[GHC] #10750: silly assembly for comparing Doubles

[GHC]

#10750: silly assembly for comparing Doubles
-------------------------------------+-------------------------------------
              Reporter:  rwbarton    |             Owner:
                  Type:  bug         |            Status:  new
              Priority:  normal      |         Milestone:
             Component:  Compiler    |           Version:  7.11
  (CodeGen)                          |
              Keywords:              |  Operating System:  Unknown/Multiple
          Architecture:              |   Type of failure:  Runtime
  Unknown/Multiple                   |  performance bug
             Test Case:              |        Blocked By:
              Blocking:              |   Related Tickets:  #10137, #10677
Differential Revisions:              |
-------------------------------------+-------------------------------------
 A function like `f x = if x > 0.0 then ... else ...` compiles to the
 assembly
 {{{
         movsd 7(%rbx),%xmm0
         xorpd %xmm1,%xmm1
         ucomisd %xmm1,%xmm0
         seta %al
         movzbl %al,%eax
         cmpq $1,%rax
         jb _c1tb
 }}}
 This `seta/movzbl/cmpq/jb` is bad, we can just generate `ja` (and 7.10
 did).

 The cause is that the code generator produces Cmm like
 {{{
            switch [0 .. 1] %MO_F_Gt_W64(_s1sv::F64, 0.0 :: W64)::I64 {
                case 0 : goto c1tb;
                case 1 : goto c1tc;
            }
 }}}
 which turns into
 {{{
       if (%MO_F_Gt_W64(_s1sv::F64,
                        0.0 :: W64) < 1) goto c1tb; else goto c1tc;
 }}}
 and then GHC is stuck. It knows how to turn `condition >= 1` into
 `condition`, and it knows how to turn `condition < 1` into a negated
 version of `condition` when possible, but there is no negated version of
 `MO_F_Gt_W64` (it's not `MO_F_Le_W64` because of NaNs and signed zeros).
 It doesn't know how to turn a negated conditional into a conditional with
 the branches swapped because it doesn't look at that much at once.

 Presumably more fallout from #10137, and maybe can be fixed simultaneously
 with #10677.

--
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/10750>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler