summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorSebastian Graf <sebastian.graf@kit.edu>2018-11-28 13:28:29 +0100
committerSebastian Graf <sebastian.graf@kit.edu>2018-11-28 13:28:29 +0100
commit4c8bf63b78f2fea56a184ef2a40e616f4b8bc59b (patch)
tree40b5419f0c7bfe8c4f0951ea26e24456a6ef7a0b
parent8f52ab9223544b756010a7a92ea52fffdf1d1c71 (diff)
downloadhaskell-4c8bf63b78f2fea56a184ef2a40e616f4b8bc59b.tar.gz
Add Note [Dead case binders in -O0]
After reverting Phab:D5358, Simon (Peyton Jones) asked for a Note summarising why we want to keep the dead case binder check in `cgCase`. Summary from mail conversation: * Phab:D5324 means that we no longer /recompute/ dead-ness of case-binders in STG-land * But TidyPgm preserves dead-ness info (see CoreTidy.tidyIdBndr) * And so we can take advantage of it to avoid a redundant load. This load would be eliminated by CmmSink, but that only happens with -O
-rw-r--r--compiler/codeGen/StgCmmExpr.hs14
1 files changed, 13 insertions, 1 deletions
diff --git a/compiler/codeGen/StgCmmExpr.hs b/compiler/codeGen/StgCmmExpr.hs
index ea64e456ce..30603ee904 100644
--- a/compiler/codeGen/StgCmmExpr.hs
+++ b/compiler/codeGen/StgCmmExpr.hs
@@ -305,7 +305,7 @@ cgCase (StgOpApp (StgPrimOp op) args _) bndr (AlgAlt tycon) alts
= do { tag_expr <- do_enum_primop op args
-- If the binder is not dead, convert the tag to a constructor
- -- and assign it.
+ -- and assign it. See Note [Dead case binders in -O0]
; unless (isDeadBinder bndr) $ do
{ dflags <- getDynFlags
; tmp_reg <- bindArgToReg (NonVoid bndr)
@@ -385,6 +385,18 @@ Now the trouble is that 's' has VoidRep, and we do not bind void
arguments in the environment; they don't live anywhere. See the
calls to nonVoidIds in various places. So we must not look up
's' in the environment. Instead, just evaluate the RHS! Simple.
+
+Note [Dead case binders in -O0]
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+Before CmmSink came to eliminate dead assignments, omitting assignment of dead
+case binders was a cheap and worthwhile optimisation. This probably also was the
+reason for occurrence hacks such as in Phab:D5339 to exist, because the
+occurrence information preserved by 'CoreTidy.tidyIdBndr' was insufficient.
+
+Nowadays, with CmmSink there's little reason to complicate the code by checking
+for dead case binders, except that CmmSink won't run with -O0. Since the
+majority of case binders are dead, this optimisation probably still has a great
+benefit-cost ratio and we want to keep it for -O0. See also Phab:D5358.
-}
cgCase (StgApp v []) _ (PrimAlt _) alts