[ISel] Keep matching state consistent when folding during X86 address match

In the X86 backend, matching an address is initiated by the 'addr' complex
pattern and its friends.  During this process we may reassociate and-of-shift
into shift-of-and (FoldMaskedShiftToScaledMask) to allow folding of the
shift into the scale of the address.

However as demonstrated by the testcase, this can trigger CSE of not only the
shift and the AND which the code is prepared for but also the underlying load
node.  In the testcase this node is sitting in the RecordedNode and MatchScope
data structures of the matcher and becomes a deleted node upon CSE.  Returning
from the complex pattern function, we try to access it again hitting an assert
because the node is no longer a load even though this was checked before.

Now obviously changing the DAG this late is bending the rules but I think it
makes sense somewhat.  Outside of addresses we prefer and-of-shift because it
may lead to smaller immediates (FoldMaskAndShiftToScale is an even better
example because it create a non-canonical node).  We currently don't recognize
addresses during DAGCombiner where arguably this canonicalization should be
performed.  On the other hand, having this in the matcher allows us to cover
all the cases where an address can be used in an instruction.

I've also talked a little bit to Dan Gohman on llvm-dev who added the RAUW for
the new shift node in FoldMaskedShiftToScaledMask.  This RAUW is responsible
for initiating the recursive CSE on users
(http://lists.cs.uiuc.edu/pipermail/llvmdev/2014-September/076903.html) but it
is not strictly necessary since the shift is hooked into the visited user.  Of
course it's safer to keep the DAG consistent at all times (e.g. for accurate
number of uses, etc.).

So rather than changing the fundamentals, I've decided to continue along the
previous patches and detect the CSE.  This patch installs a very targeted
DAGUpdateListener for the duration of a complex-pattern match and updates the
matching state accordingly.  (Previous patches used HandleSDNode to detect the
CSE but that's not practical here).  The listener is only installed on X86.

I tested that there is no measurable overhead due to this while running
through the spec2k BC files with llc.  The only thing we pay for is the
creation of the listener.  The callback never ever triggers in spec2k since
this is a corner case.

Fixes rdar://problem/18206171

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219009 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/include/llvm/CodeGen/SelectionDAGISel.h b/include/llvm/CodeGen/SelectionDAGISel.h
index 4d6e8423e6f31a569be935e27eefd99802505131..6c756495db5ad23e3758ded9db0cb72e4ca93725 100644 (file)
--- a/include/llvm/CodeGen/SelectionDAGISel.h
+++ b/include/llvm/CodeGen/SelectionDAGISel.h
@@ -239,6 +239,12 @@ public:
                            const unsigned char *MatcherTable,
                            unsigned TableSize);
 
+  /// \brief Return true if complex patterns for this target can mutate the
+  /// DAG.
+  virtual bool ComplexPatternFuncMutatesDAG() const {
+    return false;
+  }
+
 private:
 
   // Calls to these functions are generated by tblgen.

diff --git a/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp b/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp
index df2c6b8de0b2e9ae8bb907794c6603314ddaab28..31f4b068b909dbcd6f9a2566a3de779275bed9e7 100644 (file)
--- a/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp
+++ b/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp
@@ -2432,6 +2432,42 @@ struct MatchScope {
   bool HasChainNodesMatched, HasGlueResultNodesMatched;
 };
 
+/// \\brief A DAG update listener to keep the matching state
+/// (i.e. RecordedNodes and MatchScope) uptodate if the target is allowed to
+/// change the DAG while matching.  X86 addressing mode matcher is an example
+/// for this.
+class MatchStateUpdater : public SelectionDAG::DAGUpdateListener
+{
+      SmallVectorImpl<std::pair<SDValue, SDNode*> > &RecordedNodes;
+      SmallVectorImpl<MatchScope> &MatchScopes;
+public:
+  MatchStateUpdater(SelectionDAG &DAG,
+                    SmallVectorImpl<std::pair<SDValue, SDNode*> > &RN,
+                    SmallVectorImpl<MatchScope> &MS) :
+    SelectionDAG::DAGUpdateListener(DAG),
+    RecordedNodes(RN), MatchScopes(MS) { }
+
+  void NodeDeleted(SDNode *N, SDNode *E) {
+    // Some early-returns here to avoid the search if we deleted the node or
+    // if the update comes from MorphNodeTo (MorphNodeTo is the last thing we
+    // do, so it's unnecessary to update matching state at that point).
+    // Neither of these can occur currently because we only install this
+    // update listener during matching a complex patterns.
+    if (!E || E->isMachineOpcode())
+      return;
+    // Performing linear search here does not matter because we almost never
+    // run this code.  You'd have to have a CSE during complex pattern
+    // matching.
+    for (auto &I : RecordedNodes)
+      if (I.first.getNode() == N)
+        I.first.setNode(E);
+
+    for (auto &I : MatchScopes)
+      for (auto &J : I.NodeStack)
+        if (J.getNode() == N)
+          J.setNode(E);
+  }
+};
 }
 
