#14687: Investigate differences in Int-In/Outlining -------------------------------------+------------------------------------- Reporter: AndreasK | Owner: mpickering Type: task | Status: new Priority: lowest | Milestone: Component: Compiler | Version: 8.2.2 Resolution: | Keywords: Operating System: Unknown/Multiple | Architecture: | Unknown/Multiple Type of failure: None/Unknown | Test Case: Blocked By: | Blocking: Related Tickets: | Differential Rev(s): Wiki Page: | -------------------------------------+------------------------------------- Comment (by mpickering): Here is what is happening. I don't know if there are any bugs here or not. The core for `func` looks like {{{ -- RHS size: {terms: 17, types: 4, coercions: 0, joins: 0/0} func :: Int# -> Int# -> Int [LclIdX, Arity=2, Unf=Unf{Src=<vanilla>, TopLvl=True, Value=True, ConLike=True, WorkFree=True, Expandable=True, Guidance=IF_ARGS [50 40] 90 0}] func = \ (ds_dWL :: Int#) (ds_dWM :: Int#) -> case ds_dWL of { __DEFAULT -> fail_sXf void#; 1# -> case ds_dWM of { __DEFAULT -> fail_sXf void#; 1# -> lvl_sXg; 2# -> lvl_sXh } } }}} whilst the core for `foo` looks like {{{ -- RHS size: {terms: 23, types: 5, coercions: 0, joins: 0/0} foo :: Int# -> Int# -> Int [LclIdX, Arity=2, Unf=Unf{Src=<vanilla>, TopLvl=True, Value=True, ConLike=True, WorkFree=True, Expandable=True, Guidance=IF_ARGS [100 60] 130 0}] foo = \ (ds_dWG :: Int#) (ds_dWH :: Int#) -> case ds_dWG of { __DEFAULT -> fail_sXk void#; 1# -> case ds_dWH of { __DEFAULT -> fail_sXk void#; 1# -> lvl_sXl }; 2# -> case ds_dWH of { __DEFAULT -> fail_sXk void#; 1# -> lvl_sXm } } }}} It is key to note at this stage that the size of `func` is 90 whilst the size of `foo` is 130. GHC has already dutifully floated out the integers at this stage so how do they reappear? `foo` is then W/W but `func` is not. The reason for this is that GHC decides that `func` will certainly get inlined due to the following line in `CoreUnfold` {{{ , size - (10 * (arity + 1)) <= ufUseThreshold dflags }}} As the size of `func` is 90, this condition returns `True` (90-30 = 60, ufUseThreshold=60). So then what happens to `foo`? Well it is W/Wed and then the constants appear in the worker immediately in `post-worker-wrapper` simplifier run. It looks likely because the worker has a case-of-case opportunity which then leaves the variable in a case position. However I didn't verify exactly why the variable is inlined as it didn't appear in `-ddump- inlinings`. The worker is then inlined back into the wrapper and thus they reappear in the definition. -- Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/14687#comment:2 GHC http://www.haskell.org/ghc/ The Glasgow Haskell Compiler