Fix some issues in WalkChainUsers dealing with
CopyToReg/CopyFromReg/INLINEASM. These are annoying because
they have the same opcode before an after isel. Fix this by
setting their NodeID to -1 to indicate that they are selected,
just like what automatically happens when selecting things that
end up being machine nodes.
With that done, give IsLegalToFold a new flag that causes it to
ignore chains. This lets the HandleMergeInputChains routine be
the one place that validates chains after a match is successful,
enabling the new hotness in chain processing. This smarter
chain processing eliminates the need for "PreprocessRMW" in the
X86 and MSP430 backends and enables MSP to start matching it's
multiple mem operand instructions more aggressively.
I currently #if out the dead code in the X86 backend and MSP
backend, I'll remove it for real in a follow-on patch.
The testcase changes are:
test/CodeGen/X86/sse3.ll: we generate better code
test/CodeGen/X86/store_op_load_fold2.ll: PreprocessRMW was
miscompiling this before, we now generate correct code
Convert it to filecheck while I'm at it.
test/CodeGen/MSP430/Inst16mm.ll: Add a testcase for mem/mem
folding to make anton happy. :)
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@97596 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/include/llvm/CodeGen/SelectionDAGISel.h b/include/llvm/CodeGen/SelectionDAGISel.h
index af1273c..9ee8604 100644
--- a/include/llvm/CodeGen/SelectionDAGISel.h
+++ b/include/llvm/CodeGen/SelectionDAGISel.h
@@ -97,7 +97,8 @@
/// IsLegalToFold - Returns true if the specific operand node N of
/// U can be folded during instruction selection that starts at Root.
- virtual bool IsLegalToFold(SDValue N, SDNode *U, SDNode *Root) const;
+ virtual bool IsLegalToFold(SDValue N, SDNode *U, SDNode *Root,
+ bool IgnoreChains = false) const;
/// CreateTargetHazardRecognizer - Return a newly allocated hazard recognizer
/// to use for this target when scheduling the DAG.
diff --git a/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp b/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp
index ff8ab46..8e88f16 100644
--- a/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp
+++ b/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp
@@ -1376,8 +1376,8 @@
/// This function recursively traverses up the operand chain, ignoring
/// certain nodes.
static bool findNonImmUse(SDNode *Use, SDNode* Def, SDNode *ImmedUse,
- SDNode *Root,
- SmallPtrSet<SDNode*, 16> &Visited) {
+ SDNode *Root, SmallPtrSet<SDNode*, 16> &Visited,
+ bool IgnoreChains) {
// The NodeID's are given uniques ID's where a node ID is guaranteed to be
// greater than all of its (recursive) operands. If we scan to a point where
// 'use' is smaller than the node we're scanning for, then we know we will
@@ -1395,6 +1395,10 @@
return false;
for (unsigned i = 0, e = Use->getNumOperands(); i != e; ++i) {
+ // Ignore chain uses, they are validated by HandleMergeInputChains.
+ if (Use->getOperand(i).getValueType() == MVT::Other && IgnoreChains)
+ continue;
+
SDNode *N = Use->getOperand(i).getNode();
if (N == Def) {
if (Use == ImmedUse || Use == Root)
@@ -1404,7 +1408,7 @@
}
// Traverse up the operand chain.
- if (findNonImmUse(N, Def, ImmedUse, Root, Visited))
+ if (findNonImmUse(N, Def, ImmedUse, Root, Visited, IgnoreChains))
return true;
}
return false;
@@ -1419,9 +1423,10 @@
/// have one non-chain use, we only need to watch out for load/op/store
/// and load/op/cmp case where the root (store / cmp) may reach the load via
/// its chain operand.
-static inline bool isNonImmUse(SDNode *Root, SDNode *Def, SDNode *ImmedUse) {
+static inline bool isNonImmUse(SDNode *Root, SDNode *Def, SDNode *ImmedUse,
+ bool IgnoreChains) {
SmallPtrSet<SDNode*, 16> Visited;
- return findNonImmUse(Root, Def, ImmedUse, Root, Visited);
+ return findNonImmUse(Root, Def, ImmedUse, Root, Visited, IgnoreChains);
}
/// IsProfitableToFold - Returns true if it's profitable to fold the specific
@@ -1434,7 +1439,8 @@
/// IsLegalToFold - Returns true if the specific operand node N of
/// U can be folded during instruction selection that starts at Root.
