Improve CSE.combineAlts

This patch improves the way that CSE combines identical
alternatives.  See #17901.

I'm still not happy about the duplication between CSE.combineAlts
and GHC.Core.Utils.combineIdenticalAlts; see the Notes with those
functions.  But this patch is a step forward.

Metric Decrease:
    T12425
    T5642

1 files changed, 21 insertions, 4 deletions

diff --git a/compiler/GHC/Core/Utils.hs b/compiler/GHC/Core/Utils.hs
index c846d2ac2e..82d920a58e 100644
--- a/compiler/GHC/Core/Utils.hs
+++ b/compiler/GHC/Core/Utils.hs
@@ -864,10 +864,27 @@ This gave rise to a horrible sequence of cases
 
 and similarly in cascade for all the join points!
 
-NB: it's important that all this is done in [InAlt], *before* we work
-on the alternatives themselves, because Simplify.simplAlt may zap the
-occurrence info on the binders in the alternatives, which in turn
-defeats combineIdenticalAlts (see #7360).
+Note [Combine identical alternatives: wrinkles]
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+* It's important that we try to combine alternatives *before*
+  simplifying them, rather than after. Reason: because
+  Simplify.simplAlt may zap the occurrence info on the binders in the
+  alternatives, which in turn defeats combineIdenticalAlts use of
+  isDeadBinder (see #7360).
+
+  You can see this in the call to combineIdenticalAlts in
+  SimplUtils.prepareAlts.  Here the alternatives have type InAlt
+  (the "In" meaning input) rather than OutAlt.
+
+* combineIdenticalAlts does not work well for nullary constructors
+      case x of y
+         []    -> f []
+         (_:_) -> f y
+  Here we won't see that [] and y are the same.  Sigh! This problem
+  is solved in CSE, in CSE.combineAlts, which does a better version of
+  combineIdenticalAlts. But sadly it doesn't have the occurrence info
+  we have here.  See Note [Combine case alts: awkward corner] in CSE).
 
 Note [Care with impossible-constructors when combining alternatives]
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~