Re: [GHC] #10124: Simple case analyses generate too many branches

31 Mar 2015

      #10124: Simple case analyses generate too many branches
-------------------------------------+-------------------------------------
        Reporter:  bgamari           |                   Owner:
            Type:  bug               |                  Status:  new
        Priority:  normal            |               Milestone:
       Component:  Compiler          |                 Version:  7.8.4
      Resolution:                    |                Keywords:
Operating System:  Unknown/Multiple  |            Architecture:
 Type of failure:  Runtime           |  Unknown/Multiple
  performance bug                    |               Test Case:
      Blocked By:                    |                Blocking:
 Related Tickets:  #6135,            |  Differential Revisions:
  #9661,#10137                       |
-------------------------------------+-------------------------------------

Comment (by nomeata):

 Hmm, given the example from comment:10, we currently generate this code:
 {{{
 block_c3S3_info:
 _c3S3:
         movq 7(%rbx),%rax
         cmpq $10,%rax
         jb _u3Si
 _u3Sj:
         cmpq $11,%rax
         jb _c3Sh
 _u3Sk:
         cmpq $32,%rax
         jne _c3Se
 _c3Sh:
         movl $GHC.Types.True_closure+2,%ebx
         addq $8,%rbp
         jmp *(%rbp)
 _c3S7:
         movl $myIsSpace_rkB_closure,%ebx
         jmp *-8(%r13)
 _c3Se:
         movl $GHC.Types.False_closure+1,%ebx
         addq $8,%rbp
         jmp *(%rbp)
 _u3Si:
         cmpq $9,%rax
         jb _c3Se
         jmp _c3Sh
 }}}
 while my code now generates
 {{{
 _c3S3:
         movq 7(%rbx),%rax
         cmpq $32,%rax
         setne %bl
         movzbl %bl,%ebx
         cmpq $10,%rax
         setne %cl
         movzbl %cl,%ecx
         andq %rbx,%rcx
         cmpq $9,%rax
         setne %al
         movzbl %al,%eax
         andq %rcx,%rax
         testq %rax,%rax
         jne _c3Se
 _c3Sh:
         movl $GHC.Types.True_closure+2,%ebx
         addq $8,%rbp
         jmp *(%rbp)
 _c3S7:
         movl $myIsSpace_rkB_closure,%ebx
         jmp *-8(%r13)
 _c3Se:
         movl $GHC.Types.False_closure+1,%ebx
         addq $8,%rbp
         jmp *(%rbp)
         .size myIsSpace_rkB_info, .-myIsSpace_rkB_info
 }}}

 But not even this simple microbenchmark
 {{{
 main :: IO ()
 main = x `seq` return ()
   where x = length $ filter myIsSpace $ concatMap (replicate 100000000) $
 ['\001'..'z']
 }}}
 shows a change in performance. bgamari, is that expected? Is it just that
 the assembly (generated from
 {{{
 if ((_s3Rr::I64 != 9) & (_s3Rr::I64 != 10) & (_s3Rr::I64 != 32) != 0) goto
 c3Se; else goto c3Sh;
 }}}
 is too bad, or is this not a test case where this will help a lot?

--
Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/10124#comment:15
GHC http://www.haskell.org/ghc/
The Glasgow Haskell Compiler