diff options
Diffstat (limited to 'compiler/GHC/Core/Unfold.hs')
| -rw-r--r-- | compiler/GHC/Core/Unfold.hs | 17 |
1 files changed, 11 insertions, 6 deletions
diff --git a/compiler/GHC/Core/Unfold.hs b/compiler/GHC/Core/Unfold.hs index a6ec0b321e..96c1a07ffa 100644 --- a/compiler/GHC/Core/Unfold.hs +++ b/compiler/GHC/Core/Unfold.hs @@ -698,19 +698,24 @@ funSize opts top_args fun n_val_args voids conSize :: DataCon -> Int -> ExprSize conSize dc n_val_args - | n_val_args == 0 = SizeIs 0 emptyBag 10 -- Like variables - -- See Note [Unboxed tuple size and result discount] | isUnboxedTupleDataCon dc = SizeIs 0 emptyBag 10 -- See Note [Constructor size and result discount] - | otherwise = SizeIs (callSize n_val_args 0) emptyBag 10 + | otherwise = SizeIs (6 * n_val_args) emptyBag 10 {- Note [Constructor size and result discount] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ -We generally compute the size of a constructor application as if it was -a regular function call (#22317); the size of the code we generate is similar. -We have an extra case for when it's a nullary application, in which case it is +Constructor applications are generally a bit smaller than function calls, but +we have to ensure that we don't duplicate a lot of big record updates. +So every value argument contributes a size of 6, but SG's experiments concluded +that any factor between 6 and 9 would work: + * A factor of 10 means we don't inline eftWord anymore (undesirable for + T15263) + * But a factor of 5 means we get too much inlining in T22317b. +Picking 6 seems to have the lowest potential for breaking performance changes, +so that's what we go with. +A side effect of the formula is that a nullary constructor application is treated the same as a lone variable. The "result discount" is applied if the result of the call is |