[ARM/AArch64] - Remove some undefined behaviour when lowering vector shifts

[LukeCheeseman]

Summary

When lowering vector shifts a check is performed to see if the value to shift by is an immediate, in this check the value is negated and stored in and int64_t. The value can be -2^63 yet the result cannot be stored in an int64_t and this gives some undefined behaviour. The regression test included exposes this behaviour, this test fails in the current release build with assertions when built using clang as the check for an immediate passes when it should not due to this undefined behaviour. The negation is only necessary when the values is within a certain range and so it should not need to negate -2^63, this patch introduces this and also a regression test.

[rengolin]

No AArch64 tests?

test/CodeGen/ARM/neon_vshl_minint.ll

2 ↗

(On Diff #30335)

You're not using prefixes ACORE/MCORE?

[LukeCheeseman]

Removed some unnecessary check prefixes in tests.

[LukeCheeseman]

Hi Renato

Thanks for pointing out the unnecessary check prefixes. I haven't added a testcase for the AArch64 backend as it isn't possible to hit the same undefined behaviour as in the ARM backend. The negation is only done in isVShiftRImm when isIntrinsic is true and the only call site in the AArch64 backend passes this as false so the negation isn't performed.

Thanks,
Luke

[rengolin]

In D11408#210737, @LukeCheeseman wrote:

Thanks for pointing out the unnecessary check prefixes. I haven't added a testcase for the AArch64 backend as it isn't possible to hit the same undefined behaviour as in the ARM backend. The negation is only done in isVShiftRImm when isIntrinsic is true and the only call site in the AArch64 backend passes this as false so the negation isn't performed.

So, why not an an assert for now, and point to the same implementation in ARM mode for when it's needed?

--renato

[LukeCheeseman]

Do you mean assert that isIntrinsic is always false?

Thanks,
Luke

[rengolin]

In D11408#210757, @LukeCheeseman wrote:

Do you mean assert that isIntrinsic is always false?

If it's unlikely to ever be used, then yes.

[LukeCheeseman]

Removed isIntrinsic boolean from isVShift{R, L}Imm in AArch64 backend

[rengolin]

LGTM. Thanks!

llvm/trunk/lib/Target/AArch64/AArch64ISelLowering.cpp

 }
 
 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
 /// operand of a vector shift left operation.  That value must be in the range:
 ///   0 <= Value < ElementBits for a left shift; or
 ///   0 <= Value <= ElementBits for a long left shift.
 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
   assert(VT.isVector() && "vector shift count is not a vector type");
-  unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
+  int64_t ElementBits = VT.getVectorElementType().getSizeInBits();
   if (!getVShiftImm(Op, ElementBits, Cnt))
     return false;
   return (Cnt >= 0 && (isLong ? Cnt - 1 : Cnt) < ElementBits);
 }
 
 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
-/// operand of a vector shift right operation.  For a shift opcode, the value
-/// is positive, but for an intrinsic the value count must be negative. The
-/// absolute value must be in the range:
-///   1 <= |Value| <= ElementBits for a right shift; or
-///   1 <= |Value| <= ElementBits/2 for a narrow right shift.
-static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
-                         int64_t &Cnt) {
+/// operand of a vector shift right operation. The value must be in the range:
+///   1 <= Value <= ElementBits for a right shift; or
+static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, int64_t &Cnt) {
   assert(VT.isVector() && "vector shift count is not a vector type");
-  unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
+  int64_t ElementBits = VT.getVectorElementType().getSizeInBits();
   if (!getVShiftImm(Op, ElementBits, Cnt))
     return false;
-  if (isIntrinsic)
-    Cnt = -Cnt;
   return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits / 2 : ElementBits));
 }
 
 SDValue AArch64TargetLowering::LowerVectorSRA_SRL_SHL(SDValue Op,
                                                       SelectionDAG &DAG) const {
   EVT VT = Op.getValueType();
   SDLoc DL(Op);
   int64_t Cnt;
if (isVShiftLImm(Op.getOperand(1), VT, false, Cnt) && Cnt < EltSize)
                          DAG.getConstant(Cnt, DL, MVT::i32));
     return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
                        DAG.getConstant(Intrinsic::aarch64_neon_ushl, DL,
                                        MVT::i32),
                        Op.getOperand(0), Op.getOperand(1));
   case ISD::SRA:
   case ISD::SRL:
     // Right shift immediate
-    if (isVShiftRImm(Op.getOperand(1), VT, false, false, Cnt) &&
-        Cnt < EltSize) {
+    if (isVShiftRImm(Op.getOperand(1), VT, false, Cnt) && Cnt < EltSize) {
       unsigned Opc =
           (Op.getOpcode() == ISD::SRA) ? AArch64ISD::VASHR : AArch64ISD::VLSHR;
       return DAG.getNode(Opc, DL, VT, Op.getOperand(0),
                          DAG.getConstant(Cnt, DL, MVT::i32));
     }
 
     // Right shift register.  Note, there is not a shift right register
     // instruction, but the shift left register instruction takes a signed

llvm/trunk/lib/Target/ARM/ARMISelLowering.cpp

 }
 
 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
 /// operand of a vector shift left operation.  That value must be in the range:
 ///   0 <= Value < ElementBits for a left shift; or
 ///   0 <= Value <= ElementBits for a long left shift.
 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
   assert(VT.isVector() && "vector shift count is not a vector type");
-  unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
+  int64_t ElementBits = VT.getVectorElementType().getSizeInBits();
   if (! getVShiftImm(Op, ElementBits, Cnt))
     return false;
   return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
 }
 
 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
 /// operand of a vector shift right operation.  For a shift opcode, the value
 /// is positive, but for an intrinsic the value count must be negative. The
 /// absolute value must be in the range:
 ///   1 <= |Value| <= ElementBits for a right shift; or
 ///   1 <= |Value| <= ElementBits/2 for a narrow right shift.
 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
                          int64_t &Cnt) {
   assert(VT.isVector() && "vector shift count is not a vector type");
-  unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
+  int64_t ElementBits = VT.getVectorElementType().getSizeInBits();
   if (! getVShiftImm(Op, ElementBits, Cnt))
     return false;
-  if (isIntrinsic)
-    Cnt = -Cnt;
+  if (!isIntrinsic)
     return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
+  if (Cnt >= -(isNarrow ? ElementBits/2 : ElementBits) && Cnt <= -1) {
+    Cnt = -Cnt;
+    return true;
+  }
+  return false;
 }
 
 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
   unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
   switch (IntNo) {
   default:
     // Don't do anything for most intrinsics.

llvm/trunk/test/CodeGen/ARM/neon_vshl_minint.ll

+; RUN: llc < %s -mtriple=arm-none-eabi -mcpu=cortex-a8 2>&1 | FileCheck %s
+; RUN: llc < %s -mtriple=thumb-none-eabi -mcpu=cortex-a8 2>&1 | FileCheck %s
+
+define <1 x i64> @vshl_minint() #0 {
+  entry:
+    ; CHECK-LABEL: vshl_minint
+    ; CHECK: vldr
+    ; CHECK: vshl.u64
+    %vshl.i = tail call <1 x i64> @llvm.arm.neon.vshiftu.v1i64(<1 x i64> undef, <1 x i64> <i64 -9223372036854775808>)
+    ret <1 x i64> %vshl.i
+}
+
+declare <1 x i64> @llvm.arm.neon.vshiftu.v1i64(<1 x i64>, <1 x i64>)