================
@@ -652,3 +652,146 @@ void VPlanTransforms::attachCheckBlock(VPlan &Plan, Value 
*Cond,
     Term->addMetadata(LLVMContext::MD_prof, BranchWeights);
   }
 }
+
+bool VPlanTransforms::handleMaxMinNumReductionsWithoutFastMath(VPlan &Plan) {
+  VPRegionBlock *LoopRegion = Plan.getVectorLoopRegion();
+  VPReductionPHIRecipe *RedPhiR = nullptr;
+  bool HasUnsupportedPhi = false;
+
+  auto GetMinMaxCompareValue = [](VPReductionPHIRecipe *RedPhiR) -> VPValue * {
+    auto *MinMaxR = dyn_cast<VPRecipeWithIRFlags>(
+        RedPhiR->getBackedgeValue()->getDefiningRecipe());
+    if (!MinMaxR)
+      return nullptr;
+
+    auto *RepR = dyn_cast<VPReplicateRecipe>(MinMaxR);
+    if (!isa<VPWidenIntrinsicRecipe>(MinMaxR) &&
+        !(RepR && (isa<IntrinsicInst>(RepR->getUnderlyingInstr()))))
+      return nullptr;
+
+#ifndef NDEBUG
+    Intrinsic::ID RdxIntrinsicId =
+        RedPhiR->getRecurrenceKind() == RecurKind::FMaxNum ? Intrinsic::maxnum
+                                                           : Intrinsic::minnum;
+    assert((isa<VPWidenIntrinsicRecipe>(MinMaxR) &&
+            cast<VPWidenIntrinsicRecipe>(MinMaxR)->getVectorIntrinsicID() ==
+                RdxIntrinsicId) ||
+           (RepR &&
+            cast<IntrinsicInst>(RepR->getUnderlyingInstr())->getIntrinsicID() 
==
+                RdxIntrinsicId) &&
+               "Intrinsic did not match recurrence kind");
+#endif
+
+    if (MinMaxR->getOperand(0) == RedPhiR)
+      return MinMaxR->getOperand(1);
+
+    assert(MinMaxR->getOperand(1) == RedPhiR &&
+           "Reduction phi operand expected");
+    return MinMaxR->getOperand(0);
+  };
+
+  for (auto &R : LoopRegion->getEntryBasicBlock()->phis()) {
+    // TODO: Also support fixed-order recurrence phis.
+    HasUnsupportedPhi |=
+        !isa<VPCanonicalIVPHIRecipe, VPWidenIntOrFpInductionRecipe,
+             VPReductionPHIRecipe>(&R);
+    auto *Cur = dyn_cast<VPReductionPHIRecipe>(&R);
+    if (!Cur)
+      continue;
+    // For now, only a single reduction is supported.
+    // TODO: Support multiple MaxNum/MinNum reductions and other reductions.
+    if (RedPhiR)
+      return false;
+    if (Cur->getRecurrenceKind() != RecurKind::FMaxNum &&
+        Cur->getRecurrenceKind() != RecurKind::FMinNum)
+      continue;
+    RedPhiR = Cur;
+  }
+
+  if (!RedPhiR)
+    return true;
+
+  RecurKind RedPhiRK = RedPhiR->getRecurrenceKind();
+  assert((RedPhiRK == RecurKind::FMaxNum || RedPhiRK == RecurKind::FMinNum) &&
+         "unsupported reduction");
+
+  VPValue *MinMaxOp = GetMinMaxCompareValue(RedPhiR);
+  if (!MinMaxOp)
+    return false;
+
+  // We won't be able to resume execution in the scalar tail, if there are
+  // unsupported header phis or there is no scalar tail at all, due to
+  // tail-folding.
+  if (HasUnsupportedPhi || !Plan.hasScalarTail())
+    return false;
+
+  /// Check if the vector loop of \p Plan can early exit and restart
+  /// execution of last vector iteration in the scalar loop. This requires all
+  /// recipes up to early exit point be side-effect free as they are
+  /// re-executed. Currently we check that the loop is free of any recipe that
+  /// may write to memory. Expected to operate on an early VPlan w/o nested
+  /// regions.
+  for (VPBlockBase *VPB : vp_depth_first_shallow(
+           Plan.getVectorLoopRegion()->getEntryBasicBlock())) {
+    auto *VPBB = cast<VPBasicBlock>(VPB);
+    for (auto &R : *VPBB) {
+      if (R.mayWriteToMemory() &&
+          !match(&R, m_BranchOnCount(m_VPValue(), m_VPValue())))
+        return false;
+    }
+  }
+
+  VPBasicBlock *LatchVPBB = LoopRegion->getExitingBasicBlock();
+  VPBuilder Builder(LatchVPBB->getTerminator());
+  auto *LatchExitingBranch = cast<VPInstruction>(LatchVPBB->getTerminator());
+  assert(LatchExitingBranch->getOpcode() == VPInstruction::BranchOnCount &&
+         "Unexpected terminator");
+  auto *IsLatchExitTaken =
+      Builder.createICmp(CmpInst::ICMP_EQ, LatchExitingBranch->getOperand(0),
+                         LatchExitingBranch->getOperand(1));
+
+  VPValue *IsNaN = Builder.createFCmp(CmpInst::FCMP_UNO, MinMaxOp, MinMaxOp);
+  VPValue *AnyNaN = Builder.createNaryOp(VPInstruction::AnyOf, {IsNaN});
+  auto *AnyExitTaken =
+      Builder.createNaryOp(Instruction::Or, {AnyNaN, IsLatchExitTaken});
+  Builder.createNaryOp(VPInstruction::BranchOnCond, AnyExitTaken);
+  LatchExitingBranch->eraseFromParent();
+
+  // If we exit early due to NaNs, compute the final reduction result based on
+  // the reduction phi at the beginning of the last vector iteration.
+  auto *RdxResult = find_singleton<VPSingleDefRecipe>(
+      RedPhiR->users(), [](VPUser *U, bool) -> VPSingleDefRecipe * {
+        auto *VPI = dyn_cast<VPInstruction>(U);
+        if (VPI && VPI->getOpcode() == VPInstruction::ComputeReductionResult)
+          return VPI;
+        return nullptr;
+      });
+
+  auto *MiddleVPBB = Plan.getMiddleBlock();
+  Builder.setInsertPoint(MiddleVPBB, MiddleVPBB->begin());
+  auto *NewSel =
+      Builder.createSelect(AnyNaN, RedPhiR, RdxResult->getOperand(1));
+  RdxResult->setOperand(1, NewSel);
+
+  auto *ScalarPH = Plan.getScalarPreheader();
+  // Update the resume phis for inductions in the scalar preheader. If AnyNaN 
is
+  // true, the resume from the start of the last vector iteration via the
----------------
fhahn wrote:

Updated, thanks

https://github.com/llvm/llvm-project/pull/148239
_______________________________________________
lldb-commits mailing list
lldb-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/lldb-commits

Reply via email to