Eta reduction based on evaluation context (#21261)wip/T21261

I completely rewrote our Notes surrounding eta-reduction. The new entry point is
`Note [Eta reduction makes sense]`.

Then I went on to extend the Simplifier to maintain an evaluation context in the
form of a `SubDemand` inside a `SimplCont`. That `SubDemand` is useful for doing
eta reduction according to `Note [Eta reduction based on evaluation context]`,
which describes how Demand analysis, Simplifier and `tryEtaReduce` interact to
facilitate eta reduction in more scenarios.

Thus we fix #21261.

ghc/alloc perf marginally improves (-0.0%). A medium-sized win is when compiling
T3064 (-3%). It seems that haddock improves by 0.6% to 1.0%, too.

Metric Decrease:
    T3064

1 files changed, 3 insertions, 3 deletions

diff --git a/compiler/GHC/Core/SimpleOpt.hs b/compiler/GHC/Core/SimpleOpt.hs
index fa6fcda351..925eaf5841 100644
--- a/compiler/GHC/Core/SimpleOpt.hs
+++ b/compiler/GHC/Core/SimpleOpt.hs
@@ -36,7 +36,7 @@ import GHC.Types.Var      ( isNonCoVarId )
 import GHC.Types.Var.Set
 import GHC.Types.Var.Env
 import GHC.Core.DataCon
-import GHC.Types.Demand( etaConvertDmdSig )
+import GHC.Types.Demand( etaConvertDmdSig, topSubDmd )
 import GHC.Types.Tickish
 import GHC.Core.Coercion.Opt ( optCoercion, OptCoercionOpts (..) )
 import GHC.Core.Type hiding ( substTy, extendTvSubst, extendCvSubst, extendTvSubstList
@@ -295,8 +295,8 @@ simple_opt_expr env expr
        where
          (env', b') = subst_opt_bndr env b
     go_lam env bs' e
-       | Just etad_e <- tryEtaReduce bs e' = etad_e
-       | otherwise                         = mkLams bs e'
+       | Just etad_e <- tryEtaReduce bs e' topSubDmd = etad_e
+       | otherwise                                   = mkLams bs e'
        where
          bs = reverse bs'
          e' = simple_opt_expr env e