ax
        sahf
        jne     L43
        jp      L43
        fstp    %st(0)
        fldz
        fxch    %st(1)
L43:    fstp    %st(1)
        popl    %ebp

[yuyichao]

Close because this is a known issue with MARCH=i686