From cc5fc44d0203c25729c631e273a45efdbb7b5d98 Mon Sep 17 00:00:00 2001
From: Dale Johannesen <dalej@apple.com>
Date: Wed, 11 Feb 2009 19:19:41 +0000
Subject: [PATCH] Make a transformation added in 63266 a bit less aggressive.
 It was transforming (x&y)==y to (x&y)!=0 in the case where y is variable and
 known to have at most one bit set (e.g. z&1). This is not correct; the
 expressions are not equivalent when y==0. I believe this patch salvages what
 can be salvaged, including all the cases in bt.ll.  Dan, please review. Fixes
 gcc.c-torture/execute/20040709-[12].c

llvm-svn: 64314
---
 .../CodeGen/SelectionDAG/TargetLowering.cpp   | 14 ++--
 llvm/test/CodeGen/X86/nobt.ll                 | 70 +++++++++++++++++++
 2 files changed, 79 insertions(+), 5 deletions(-)
 create mode 100644 llvm/test/CodeGen/X86/nobt.ll

diff --git a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
index d7008923c159..fce57f93b7f8 100644
--- a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
@@ -1346,20 +1346,21 @@ unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
   return 1;
 }
 
-static bool ValueHasAtMostOneBitSet(SDValue Val, const SelectionDAG &DAG) {
+static bool ValueHasExactlyOneBitSet(SDValue Val, const SelectionDAG &DAG) {
   // Logical shift right or left won't ever introduce new set bits.
   // We check for this case because we don't care which bits are
   // set, but ComputeMaskedBits won't know anything unless it can
   // determine which specific bits may be set.
   if (Val.getOpcode() == ISD::SHL || Val.getOpcode() == ISD::SRL)
-    return ValueHasAtMostOneBitSet(Val.getOperand(0), DAG);
+    return ValueHasExactlyOneBitSet(Val.getOperand(0), DAG);
 
   MVT OpVT = Val.getValueType();
   unsigned BitWidth = OpVT.getSizeInBits();
   APInt Mask = APInt::getAllOnesValue(BitWidth);
   APInt KnownZero, KnownOne;
   DAG.ComputeMaskedBits(Val, Mask, KnownZero, KnownOne);
-  return KnownZero.countPopulation() == BitWidth - 1;
+  return (KnownZero.countPopulation() == BitWidth - 1) &&
+         (KnownOne.countPopulation() == 1);
 }
 
 /// SimplifySetCC - Try to simplify a setcc built with the specified operands 
@@ -1832,9 +1833,12 @@ TargetLowering::SimplifySetCC(MVT VT, SDValue N0, SDValue N1,
     }
 
     // Simplify x&y == y to x&y != 0 if y has exactly one bit set.
+    // Note that where y is variable and is known to have at most
+    // one bit set (for example, if it is z&1) we cannot do this;
+    // the expressions are not equivalent when y==0.
     if (N0.getOpcode() == ISD::AND)
       if (N0.getOperand(0) == N1 || N0.getOperand(1) == N1) {
-        if (ValueHasAtMostOneBitSet(N1, DAG)) {
+        if (ValueHasExactlyOneBitSet(N1, DAG)) {
           Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true);
           SDValue Zero = DAG.getConstant(0, N1.getValueType());
           return DAG.getSetCC(dl, VT, N0, Zero, Cond);
@@ -1842,7 +1846,7 @@ TargetLowering::SimplifySetCC(MVT VT, SDValue N0, SDValue N1,
       }
     if (N1.getOpcode() == ISD::AND)
       if (N1.getOperand(0) == N0 || N1.getOperand(1) == N0) {
-        if (ValueHasAtMostOneBitSet(N0, DAG)) {
+        if (ValueHasExactlyOneBitSet(N0, DAG)) {
           Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true);
           SDValue Zero = DAG.getConstant(0, N0.getValueType());
           return DAG.getSetCC(dl, VT, N1, Zero, Cond);
diff --git a/llvm/test/CodeGen/X86/nobt.ll b/llvm/test/CodeGen/X86/nobt.ll
new file mode 100644
index 000000000000..55294280f5c8
--- /dev/null
+++ b/llvm/test/CodeGen/X86/nobt.ll
@@ -0,0 +1,70 @@
+; RUN: llvm-as < %s | llc -march=x86 | not grep btl
+
+; This tests some cases where BT must not be generated.  See also bt.ll.
+; Fixes 20040709-[12].c in gcc testsuite.
+
+define void @test2(i32 %x, i32 %n) nounwind {
+entry:
+        %tmp1 = and i32 %x, 1
+        %tmp2 = urem i32 %tmp1, 15
+	%tmp3 = and i32 %tmp2, 1		; <i32> [#uses=1]
+	%tmp4 = icmp eq i32 %tmp3, %tmp2	; <i1> [#uses=1]
+	br i1 %tmp4, label %bb, label %UnifiedReturnBlock
+
+bb:		; preds = %entry
+	call void @foo()
+	ret void
+
+UnifiedReturnBlock:		; preds = %entry
+	ret void
+}
+
+define void @test3(i32 %x, i32 %n) nounwind {
+entry:
+        %tmp1 = and i32 %x, 1
+        %tmp2 = urem i32 %tmp1, 15
+	%tmp3 = and i32 %tmp2, 1		; <i32> [#uses=1]
+	%tmp4 = icmp eq i32 %tmp2, %tmp3	; <i1> [#uses=1]
+	br i1 %tmp4, label %bb, label %UnifiedReturnBlock
+
+bb:		; preds = %entry
+	call void @foo()
+	ret void
+
+UnifiedReturnBlock:		; preds = %entry
+	ret void
+}
+
+define void @test4(i32 %x, i32 %n) nounwind {
+entry:
+        %tmp1 = and i32 %x, 1
+        %tmp2 = urem i32 %tmp1, 15
+	%tmp3 = and i32 %tmp2, 1		; <i32> [#uses=1]
+	%tmp4 = icmp ne i32 %tmp2, %tmp3	; <i1> [#uses=1]
+	br i1 %tmp4, label %bb, label %UnifiedReturnBlock
+
+bb:		; preds = %entry
+	call void @foo()
+	ret void
+
+UnifiedReturnBlock:		; preds = %entry
+	ret void
+}
+
+define void @test5(i32 %x, i32 %n) nounwind {
+entry:
+        %tmp1 = and i32 %x, 1
+        %tmp2 = urem i32 %tmp1, 15
+	%tmp3 = and i32 %tmp2, 1		; <i32> [#uses=1]
+	%tmp4 = icmp ne i32 %tmp2, %tmp3	; <i1> [#uses=1]
+	br i1 %tmp4, label %bb, label %UnifiedReturnBlock
+
+bb:		; preds = %entry
+	call void @foo()
+	ret void
+
+UnifiedReturnBlock:		; preds = %entry
+	ret void
+}
+
+declare void @foo()