[llvm] r245544 - [X86] Fix the (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2)) fold

[David Majnemer via llvm-commits]

Author: majnemer
Date: Thu Aug 20 04:00:56 2015
New Revision: 245544

URL: http://llvm.org/viewvc/llvm-project?rev=245544&view=rev
Log:
[X86] Fix the (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2)) fold

We didn't check for the necessary preconditions before folding a
mask/shift into a single mask.

This fixes PR24516.

Modified:
    llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
    llvm/trunk/test/CodeGen/X86/setcc.ll

Modified: llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86ISelLowering.cpp?rev=245544&r1=245543&r2=245544&view=diff
==============================================================================
--- llvm/trunk/lib/Target/X86/X86ISelLowering.cpp (original)
+++ llvm/trunk/lib/Target/X86/X86ISelLowering.cpp Thu Aug 20 04:00:56 2015
@@ -23435,18 +23435,34 @@ static SDValue PerformSHLCombine(SDNode
       N1C && N0.getOpcode() == ISD::AND &&
       N0.getOperand(1).getOpcode() == ISD::Constant) {
     SDValue N00 = N0.getOperand(0);
-    if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
-        ((N00.getOpcode() == ISD::ANY_EXTEND ||
-          N00.getOpcode() == ISD::ZERO_EXTEND) &&
-         N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
-      APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
-      APInt ShAmt = N1C->getAPIntValue();
-      Mask = Mask.shl(ShAmt);
-      if (Mask != 0) {
-        SDLoc DL(N);
-        return DAG.getNode(ISD::AND, DL, VT,
-                           N00, DAG.getConstant(Mask, DL, VT));
-      }
+    APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
+    APInt ShAmt = N1C->getAPIntValue();
+    Mask = Mask.shl(ShAmt);
+    bool MaskOK = false;
+    // We can handle cases concerning bit-widening nodes containing setcc_c if
+    // we carefully interrogate the mask to make sure we are semantics
+    // preserving.
+    // The transform is not safe if the result of C1 << C2 exceeds the bitwidth
+    // of the underlying setcc_c operation if the setcc_c was zero extended.
+    // Consider the following example:
+    //   zext(setcc_c)                 -> i32 0x0000FFFF
+    //   c1                            -> i32 0x0000FFFF
+    //   c2                            -> i32 0x00000001
+    //   (shl (and (setcc_c), c1), c2) -> i32 0x0001FFFE
+    //   (and setcc_c, (c1 << c2))     -> i32 0x0000FFFE
+    if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
+      MaskOK = true;
+    } else if (N00.getOpcode() == ISD::SIGN_EXTEND &&
+               N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
+      MaskOK = true;
+    } else if ((N00.getOpcode() == ISD::ZERO_EXTEND ||
+                N00.getOpcode() == ISD::ANY_EXTEND) &&
+               N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
+      MaskOK = Mask.isIntN(N00.getOperand(0).getValueSizeInBits());
+    }
+    if (MaskOK && Mask != 0) {
+      SDLoc DL(N);
+      return DAG.getNode(ISD::AND, DL, VT, N00, DAG.getConstant(Mask, DL, VT));
     }
   }
 

Modified: llvm/trunk/test/CodeGen/X86/setcc.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/setcc.ll?rev=245544&r1=245543&r2=245544&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/X86/setcc.ll (original)
+++ llvm/trunk/test/CodeGen/X86/setcc.ll Thu Aug 20 04:00:56 2015
@@ -34,3 +34,23 @@ entry:
   %iftmp.2.0 = select i1 %0, i64 64, i64 0        ; <i64> [#uses=1]
   ret i64 %iftmp.2.0
 }
+
+ at v4 = common global i32 0, align 4
+
+define i32 @t4(i32 %a) {
+entry:
+; CHECK-LABEL: t4:
+; CHECK:  movq    _v4 at GOTPCREL(%rip), %rax
+; CHECK:  cmpl    $1, (%rax)
+; CHECK:  sbbl    %eax, %eax
+; CHECK:  andl    $32768, %eax
+; CHECK:  leal    65536(%rax,%rax), %eax
+  %0 = load i32, i32* @v4, align 4
+  %not.tobool = icmp eq i32 %0, 0
+  %conv.i = sext i1 %not.tobool to i16
+  %call.lobit = lshr i16 %conv.i, 15
+  %add.i.1 = add nuw nsw i16 %call.lobit, 1
+  %conv4.2 = zext i16 %add.i.1 to i32
+  %add = shl nuw nsw i32 %conv4.2, 16
+  ret i32 %add
+}