-bool SelectionDAGISel::IsLegalToFold(SDValue N, SDNode *U, SDNode *Root) const {
+bool SelectionDAGISel::IsLegalToFold(SDValue N, SDNode *U, SDNode *Root,
+ bool IgnoreChains) const {
if (OptLevel == CodeGenOpt::None) return false;
// If Root use can somehow reach N through a path that that doesn't contain
@@ -1488,7 +1494,7 @@
VT = Root->getValueType(Root->getNumValues()-1);
}
- return !isNonImmUse(Root, N.getNode(), U);
+ return !isNonImmUse(Root, N.getNode(), U, IgnoreChains);
}
SDNode *SelectionDAGISel::Select_INLINEASM(SDNode *N) {
@@ -1500,6 +1506,7 @@
VTs.push_back(MVT::Flag);
SDValue New = CurDAG->getNode(ISD::INLINEASM, N->getDebugLoc(),
VTs, &Ops[0], Ops.size());
+ New->setNodeId(-1);
return New.getNode();
}
@@ -1636,11 +1643,17 @@
// pattern that we're selecting down into the already selected chunk of the
// DAG.
if (User->isMachineOpcode() ||
- User->getOpcode() == ISD::CopyToReg ||
- User->getOpcode() == ISD::CopyFromReg ||
- User->getOpcode() == ISD::INLINEASM ||
User->getOpcode() == ISD::HANDLENODE) // Root of the graph.
continue;
+
+ if (User->getOpcode() == ISD::CopyToReg ||
+ User->getOpcode() == ISD::CopyFromReg ||
+ User->getOpcode() == ISD::INLINEASM) {
+ // If their node ID got reset to -1 then they've already been selected.
+ // Treat them like a MachineOpcode.
+ if (User->getNodeId() == -1)
+ continue;
+ }
// If we have a TokenFactor, we handle it specially.
if (User->getOpcode() != ISD::TokenFactor) {
@@ -1876,6 +1889,7 @@
case ISD::TokenFactor:
case ISD::CopyFromReg:
case ISD::CopyToReg:
+ NodeToMatch->setNodeId(-1); // Mark selected.
return 0;
case ISD::AssertSext:
case ISD::AssertZext:
@@ -2172,7 +2186,7 @@
if (!IsProfitableToFold(N, NodeStack[NodeStack.size()-2].getNode(),
NodeToMatch) ||
!IsLegalToFold(N, NodeStack[NodeStack.size()-2].getNode(),
- NodeToMatch))
+ NodeToMatch, true/*We validate our own chains*/))
break;
continue;
diff --git a/lib/Target/MSP430/MSP430ISelDAGToDAG.cpp b/lib/Target/MSP430/MSP430ISelDAGToDAG.cpp
index 54062a0..2ca184e 100644
--- a/lib/Target/MSP430/MSP430ISelDAGToDAG.cpp
+++ b/lib/Target/MSP430/MSP430ISelDAGToDAG.cpp
@@ -125,7 +125,9 @@
bool MatchWrapper(SDValue N, MSP430ISelAddressMode &AM);
bool MatchAddressBase(SDValue N, MSP430ISelAddressMode &AM);
+#if 0
bool IsLegalToFold(SDValue N, SDNode *U, SDNode *Root) const;
+#endif
virtual bool
SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode,
@@ -323,6 +325,7 @@
return false;
}
+#if 0
bool MSP430DAGToDAGISel::IsLegalToFold(SDValue N, SDNode *U,
SDNode *Root) const {
if (OptLevel == CodeGenOpt::None) return false;
@@ -357,6 +360,7 @@
// Proceed to 'generic' cycle finder code
return SelectionDAGISel::IsLegalToFold(N, U, Root);
}
+#endif
/// MoveBelowTokenFactor - Replace TokenFactor operand with load's chain operand
@@ -516,6 +520,7 @@
/// This allows selection of mem-mem instructions. Yay!
void MSP430DAGToDAGISel::PreprocessForRMW() {
+ return;
for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
E = CurDAG->allnodes_end(); I != E; ++I) {
if (!ISD::isNON_TRUNCStore(I))
diff --git a/lib/Target/X86/X86ISelDAGToDAG.cpp b/lib/Target/X86/X86ISelDAGToDAG.cpp
index e162f4e..5ec5065 100644
--- a/lib/Target/X86/X86ISelDAGToDAG.cpp
+++ b/lib/Target/X86/X86ISelDAGToDAG.cpp
@@ -467,46 +467,6 @@
}
-/// PreprocessForRMW - Preprocess the DAG to make instruction selection better.
-/// This is only run if not in -O0 mode.
