[llvm] r243100 - When lowering vector shifts a check is performed to see if the value to shift by

[Luke Cheeseman]

Author: lukecheeseman
Date: Fri Jul 24 04:31:48 2015
New Revision: 243100

URL: http://llvm.org/viewvc/llvm-project?rev=243100&view=rev
Log:
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 causing failures. 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.

Differential Revision: http://reviews.llvm.org/D11408


Added:
    llvm/trunk/test/CodeGen/ARM/neon_vshl_minint.ll
Modified:
    llvm/trunk/lib/Target/AArch64/AArch64ISelLowering.cpp
    llvm/trunk/lib/Target/ARM/ARMISelLowering.cpp

Modified: llvm/trunk/lib/Target/AArch64/AArch64ISelLowering.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/AArch64/AArch64ISelLowering.cpp?rev=243100&r1=243099&r2=243100&view=diff
==============================================================================
--- llvm/trunk/lib/Target/AArch64/AArch64ISelLowering.cpp (original)
+++ llvm/trunk/lib/Target/AArch64/AArch64ISelLowering.cpp Fri Jul 24 04:31:48 2015
@@ -6440,26 +6440,20 @@ static bool getVShiftImm(SDValue Op, uns
 ///   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));
 }
 
@@ -6488,8 +6482,7 @@ SDValue AArch64TargetLowering::LowerVect
   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),

Modified: llvm/trunk/lib/Target/ARM/ARMISelLowering.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/ARM/ARMISelLowering.cpp?rev=243100&r1=243099&r2=243100&view=diff
==============================================================================
--- llvm/trunk/lib/Target/ARM/ARMISelLowering.cpp (original)
+++ llvm/trunk/lib/Target/ARM/ARMISelLowering.cpp Fri Jul 24 04:31:48 2015
@@ -9691,7 +9691,7 @@ static bool getVShiftImm(SDValue Op, uns
 ///   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);
@@ -9706,12 +9706,16 @@ static bool isVShiftLImm(SDValue Op, EVT
 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)
+  if (!isIntrinsic)
+    return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
+  if (Cnt >= -(isNarrow ? ElementBits/2 : ElementBits) && Cnt <= -1) {
     Cnt = -Cnt;
-  return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
+    return true;
+  }
+  return false;
 }
 
 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.

Added: llvm/trunk/test/CodeGen/ARM/neon_vshl_minint.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/ARM/neon_vshl_minint.ll?rev=243100&view=auto
==============================================================================
--- llvm/trunk/test/CodeGen/ARM/neon_vshl_minint.ll (added)
+++ llvm/trunk/test/CodeGen/ARM/neon_vshl_minint.ll Fri Jul 24 04:31:48 2015
@@ -0,0 +1,13 @@
+; 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>)