412016 - TT: remove redundant NaN checks

Status: VERIFIED DUPLICATE of bug 416398

(Tamarin Graveyard :: Tracing Virtual Machine, defect, P2)

Description

[Edwin Smith]

say x = a+(b*c)+1

each operation that returns an intermediate double also contains a NaN check as part of the dbox operation used on the double.  However those nan checks don't amount to anything if the value is immediately used as a double; even if the temp is a NaN, the code flow would not have changed.

here is a snip from math-spectral-norm:

    247 fadd  106,111
    250 eq    247,247      NaN check 1
    254 xf    250

    333 quad  #3FF00000:0
    334 fadd  247,333
    337 eq    334,334      NaN check 2 (obsoletes check 1)
    341 xf    337

    402 fmul  75,334
    405 eq    402,402      NaN check 3 (obsoletes check 2)
    409 xf    405

    476 quad  #40000000:0
    477 fdiv  402,476
    480 eq    477,477      NaN check 4 (obsoletes check 3)
    484 xf    480

    535 xf    118

    633 fadd  477,106
    636 eq    633,633      NaN check 5 (see a pattern?)
    640 xf    636

    716 quad  #3FF00000:0
    717 fadd  633,716
    720 eq    717,717      NaN check 6 (and so on)
    724 xf    720

Comment 1

[Edwin Smith]

x (op) const is not nan when x is i2d() or u2d() and const is an integer valued double.