-/// This allows the instruction selector to pick more read-modify-write
-/// instructions. This is a common case:
-///
-/// [Load chain]
-/// ^
-/// |
-/// [Load]
-/// ^ ^
-/// | |
-/// / \-
-/// / |
-/// [TokenFactor] [Op]
-/// ^ ^
-/// | |
-/// \ /
-/// \ /
-/// [Store]
-///
-/// The fact the store's chain operand != load's chain will prevent the
-/// (store (op (load))) instruction from being selected. We can transform it to:
-///
-/// [Load chain]
-/// ^
-/// |
-/// [TokenFactor]
-/// ^
-/// |
-/// [Load]
-/// ^ ^
-/// | |
-/// | \-
-/// | |
-/// | [Op]
-/// | ^
-/// | |
-/// \ /
-/// \ /
-/// [Store]
void X86DAGToDAGISel::PreprocessForRMW() {
for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
E = CurDAG->allnodes_end(); I != E; ++I) {
@@ -538,6 +498,9 @@
++NumLoadMoved;
continue;
}
+
+ continue;
+
if (!ISD::isNON_TRUNCStore(I))
continue;
@@ -1415,11 +1378,12 @@
SDValue &Base, SDValue &Scale,
SDValue &Index, SDValue &Disp,
SDValue &Segment) {
- if (ISD::isNON_EXTLoad(N.getNode()) &&
- IsProfitableToFold(N, P, P) &&
- IsLegalToFold(N, P, P))
- return SelectAddr(P, N.getOperand(1), Base, Scale, Index, Disp, Segment);
- return false;
+ if (!ISD::isNON_EXTLoad(N.getNode()) ||
+ !IsProfitableToFold(N, P, P) ||
+ !IsLegalToFold(N, P, P))
+ return false;
+
+ return SelectAddr(P, N.getOperand(1), Base, Scale, Index, Disp, Segment);
}
/// getGlobalBaseReg - Return an SDNode that returns the value of
diff --git a/test/CodeGen/MSP430/Inst16mm.ll b/test/CodeGen/MSP430/Inst16mm.ll
index 510afe3..2337c2c 100644
--- a/test/CodeGen/MSP430/Inst16mm.ll
+++ b/test/CodeGen/MSP430/Inst16mm.ll
@@ -1,4 +1,4 @@
-; RUN: llc -march=msp430 < %s | FileCheck %s
+; RUN: llc -march=msp430 -combiner-alias-analysis < %s | FileCheck %s
target datalayout = "e-p:16:8:8-i8:8:8-i16:8:8-i32:8:8"
target triple = "msp430-generic-generic"
@foo = common global i16 0, align 2
@@ -52,3 +52,18 @@
ret void
}
+define i16 @mov2() nounwind {
+entry:
+ %retval = alloca i16 ; <i16*> [#uses=3]
+ %x = alloca i32, align 2 ; <i32*> [#uses=1]
+ %y = alloca i32, align 2 ; <i32*> [#uses=1]
+ store i16 0, i16* %retval
+ %tmp = load i32* %y ; <i32> [#uses=1]
+ store i32 %tmp, i32* %x
+ store i16 0, i16* %retval
+ %0 = load i16* %retval ; <i16> [#uses=1]
+ ret i16 %0
+; CHECK: mov2:
+; CHECK: mov.w 0(r1), 4(r1)
+; CHECK: mov.w 2(r1), 6(r1)
+}
diff --git a/test/CodeGen/X86/sse3.ll b/test/CodeGen/X86/sse3.ll
index b2af7c9..921161e 100644
--- a/test/CodeGen/X86/sse3.ll
+++ b/test/CodeGen/X86/sse3.ll
@@ -144,10 +144,9 @@
store <4 x float> %tmp13, <4 x float>* %r
ret void
; X64: t9:
-; X64: movsd (%rsi), %xmm0
-; X64: movaps (%rdi), %xmm1
-; X64: movlhps %xmm0, %xmm1
-; X64: movaps %xmm1, (%rdi)
+; X64: movaps (%rdi), %xmm0
+; X64: movhps (%rsi), %xmm0
+; X64: movaps %xmm0, (%rdi)
; X64: ret
}
diff --git a/test/CodeGen/X86/store_op_load_fold2.ll b/test/CodeGen/X86/store_op_load_fold2.ll
index e862830..46e59e9 100644
--- a/test/CodeGen/X86/store_op_load_fold2.ll
+++ b/test/CodeGen/X86/store_op_load_fold2.ll
@@ -1,5 +1,4 @@
-; RUN: llc < %s -march=x86 -x86-asm-syntax=intel | \
-; RUN: grep {and DWORD PTR} | count 2
+; RUN: llc < %s -march=x86 -x86-asm-syntax=intel | FileCheck %s
target datalayout = "e-p:32:32"
%struct.Macroblock = type { i32, i32, i32, i32, i32, [8 x i32], %struct.Macroblock*, %struct.Macroblock*, i32, [2 x [4 x [4 x [2 x i32]]]], [16 x i8], [16 x i8], i32, i64, [4 x i32], [4 x i32], i64, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i16, double, i32, i32, i32, i32, i32, i32, i32, i32, i32 }
@@ -16,5 +15,10 @@
%tmp2676.us.us = and i64 %tmp2667.us.us, %tmp2675not.us.us ; <i64> [#uses=1]
store i64 %tmp2676.us.us, i64* %tmp2666
ret i32 0
+
+; CHECK: and {{E..}}, DWORD PTR [360]
+; CHECK: and DWORD PTR [356], {{E..}}
+; CHECK: mov DWORD PTR [360], {{E..}}
+
}