Add code for lowering v32i8 shifts by a splat to AVX2 immediate shift instructions. Remove 256-bit splat handling from LowerShift as it was already handled by PerformShiftCombine.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@145005 91177308-0d34-0410-b5e6-96231b3b80d8

2 files changed, 45 insertions, 45 deletions

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 61d9246..470a115 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -10338,47 +10338,48 @@ SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
         return Res;
       }
 
-      if (Subtarget->hasAVX2()) {
-        if (VT == MVT::v4i64 && Op.getOpcode() == ISD::SHL)
-         return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
-                       DAG.getConstant(Intrinsic::x86_avx2_pslli_q, MVT::i32),
-                       R, DAG.getConstant(ShiftAmt, MVT::i32));
-
-        if (VT == MVT::v8i32 && Op.getOpcode() == ISD::SHL)
-         return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
-                       DAG.getConstant(Intrinsic::x86_avx2_pslli_d, MVT::i32),
-                       R, DAG.getConstant(ShiftAmt, MVT::i32));
-
-        if (VT == MVT::v16i16 && Op.getOpcode() == ISD::SHL)
-         return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
-                       DAG.getConstant(Intrinsic::x86_avx2_pslli_w, MVT::i32),
-                       R, DAG.getConstant(ShiftAmt, MVT::i32));
-
-        if (VT == MVT::v4i64 && Op.getOpcode() == ISD::SRL)
-         return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
-                       DAG.getConstant(Intrinsic::x86_avx2_psrli_q, MVT::i32),
-                       R, DAG.getConstant(ShiftAmt, MVT::i32));
-
-        if (VT == MVT::v8i32 && Op.getOpcode() == ISD::SRL)
-         return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
-                       DAG.getConstant(Intrinsic::x86_avx2_psrli_d, MVT::i32),
-                       R, DAG.getConstant(ShiftAmt, MVT::i32));
-
-        if (VT == MVT::v16i16 && Op.getOpcode() == ISD::SRL)
-         return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
-                       DAG.getConstant(Intrinsic::x86_avx2_psrli_w, MVT::i32),
-                       R, DAG.getConstant(ShiftAmt, MVT::i32));
-
-        if (VT == MVT::v8i32 && Op.getOpcode() == ISD::SRA)
-         return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
-                       DAG.getConstant(Intrinsic::x86_avx2_psrai_d, MVT::i32),
-                       R, DAG.getConstant(ShiftAmt, MVT::i32));
-
-        if (VT == MVT::v16i16 && Op.getOpcode() == ISD::SRA)
-         return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
-                       DAG.getConstant(Intrinsic::x86_avx2_psrai_w, MVT::i32),
-                       R, DAG.getConstant(ShiftAmt, MVT::i32));
+      if (Subtarget->hasAVX2() && VT == MVT::v32i8) {
+        if (Op.getOpcode() == ISD::SHL) {
+          // Make a large shift.
+          SDValue SHL =
+            DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
+                        DAG.getConstant(Intrinsic::x86_avx2_pslli_w, MVT::i32),
+                        R, DAG.getConstant(ShiftAmt, MVT::i32));
+          // Zero out the rightmost bits.
+          SmallVector<SDValue, 32> V(32, DAG.getConstant(uint8_t(-1U << ShiftAmt),
+                                                         MVT::i8));
+          return DAG.getNode(ISD::AND, dl, VT, SHL,
+                             DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &V[0], 32));
         }
+        if (Op.getOpcode() == ISD::SRL) {
+          // Make a large shift.
+          SDValue SRL =
+            DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
+                        DAG.getConstant(Intrinsic::x86_avx2_psrli_w, MVT::i32),
+                        R, DAG.getConstant(ShiftAmt, MVT::i32));
+          // Zero out the leftmost bits.
+          SmallVector<SDValue, 32> V(32, DAG.getConstant(uint8_t(-1U) >> ShiftAmt,
+                                                         MVT::i8));
+          return DAG.getNode(ISD::AND, dl, VT, SRL,
+                             DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &V[0], 32));
+        }
+        if (Op.getOpcode() == ISD::SRA) {
+          if (ShiftAmt == 7) {
+            // R s>> 7  ===  R s< 0
+            SDValue Zeros = getZeroVector(VT, true /* HasXMMInt */, DAG, dl);
+            return DAG.getNode(X86ISD::PCMPGTB, dl, VT, Zeros, R);
+          }
+
+          // R s>> a === ((R u>> a) ^ m) - m
+          SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
+          SmallVector<SDValue, 32> V(32, DAG.getConstant(128 >> ShiftAmt,
+                                                         MVT::i8));
+          SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &V[0], 32);
+          Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
+          Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
+          return Res;
+        }
+      }
     }
   }
 
diff --git a/lib/Target/X86/X86InstrSSE.td b/lib/Target/X86/X86InstrSSE.td
index 1e74da1..68010b7 100644
--- a/lib/Target/X86/X86InstrSSE.td
+++ b/lib/Target/X86/X86InstrSSE.td
@@ -311,17 +311,16 @@ def : Pat<(bc_v4i64 (v8f32 immAllZerosV)),
 // JIT implementation, it does not expand the instructions below like
 // X86MCInstLower does.
 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
-    isCodeGenOnly = 1, ExeDomain = SSEPackedInt in
+    isCodeGenOnly = 1, ExeDomain = SSEPackedInt in {
   def V_SETALLONES : PDI<0x76, MRMInitReg, (outs VR128:$dst), (ins), "",
                          [(set VR128:$dst, (v4i32 immAllOnesV))]>;
-let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
-    isCodeGenOnly = 1, ExeDomain = SSEPackedInt, Predicates = [HasAVX] in
+  let Predicates = [HasAVX] in
   def AVX_SETALLONES : PDI<0x76, MRMInitReg, (outs VR128:$dst), (ins), "",
                          [(set VR128:$dst, (v4i32 immAllOnesV))]>, VEX_4V;
-let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
-    isCodeGenOnly = 1, ExeDomain = SSEPackedInt, Predicates = [HasAVX2] in
+  let Predicates = [HasAVX2] in
   def AVX2_SETALLONES : PDI<0x76, MRMInitReg, (outs VR256:$dst), (ins), "",
                           [(set VR256:$dst, (v8i32 immAllOnesV))]>, VEX_4V;
+}
 
 
 //===----------------------------------------------------------------------===//