Re: [GHC] #8131: T7571 with WAY=llvm fails, but not WAY=optllvm

9 Sep 2013

      #8131: T7571 with WAY=llvm fails, but not WAY=optllvm
----------------------------------------------+----------------------------
        Reporter:  thoughtpolice              |            Owner:
            Type:  bug                        |           Status:  new
        Priority:  high                       |        Milestone:
       Component:  Compiler                   |          Version:  7.7
      Resolution:                             |         Keywords:
Operating System:  Unknown/Multiple           |     Architecture:
 Type of failure:  None/Unknown               |  Unknown/Multiple
       Test Case:  llvm/should_compile/T8131  |       Difficulty:  Unknown
        Blocking:                             |       Blocked By:
                                              |  Related Tickets:
----------------------------------------------+----------------------------

Comment (by rwbarton):

 Can also trigger this from Haskell.
 {{{
 {-# LANGUAGE MagicHash, UnboxedTuples #-}
 import GHC.Prim
 import GHC.IO

 main = IO $ \s ->
   let (# s1, p0 #) = newByteArray# 10# s
       (# s2, p #) = unsafeFreezeByteArray# p0 s1
       (# s3, q #) = newByteArray# 10# s2
   in (# copyByteArray# p 0# q 0# 10# s, () #)
 }}}

 {{{
 rwbarton@adjunction:/tmp$ ~/ghc-head/bin/ghc m -fllvm -fforce-recomp
 [1 of 1] Compiling Main             ( m.hs, m.o )
 WARNING: Non constant alignment value given for memcpy! Please report to
 GHC developers
 Linking m ...
 }}}

 The `llvm.memcpy` intrinsic requires that its `align` argument be a
 literal. With `-fllvm` but not `-O`, somewhere along the way the alignment
 argument to the `MO_Memcpy` gets stored in a Cmm register and the register
 gets used for the `llvm.memcpy` call, which isn't good enough. With
 (`-fllvm` and) `-O`, the store to a register gets eliminated and the
 `llvm.memcpy` call does use a literal alignment argument. That's why T7571
 and the Haskell program above work with `-O` but not without. (It doesn't
 have anything to do with the `if (1)` in T7571.) Fragile.

 Your T8131 is expected to not compile, I think, because there the
 alignment argument really is nonconstant.

 Given that the alignment for `MO_Memcpy` is only used in the LLVM backend,
 and there it is required to be constant, I would be inclined to move it
 from a MachOp argument to a parameter of the MachOp constructor, like
 {{{
 data CallishMachOp = ...
   | MO_Memcpy Alignment -- or would another type be more appropriate?
 }}}

 I have a half-finished patch that implements this. Besides eliminating the
 requirement of ensuring the alignment argument is a literal, it also
 removes some special cases in the other backends that have to throw away
 the alignment argument before generating a `memcpy` function call.

-- 
Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/8131#comment:2
GHC http://www.haskell.org/ghc/
The Glasgow Haskell Compiler