[PATCH] D71984: [X86] Call SimplifyMultipleUseDemandedBits from combineVSelectToBLENDV if the condition is used by something other than select conditions.

[Craig Topper via Phabricator via llvm-commits]

craig.topper created this revision.
craig.topper added reviewers: RKSimon, spatel.
Herald added a subscriber: hiraditya.
Herald added a project: LLVM.

We might be able to bypass some nodes on the condition path.

This interacts with D71979 <https://reviews.llvm.org/D71979> and D71956 <https://reviews.llvm.org/D71956>. I'll rebase whichever gets
in first.


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D71984

Files:
  llvm/lib/Target/X86/X86ISelLowering.cpp
  llvm/test/CodeGen/X86/vselect.ll


Index: llvm/test/CodeGen/X86/vselect.ll
===================================================================
--- llvm/test/CodeGen/X86/vselect.ll
+++ llvm/test/CodeGen/X86/vselect.ll
@@ -536,8 +536,8 @@
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    psllq $63, %xmm0
 ; SSE41-NEXT:    psrad $31, %xmm0
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
 ; SSE41-NEXT:    blendvpd %xmm0, %xmm1, %xmm2
+; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
 ; SSE41-NEXT:    paddq %xmm2, %xmm0
 ; SSE41-NEXT:    retq
 ;
Index: llvm/lib/Target/X86/X86ISelLowering.cpp
===================================================================
--- llvm/lib/Target/X86/X86ISelLowering.cpp
+++ llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -37552,37 +37552,55 @@
   if (VT.is512BitVector())
     return SDValue();
 
-  // TODO: Add other opcodes eventually lowered into BLEND.
-  for (SDNode::use_iterator UI = Cond->use_begin(), UE = Cond->use_end();
-       UI != UE; ++UI)
-    if ((UI->getOpcode() != ISD::VSELECT &&
-         UI->getOpcode() != X86ISD::BLENDV) ||
-        UI.getOperandNo() != 0)
+  auto OnlyUsedAsSelectCond = [](SDValue Cond) {
+    for (SDNode::use_iterator UI = Cond->use_begin(), UE = Cond->use_end();
+         UI != UE; ++UI)
+      if ((UI->getOpcode() != ISD::VSELECT &&
+           UI->getOpcode() != X86ISD::BLENDV) ||
+          UI.getOperandNo() != 0)
+        return false;
+
+    return true;
+  };
+
+  if (OnlyUsedAsSelectCond(Cond)) {
+    APInt DemandedMask(APInt::getSignMask(BitWidth));
+    KnownBits Known;
+    TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
+                                          !DCI.isBeforeLegalizeOps());
+    if (!TLI.SimplifyDemandedBits(Cond, DemandedMask, Known, TLO, 0, true))
       return SDValue();
 
+    // If we changed the computation somewhere in the DAG, this change will
+    // affect all users of Cond. Update all the nodes so that we do not use
+    // the generic VSELECT anymore. Otherwise, we may perform wrong
+    // optimizations as we messed with the actual expectation for the vector
+    // boolean values.
+    for (SDNode *U : Cond->uses()) {
+      if (U->getOpcode() == X86ISD::BLENDV)
+        continue;
+
+      SDValue SB = DAG.getNode(X86ISD::BLENDV, SDLoc(U), U->getValueType(0),
+                               Cond, U->getOperand(1), U->getOperand(2));
+      DAG.ReplaceAllUsesOfValueWith(SDValue(U, 0), SB);
+      DCI.AddToWorklist(U);
+    }
+    DCI.CommitTargetLoweringOpt(TLO);
+    return SDValue(N, 0);
+  }
+
+  // Otherwise we can still at least try to simplify multiple use bits.
   APInt DemandedMask(APInt::getSignMask(BitWidth));
+  APInt DemandedElts(APInt::getAllOnesValue(VT.getVectorNumElements()));
   KnownBits Known;
   TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
                                         !DCI.isBeforeLegalizeOps());
-  if (!TLI.SimplifyDemandedBits(Cond, DemandedMask, Known, TLO, 0, true))
-    return SDValue();
+  if (SDValue V = TLI.SimplifyMultipleUseDemandedBits(Cond, DemandedMask,
+                                                      DemandedElts, DAG, 0))
+    return DAG.getNode(X86ISD::BLENDV, SDLoc(N), N->getValueType(0),
+                       V, N->getOperand(1), N->getOperand(2));
 
-  // If we changed the computation somewhere in the DAG, this change will
-  // affect all users of Cond. Update all the nodes so that we do not use
-  // the generic VSELECT anymore. Otherwise, we may perform wrong
-  // optimizations as we messed with the actual expectation for the vector
-  // boolean values.
-  for (SDNode *U : Cond->uses()) {
-    if (U->getOpcode() == X86ISD::BLENDV)
-      continue;
-
-    SDValue SB = DAG.getNode(X86ISD::BLENDV, SDLoc(U), U->getValueType(0),
-                             Cond, U->getOperand(1), U->getOperand(2));
-    DAG.ReplaceAllUsesOfValueWith(SDValue(U, 0), SB);
-    DCI.AddToWorklist(U);
-  }
-  DCI.CommitTargetLoweringOpt(TLO);
-  return SDValue(N, 0);
+  return SDValue();
 }
 
 /// Do target-specific dag combines on SELECT and VSELECT nodes.


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D71984.235533.patch
Type: text/x-patch
Size: 4083 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20191230/392708a8/attachment.bin>