X86: When lowering shl_parts, don't emit shift amounts larger than the bit width.

While it's safe for the X86-specific shift nodes, dag combining will
kill generic nodes. Insert an AND to make it safe, isel will nuke it
as x86's shift instructions have an implicit AND.

Fixes PR16108, which contains a contraption to hit this case in between
constant folders.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@197228 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index e08adc2384ba15a900bc5fb0f8710d2db4dda825..538a116902e3b4118e4d5dc9cf924929815bc866 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -8459,6 +8459,11 @@ SDValue X86TargetLowering::LowerShiftParts(SDValue Op, SelectionDAG &DAG) const{
   SDValue ShOpLo = Op.getOperand(0);
   SDValue ShOpHi = Op.getOperand(1);
   SDValue ShAmt  = Op.getOperand(2);
+  // X86ISD::SHLD and X86ISD::SHRD have defined overflow behavior but the
+  // generic ISD nodes haven't. Insert an AND to be safe, it's optimized away
+  // during isel.
+  SDValue SafeShAmt = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
+                                  DAG.getConstant(VTBits - 1, MVT::i8));
   SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
                                      DAG.getConstant(VTBits - 1, MVT::i8))
                        : DAG.getConstant(0, VT);
@@ -8466,12 +8471,15 @@ SDValue X86TargetLowering::LowerShiftParts(SDValue Op, SelectionDAG &DAG) const{
   SDValue Tmp2, Tmp3;
   if (Op.getOpcode() == ISD::SHL_PARTS) {
     Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
-    Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
+    Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, SafeShAmt);
   } else {
     Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
-    Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
+    Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, SafeShAmt);
   }
 
+  // If the shift amount is larger or equal than the width of a part we can't
+  // rely on the results of shld/shrd. Insert a test and select the appropriate
+  // values for large shift amounts.
   SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
                                 DAG.getConstant(VTBits, MVT::i8));
   SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,

diff --git a/test/CodeGen/X86/legalize-shift-64.ll b/test/CodeGen/X86/legalize-shift-64.ll
index 77364688afe4391360e92c8163a8beec46123249..64460bb91186c6bc5403b226b7a5aeabf68cc0c8 100644 (file)
--- a/test/CodeGen/X86/legalize-shift-64.ll
+++ b/test/CodeGen/X86/legalize-shift-64.ll
@@ -64,3 +64,31 @@ define <2 x i64> @test5(<2 x i64> %A, <2 x i64> %B) {
 ; CHECK: shl
 ; CHECK: shldl
 }
+
+; PR16108
+define i32 @test6() {
+  %x = alloca i32, align 4
+  %t = alloca i64, align 8
+  store i32 1, i32* %x, align 4
+  store i64 1, i64* %t, align 8  ;; DEAD
+  %load = load i32* %x, align 4
+  %shl = shl i32 %load, 8
+  %add = add i32 %shl, -224
+  %sh_prom = zext i32 %add to i64
+  %shl1 = shl i64 1, %sh_prom
+  %cmp = icmp ne i64 %shl1, 4294967296
+  br i1 %cmp, label %if.then, label %if.end
+
+if.then:                                          ; preds = %entry
+  ret i32 1
+
+if.end:                                           ; preds = %entry
+  ret i32 0
+
+; CHECK-LABEL: test6:
+; CHECK-NOT: andb $31
+; CHECK: sete
+; CHECK: movzbl
+; CHECK: xorl $1
+; CHECK: orl
+}