#4308: LLVM compiles Updates.cmm badly
---------------------------------+------------------------------------------
Reporter: dterei | Owner: dterei
Type: bug | Status: new
Priority: normal | Milestone: 7.2.1
Component: Compiler (LLVM) | Version: 6.13
Keywords: | Testcase:
Blockedby: | Difficulty:
Os: Unknown/Multiple | Blocking:
Architecture: x86_64 (amd64) | Failure: Runtime performance bug
---------------------------------+------------------------------------------
Comment(by dterei):
OK so problem with C call in a branch is due to the LLVM optimiser. In one
branch the <Rn> stg registers are the <Rn(old)> value while in the branch
with the C call they're 'undef'. So you would expect there to be a phi
node like this:
{{{
join:
R3 = phi [lbranch, undef], [rbranch, R3_old]
...
}}}
However the optimiser doesn't do this as it simplifies phi nodes with
'undef' values by removing the undef value:
{{{
join:
R3 = R3_old
}}}
So this is why we get the bad code. Our undef's are removed by the
optimiser and so the code generator doesn't know it can let the stg
registers be trashed.
When the C call occurs in a non-branch block the 'undef' technique works
as the llvm optimiser bubbles down the undef value to the tail call, eg:
{{{
mid_block:
store i64 undef R3
...
%0 = call ccc sin()
...
br %final
final:
tail call g (Base, Hp, Sp, R1, R2, R3, ...)
}}}
Becomes:
{{{
mid_block:
%0 = call ccc sin()
...
br %final
final:
tail call g (Base, Hp, Sp, R1, R2, undef, ...)
}}}
So how to fix? well could try to change the semantics of 'undef' in LLVM.
But that seems a risky approach as I don't think the LLVM developers will
want to do that. Also there is a better way. We change the Cmm
representation to carry with all calls and jumps the live STG register
information. Then I don't need to keep passing the stg registers around
for every call and can free up some registers for intermediate code. Also
will need to stop the saving and restoring of caller save variables being
done in Cmm and let LLVM do it instead.
--
Ticket URL: <http://hackage.haskell.org/trac/ghc/ticket/4308#comment:6>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler
_______________________________________________
Glasgow-haskell-bugs mailing list
[email protected]
http://www.haskell.org/mailman/listinfo/glasgow-haskell-bugs
