[DAGCombiner] Use SmallDenseMap (NFC)

[kazutakahirata]

The use of SmallDenseMap saves 0.48% of heap allocations during the
compilation of a large preprocessed file, namely X86ISelLowering.cpp,
for the X86 target. During the experiment, the maximum size of
WorklistMap was 24 or less 74% of the time. (Note that DenseMap has
the maximum occupancy rate of 3/4.)

[llvmbot]

@llvm/pr-subscribers-llvm-selectiondag

Author: Kazu Hirata (kazutakahirata)

Changes

The use of SmallDenseMap saves 0.48% of heap allocations during the
compilation of a large preprocessed file, namely X86ISelLowering.cpp,
for the X86 target. During the experiment, the maximum size of
WorklistMap was 24 or less 74% of the time. (Note that DenseMap has
the maximum occupancy rate of 3/4.)

---

Full diff: https://github.com/llvm/llvm-project/pull/79681.diff

1 Files Affected:

• (modified) llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp (+1-1)

diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 87184fe409eade..ab572c5eca7399 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -174,7 +174,7 @@ namespace {
     /// This is used to find and remove nodes from the worklist (by nulling
     /// them) when they are deleted from the underlying DAG. It relies on
     /// stable indices of nodes within the worklist.
-    DenseMap<SDNode *, unsigned> WorklistMap;
+    SmallDenseMap<SDNode *, unsigned, 32> WorklistMap;
 
     /// This records all nodes attempted to be added to the worklist since we
     /// considered a new worklist entry. As we keep do not add duplicate nodes

[nikic]

LGTM

[nikic]

Looks like this causes a regression rather than an improvement in practice: https://llvm-compile-time-tracker.com/compare.php?from=57a20d2d09bbd3cc501fc6d8b4746be2040c99b7&to=863b2c84c0fbcfb02d969fa36af4932d410a827b&stat=instructions%3Au

I think this is because SmallDenseMap (unlike SmallVector) actually incurs an extra cost because it has to always check the two representations dynamically.

[kazutakahirata]

Thanks. I've reverted the patch for now: faf555f

I guess if we are to use an inline buffer for SmallDenseMap, the size of the inline buffer must accommodate the vast majority (90%+?) of the use cases.

[nikic]

@kazutakahirata It probably depends a lot on how often the DenseMap is accessed. If it is accessed much more often that it is grown, that may outweigh the allocation overhead.