Re: [GHC] #7814: panic in PPC NCG

6 Apr 2013

      #7814: panic in PPC NCG
--------------------------------+-------------------------------------------
Reporter:  heisenbug            |          Owner:          
    Type:  bug                  |         Status:  new     
Priority:  normal               |      Component:  Compiler
 Version:  7.7                  |       Keywords:          
      Os:  Unknown/Multiple     |   Architecture:  powerpc 
 Failure:  Building GHC failed  |      Blockedby:          
Blocking:                       |        Related:          
--------------------------------+-------------------------------------------

Comment(by heisenbug):

 The problem is in !JoinToTargets:
 {{{
 ==================== Registers allocated
 ====================joinToTargets: making fixup code
             in block: cl
      jmp instruction:   bne     _co
       src assignment: [(cf, InReg (RealRegSingle 30)),
                        (nv, InReg (RealRegSingle 31))]
      dest assignment: [(cf, InReg (RealRegSingle 31))]
       movement graph: [(cf,
                         InReg (RealRegSingle 30),
                         [InReg (RealRegSingle 31)])]
        sccs of graph: [NONREC
                           (cf, InReg (RealRegSingle 30), [InReg
 (RealRegSingle 31)])]

 joinToTargets: fixup code is:
     nI:         mr      31, 30
                 b       _co
 }}}

 Here is the entire output when I disable the panic:

 {{{
 stg_sel_0_upd_info:
 .Lcb:
         bcl     20,31,1f
 1:      mflr    31
         lwz     30, .LnF-(1b)(31)
         add     31, 31, 30
         mr      30, 14
         addi    29, 22, -12
         cmplw   29, 24
         blt     .Lcj
 .Lcl:
         addis   29, 31, .LC_stg_upd_frame_info-(.LCTOC1)+0@ha
         lwz     29, .LC_stg_upd_frame_info-(.LCTOC1)+0@l(29)
         stw     29, -8(22)
         stw     30, -4(22)
         lwz     30, 8(30)
         andi.   29, 30, 3
         cmpwi   29, 0
         bne     .LnI
 .Lcn:
         addis   31, 31, .LC_ce_info-(.LCTOC1)+0@ha
         lwz     31, .LC_ce_info-(.LCTOC1)+0@l(31)
         stw     31, -12(22)
         mr      14, 30
         addi    22, 22, -12
         lwz     31, 0(30)
         mtctr   31
         bctr
 .text
 .align 2
         .long   0
         .long   2097152
 ce_info:
 .Lce:
         addi    22, 22, 12
         mr      31, 14
 .Lco:
         li      30, -4
         and     31, 31, 30
         lwz     14, 4(31)
         addi    22, 22, -8
         addis   31, %vI_nv, .LC_stg_ap_0_fast-(.LCTOC1)+0@ha
         lwz     31, .LC_stg_ap_0_fast-(.LCTOC1)+0@l(31)
         mtctr   31
         bctr
 .Lcj:
         mr      14, 30
         lwz     31, -12(27)
         mtctr   31
         bctr
 .LnI:
         mr      31, 30
         b       .Lco
 }}}

 Apparently the fixup code is in pace (as '.LnI'), but

  1. why is it needed at all? (there are two branches to '_co', from 'ce'
 and 'cl')
  2. in 'cl' (%vI_nv => r31) and in 'ce' %vI_nv is not used but marked as
 live in, dying on exit
  3. who calls 'ce'? appears to have block_ce_info

-- 
Ticket URL: http://hackage.haskell.org/trac/ghc/ticket/7814#comment:5
GHC http://www.haskell.org/ghc/
The Glasgow Haskell Compiler