When x86 addresses matching exceeds its recursion limit, check to

see if the base register is already occupied before assuming it can be
used. This fixes bogus code generation in the accompanying testcase.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41049 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86ISelDAGToDAG.cpp b/lib/Target/X86/X86ISelDAGToDAG.cpp
index f0331b83a1a9f53488d8a1c80ecd5d7a2604a517..812d2ecbd2d9f1bec555a49c7d8e97f47d219c07 100644 (file)
--- a/lib/Target/X86/X86ISelDAGToDAG.cpp
+++ b/lib/Target/X86/X86ISelDAGToDAG.cpp
@@ -142,6 +142,8 @@ namespace {
 
     bool MatchAddress(SDOperand N, X86ISelAddressMode &AM,
                       bool isRoot = true, unsigned Depth = 0);
+    bool MatchAddressBase(SDOperand N, X86ISelAddressMode &AM,
+                          bool isRoot, unsigned Depth);
     bool SelectAddr(SDOperand Op, SDOperand N, SDOperand &Base,
                     SDOperand &Scale, SDOperand &Index, SDOperand &Disp);
     bool SelectLEAAddr(SDOperand Op, SDOperand N, SDOperand &Base,
@@ -572,12 +574,9 @@ void X86DAGToDAGISel::EmitFunctionEntryCode(Function &Fn, MachineFunction &MF) {
 /// addressing mode
 bool X86DAGToDAGISel::MatchAddress(SDOperand N, X86ISelAddressMode &AM,
                                    bool isRoot, unsigned Depth) {
-  if (Depth > 5) {
-    // Default, generate it as a register.
-    AM.BaseType = X86ISelAddressMode::RegBase;
-    AM.Base.Reg = N;
-    return false;
-  }
+  // Limit recursion.
+  if (Depth > 5)
+    return MatchAddressBase(N, AM, isRoot, Depth);
   
   // RIP relative addressing: %rip + 32-bit displacement!
   if (AM.isRIPRel) {
@@ -763,6 +762,13 @@ bool X86DAGToDAGISel::MatchAddress(SDOperand N, X86ISelAddressMode &AM,
     break;
   }
 
+  return MatchAddressBase(N, AM, isRoot, Depth);
+}
+
+/// MatchAddressBase - Helper for MatchAddress. Add the specified node to the
+/// specified addressing mode without any further recursion.
+bool X86DAGToDAGISel::MatchAddressBase(SDOperand N, X86ISelAddressMode &AM,
+                                       bool isRoot, unsigned Depth) {
   // Is the base register already occupied?
   if (AM.BaseType != X86ISelAddressMode::RegBase || AM.Base.Reg.Val) {
     // If so, check to see if the scale index register is set.

diff --git a/test/CodeGen/X86/lea-recursion.ll b/test/CodeGen/X86/lea-recursion.ll
new file mode 100644 (file)
index 0000000..ca7b367
--- /dev/null
+++ b/test/CodeGen/X86/lea-recursion.ll
@@ -0,0 +1,47 @@
+; RUN: llvm-as < %s | llc -march=x86-64 | grep lea | wc -l | grep 12
+
+; This testcase was written to demonstrate an instruction-selection problem,
+; however it also happens to expose a limitation in the DAGCombiner's
+; expression reassociation which causes it to miss opportunities for
+; constant folding due to the intermediate adds having multiple uses.
+; The Reassociate pass has similar limitations. If these limitations are
+; fixed, the test commands above will need to be updated to expect fewer
+; lea instructions.
+
+@g0 = weak global [1000 x i32] zeroinitializer, align 32               ; <[1000 x i32]*> [#uses=8]
+@g1 = weak global [1000 x i32] zeroinitializer, align 32               ; <[1000 x i32]*> [#uses=7]
+
+define void @foo() {
+entry:
+       %tmp4 = load i32* getelementptr ([1000 x i32]* @g0, i32 0, i32 0)               ; <i32> [#uses=1]
+       %tmp8 = load i32* getelementptr ([1000 x i32]* @g1, i32 0, i32 0)               ; <i32> [#uses=1]
+       %tmp9 = add i32 %tmp4, 1                ; <i32> [#uses=1]
+       %tmp10 = add i32 %tmp9, %tmp8           ; <i32> [#uses=2]
+       store i32 %tmp10, i32* getelementptr ([1000 x i32]* @g0, i32 0, i32 1)
+       %tmp8.1 = load i32* getelementptr ([1000 x i32]* @g1, i32 0, i32 1)             ; <i32> [#uses=1]
+       %tmp9.1 = add i32 %tmp10, 1             ; <i32> [#uses=1]
+       %tmp10.1 = add i32 %tmp9.1, %tmp8.1             ; <i32> [#uses=2]
+       store i32 %tmp10.1, i32* getelementptr ([1000 x i32]* @g0, i32 0, i32 2)
+       %tmp8.2 = load i32* getelementptr ([1000 x i32]* @g1, i32 0, i32 2)             ; <i32> [#uses=1]
+       %tmp9.2 = add i32 %tmp10.1, 1           ; <i32> [#uses=1]
+       %tmp10.2 = add i32 %tmp9.2, %tmp8.2             ; <i32> [#uses=2]
+       store i32 %tmp10.2, i32* getelementptr ([1000 x i32]* @g0, i32 0, i32 3)
+       %tmp8.3 = load i32* getelementptr ([1000 x i32]* @g1, i32 0, i32 3)             ; <i32> [#uses=1]
+       %tmp9.3 = add i32 %tmp10.2, 1           ; <i32> [#uses=1]
+       %tmp10.3 = add i32 %tmp9.3, %tmp8.3             ; <i32> [#uses=2]
+       store i32 %tmp10.3, i32* getelementptr ([1000 x i32]* @g0, i32 0, i32 4)
+       %tmp8.4 = load i32* getelementptr ([1000 x i32]* @g1, i32 0, i32 4)             ; <i32> [#uses=1]
+       %tmp9.4 = add i32 %tmp10.3, 1           ; <i32> [#uses=1]
+       %tmp10.4 = add i32 %tmp9.4, %tmp8.4             ; <i32> [#uses=2]
+       store i32 %tmp10.4, i32* getelementptr ([1000 x i32]* @g0, i32 0, i32 5)
+       %tmp8.5 = load i32* getelementptr ([1000 x i32]* @g1, i32 0, i32 5)             ; <i32> [#uses=1]
+       %tmp9.5 = add i32 %tmp10.4, 1           ; <i32> [#uses=1]
+       %tmp10.5 = add i32 %tmp9.5, %tmp8.5             ; <i32> [#uses=2]
+       store i32 %tmp10.5, i32* getelementptr ([1000 x i32]* @g0, i32 0, i32 6)
+       %tmp8.6 = load i32* getelementptr ([1000 x i32]* @g1, i32 0, i32 6)             ; <i32> [#uses=1]
+       %tmp9.6 = add i32 %tmp10.5, 1           ; <i32> [#uses=1]
+       %tmp10.6 = add i32 %tmp9.6, %tmp8.6             ; <i32> [#uses=1]
+       store i32 %tmp10.6, i32* getelementptr ([1000 x i32]* @g0, i32 0, i32 7)
+       ret void
+}
+