[PATCH] D45203: [X86] VRNDSCALE* folding from masked and single-value ffloor and fceil patterns

[Mikhail Dvoretckii via Phabricator via llvm-commits]

mike.dvoretsky created this revision.
mike.dvoretsky added reviewers: craig.topper, spatel, RKSimon.
Herald added subscribers: llvm-commits, hiraditya.

This patch handles back-end folding of generic patterns created by https://reviews.llvm.org/D45202 in cases where the instruction isn't a straightforward packed values rounding operation, but a masked operation or a single-value operation.


Repository:
  rL LLVM

https://reviews.llvm.org/D45203

Files:
  llvm/lib/Target/X86/X86ISelLowering.cpp
  llvm/lib/Target/X86/X86InstrAVX512.td


Index: llvm/lib/Target/X86/X86InstrAVX512.td
===================================================================
--- llvm/lib/Target/X86/X86InstrAVX512.td
+++ llvm/lib/Target/X86/X86InstrAVX512.td
@@ -9370,21 +9370,29 @@
 let Predicates = [HasAVX512] in {
 def : Pat<(v16f32 (ffloor VR512:$src)),
           (VRNDSCALEPSZrri VR512:$src, (i32 0x9))>;
+def : Pat<(v16f32 (vselect v16f32_info.KRCWM:$mask, (ffloor VR512:$src), VR512:$dst)),
+          (VRNDSCALEPSZrrik VR512:$dst, v16f32_info.KRCWM:$mask, VR512:$src, (i32 0x9))>;
 def : Pat<(v16f32 (fnearbyint VR512:$src)),
           (VRNDSCALEPSZrri VR512:$src, (i32 0xC))>;
 def : Pat<(v16f32 (fceil VR512:$src)),
           (VRNDSCALEPSZrri VR512:$src, (i32 0xA))>;
+def : Pat<(v16f32 (vselect v16f32_info.KRCWM:$mask, (fceil VR512:$src), VR512:$dst)),
+          (VRNDSCALEPSZrrik VR512:$dst, v16f32_info.KRCWM:$mask, VR512:$src, (i32 0xA))>;
 def : Pat<(v16f32 (frint VR512:$src)),
           (VRNDSCALEPSZrri VR512:$src, (i32 0x4))>;
 def : Pat<(v16f32 (ftrunc VR512:$src)),
           (VRNDSCALEPSZrri VR512:$src, (i32 0xB))>;
 
 def : Pat<(v8f64 (ffloor VR512:$src)),
           (VRNDSCALEPDZrri VR512:$src, (i32 0x9))>;
+def : Pat<(v8f64 (vselect v8f64_info.KRCWM:$mask, (ffloor VR512:$src), VR512:$dst)),
+          (VRNDSCALEPDZrrik VR512:$dst, v8f64_info.KRCWM:$mask, VR512:$src, (i32 0x9))>;
 def : Pat<(v8f64 (fnearbyint VR512:$src)),
           (VRNDSCALEPDZrri VR512:$src, (i32 0xC))>;
 def : Pat<(v8f64 (fceil VR512:$src)),
           (VRNDSCALEPDZrri VR512:$src, (i32 0xA))>;
+def : Pat<(v8f64 (vselect v8f64_info.KRCWM:$mask, (fceil VR512:$src), VR512:$dst)),
+          (VRNDSCALEPDZrrik VR512:$dst, v8f64_info.KRCWM:$mask, VR512:$src, (i32 0xA))>;
 def : Pat<(v8f64 (frint VR512:$src)),
           (VRNDSCALEPDZrri VR512:$src, (i32 0x4))>;
 def : Pat<(v8f64 (ftrunc VR512:$src)),
Index: llvm/lib/Target/X86/X86ISelLowering.cpp
===================================================================
--- llvm/lib/Target/X86/X86ISelLowering.cpp
+++ llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -30480,6 +30480,37 @@
   return DAG.getNode(X86ISD::ADDSUB, DL, VT, Opnd0, Opnd1);
 }
 
+// For SSE4.1 and AVX512, we may want to combine VRNDSCALES from
+// vector_shuffle<{0,3}|{0,5,6,7}> (fceil|ffloor A), B
+// patterns.
+static SDValue combineShuffleFloorCeil(SDNode *N, SelectionDAG &DAG,
+                                       const X86Subtarget &Subtarget) {
+  if (!isa<ShuffleVectorSDNode>(N))
+    return SDValue();
+  EVT VT = N->getValueType(0);
+  unsigned Num = VT.getVectorNumElements();
+  if (Num * VT.getScalarSizeInBits() != 128 || !Subtarget.hasSSE41())
+    return SDValue();
+  SDValue N0 = N->getOperand(0);
+  SDValue N1 = N->getOperand(1);
+  int Op = N0.getOpcode();
+  if ((Num != 2 && Num != 4) || (Op != ISD::FCEIL && Op != ISD::FFLOOR))
+    return SDValue();
+
+  // The mask being matched here is equivalent to a 0...01 select mask.
+  ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
+  if (SVOp->getMaskElt(0) != 0)
+    return SDValue();
+  for (unsigned i = 1; i < Num; ++i)
+    if (SVOp->getMaskElt(i) != Num + i)
+      return SDValue();
+
+  int Imm = (Op == ISD::FCEIL) ? 2 : 1;
+  SDLoc DL(N);
+  return DAG.getNode(X86ISD::VRNDSCALES, DL, VT, N1, N0.getOperand(0),
+                     DAG.getConstant(Imm, DL, MVT::i32));
+}
+
 // We are looking for a shuffle where both sources are concatenated with undef
 // and have a width that is half of the output's width. AVX2 has VPERMD/Q, so
 // if we can express this as a single-source shuffle, that's preferable.
@@ -30653,6 +30684,9 @@
             EltsFromConsecutiveLoads(VT, Elts, dl, DAG, Subtarget, true))
       return LD;
 
+  if (SDValue RndScale = combineShuffleFloorCeil(N, DAG, Subtarget))
+    return RndScale;
+
   // For AVX2, we sometimes want to combine
   // (vector_shuffle <mask> (concat_vectors t1, undef)
   //                        (concat_vectors t2, undef))


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D45203.140746.patch
Type: text/x-patch
Size: 3944 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20180403/27938f8d/attachment.bin>