#5237: Inefficient code generated for x^2
-------------------------------+--------------------------------------------
Reporter: scpmw | Owner:
Type: bug | Status: patch
Priority: normal | Milestone: 7.4.1
Component: libraries/base | Version: 7.0.3
Keywords: | Testcase:
Blockedby: | Difficulty:
Os: Linux | Blocking:
Architecture: x86_64 (amd64) | Failure: Runtime performance bug
-------------------------------+--------------------------------------------
Comment(by daniel.is.fischer):
Not sure I understand correctly. What we want is that occurrences of `expr
^ 2` or `expr ^ (2 :: Type)` in source code are rewritten to `expr * expr`
(similarly for exponents 3 and 4).
With the rules, that works fine if the exponent isn't given an explicit
type or the type is `Int` or `Integer` - unless the type of `expr` is
`Integer` (or `Int` if the exponent is specified as `Int`). For those
types, `(^)` is specialised, and in the core replaced by the specialised
version, then `GHC.Real.^_^2 expr 2` doesn't match the rule anymore^1^.
Making `(^)` `{-# NOINLINE [1] #-}` instead of `{-# INLINABLE #-}` would
cost a lot in other places, so that's not a good option; out of sheer
morbid curiosity I tried it anyway, it didn't help. But if `INLINABLE`
supports phase control (that isn't mentioned in the users guide, the
syntax is accepted, however), perhaps `(^)` should become `{-# INLINABLE
[1] #-}`?
With `Word` exponents, the problem seems to be that the exponent is moved
to a top-level binding and the rule-matcher sees `expr ^ lvl_fooBar`
instead of `expr ^ (3 :: Word)`.
But I don't think specifying the exponent as `Word` is common, mostly the
exponent will not be given an explicit type and defaulted to `Integer` or
specified as `Int` for better performance.
So I'd just leave out the rules for `Word`.
^1^ Perhaps the compiler just considers the specialisation more specific
than the rules and hence that wins? Compiling with extra rules for the
specialised types to check.
--
Ticket URL: <http://hackage.haskell.org/trac/ghc/ticket/5237#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