 SDNode *SelectionDAGISel::
@@ -2686,6 +2722,14 @@ SelectCodeCommon(SDNode *NodeToMatch, const unsigned char *MatcherTable,
       unsigned CPNum = MatcherTable[MatcherIndex++];
       unsigned RecNo = MatcherTable[MatcherIndex++];
       assert(RecNo < RecordedNodes.size() && "Invalid CheckComplexPat");
+
+      // If target can modify DAG during matching, keep the matching state
+      // consistent.
+      std::unique_ptr<MatchStateUpdater> MSU;
+      if (ComplexPatternFuncMutatesDAG())
+        MSU.reset(new MatchStateUpdater(*CurDAG, RecordedNodes,
+                                        MatchScopes));
+
       if (!CheckComplexPattern(NodeToMatch, RecordedNodes[RecNo].second,
                                RecordedNodes[RecNo].first, CPNum,
                                RecordedNodes))

diff --git a/lib/Target/X86/X86ISelDAGToDAG.cpp b/lib/Target/X86/X86ISelDAGToDAG.cpp
index a8587cd9a2a7ebf18cd7dfdea1b45a7beec0322b..429c8bdbdf0a503b41e8a2d9f2f7e95b47e052a2 100644 (file)
--- a/lib/Target/X86/X86ISelDAGToDAG.cpp
+++ b/lib/Target/X86/X86ISelDAGToDAG.cpp
@@ -299,6 +299,13 @@ namespace {
     const X86InstrInfo *getInstrInfo() const {
       return getTargetMachine().getSubtargetImpl()->getInstrInfo();
     }
+
+    /// \brief Address-mode matching performs shift-of-and to and-of-shift
+    /// reassociation in order to expose more scaled addressing
+    /// opportunities.
+    bool ComplexPatternFuncMutatesDAG() const override {
+      return true;
+    }
   };
 }
 

diff --git a/test/CodeGen/X86/addr-mode-matcher.ll b/test/CodeGen/X86/addr-mode-matcher.ll
new file mode 100644 (file)
index 0000000..d592091
--- /dev/null
+++ b/test/CodeGen/X86/addr-mode-matcher.ll
@@ -0,0 +1,62 @@
+; RUN: llc < %s | FileCheck %s
+
+; This testcase used to hit an assert during ISel.  For details, see the big
+; comment inside the function.
+
+; CHECK-LABEL: foo:
+; The AND should be turned into a subreg access.
+; CHECK-NOT: and
+; The shift (leal) should be folded into the scale of the address in the load.
+; CHECK-NOT: leal
+; CHECK: movl {{.*}},4),
+
+target datalayout = "e-m:o-p:32:32-f64:32:64-f80:128-n8:16:32-S128"
+target triple = "i386-apple-macosx10.6.0"
+
+define void @foo(i32 %a) {
+bb:
+  br label %bb1692
+
+bb1692:
+  %tmp1694 = phi i32 [ 0, %bb ], [ %tmp1745, %bb1692 ]
+  %xor = xor i32 0, %tmp1694
+
+; %load1 = (load (and (shl %xor, 2), 1020))
+  %tmp1701 = shl i32 %xor, 2
+  %tmp1702 = and i32 %tmp1701, 1020
+  %tmp1703 = getelementptr inbounds [1028 x i8]* null, i32 0, i32 %tmp1702
+  %tmp1704 = bitcast i8* %tmp1703 to i32*
+  %load1 = load i32* %tmp1704, align 4
+
+; %load2 = (load (shl (and %xor, 255), 2))
+  %tmp1698 = and i32 %xor, 255
+  %tmp1706 = shl i32 %tmp1698, 2
+  %tmp1707 = getelementptr inbounds [1028 x i8]* null, i32 0, i32 %tmp1706
+  %tmp1708 = bitcast i8* %tmp1707 to i32*
+  %load2 = load i32* %tmp1708, align 4
+
+  %tmp1710 = or i32 %load2, %a
+
+; While matching xor we address-match %load1.  The and-of-shift reassocication
+; in address matching transform this into into a shift-of-and and the resuting
+; node becomes identical to %load2.  CSE replaces %load1 which leaves its
+; references in MatchScope and RecordedNodes stale.
+  %tmp1711 = xor i32 %load1, %tmp1710
+
+  %tmp1744 = getelementptr inbounds [256 x i32]* null, i32 0, i32 %tmp1711
+  store i32 0, i32* %tmp1744, align 4
+  %tmp1745 = add i32 %tmp1694, 1
+  indirectbr i8* undef, [label %bb1756, label %bb1692]
+
+bb1756:
+  br label %bb2705
+
+bb2705:
+  indirectbr i8* undef, [label %bb5721, label %bb5736]
+
+bb5721:
+  br label %bb2705
+
+bb5736:
+  ret void
+}