[llvm] [LV] Support argmin/argmax with strict predicates. (PR #170223)

[via llvm-commits]

================
@@ -1120,6 +1122,129 @@ bool VPlanTransforms::handleMaxMinNumReductions(VPlan &Plan) {
   return true;
 }
 
+/// For argmin/argmax reductions with strict predicates, convert the existing
+/// FindLastIV reduction to a new UMin reduction of a wide canonical IV. If the
+/// original IV was not canonical, a new canonical wide IV is added, and the
+/// final result is scaled back to the original IV.
+static bool handleFirstArgMinArgMax(VPlan &Plan,
+                                    VPReductionPHIRecipe *MinMaxPhiR,
+                                    VPReductionPHIRecipe *FindIVPhiR,
+                                    VPWidenIntOrFpInductionRecipe *WideIV,
+                                    VPInstruction *MinMaxResult) {
+  Type *Ty = Plan.getVectorLoopRegion()->getCanonicalIVType();
+  // TODO: Support different IV types.
+  if (Ty != VPTypeAnalysis(Plan).inferScalarType(FindIVPhiR))
+    return false;
+
+  // If the original wide IV is not canonical, create a new one. The wide IV is
+  // guaranteed to not wrap for all lanes that are active in the vector loop.
+  if (!WideIV->isCanonical()) {
+    VPValue *Zero = Plan.getConstantInt(Ty, 0);
+    VPValue *One = Plan.getConstantInt(Ty, 1);
+    auto *WidenCanIV = new VPWidenIntOrFpInductionRecipe(
+        nullptr, Zero, One, WideIV->getVFValue(),
+        WideIV->getInductionDescriptor(),
+        VPIRFlags::WrapFlagsTy(/*HasNUW=*/true, /*HasNSW=*/false),
+        WideIV->getDebugLoc());
+    WidenCanIV->insertBefore(WideIV);
+
+    // Update the select to use the wide canonical IV.
+    auto *SelectR = cast<VPSingleDefRecipe>(
+        FindIVPhiR->getBackedgeValue()->getDefiningRecipe());
+    assert(match(SelectR, m_Select(m_VPValue(), m_VPValue(), m_VPValue())) &&
+           "backedge value must be a select");
+    WideIV->replaceUsesWithIf(WidenCanIV, [SelectR](const VPUser &U, unsigned) {
+      return SelectR == &U;
+    });
+  }
+
+  // Create the new UMin reduction recipe to track the minimum index.
+  assert(!FindIVPhiR->isInLoop() && !FindIVPhiR->isOrdered() &&
+         "inloop and ordered reductions not supported");
+  VPValue *MaxInt =
+      Plan.getConstantInt(APInt::getMaxValue(Ty->getIntegerBitWidth()));
+  assert(FindIVPhiR->getVFScaleFactor() == 1 &&
+         "FindIV reduction must not be scaled");
+  ReductionStyle Style = RdxUnordered{1};
+  auto *FirstIdxPhiR = new VPReductionPHIRecipe(
+      dyn_cast_or_null<PHINode>(FindIVPhiR->getUnderlyingValue()),
+      RecurKind::UMin, *MaxInt, *FindIVPhiR->getBackedgeValue(), Style,
+      FindIVPhiR->hasUsesOutsideReductionChain());
+  FirstIdxPhiR->insertBefore(FindIVPhiR);
+
+  VPInstruction *FindLastIVResult =
+      findUserOf<VPInstruction::ComputeFindIVResult>(FindIVPhiR);
+  MinMaxResult->moveBefore(*FindLastIVResult->getParent(),
+                           FindLastIVResult->getIterator());
+
+  // The reduction using MinMaxPhiR needs adjusting to compute the correct
+  // result:
+  //  1. Find the first canonical indices corresponding to partial min/max
+  //     values, using loop reductions.
+  //  2. Find which of the partial min/max values are equal to the overall
+  //      min/max value.
+  //  3. Select among the canonical indices those corresponding to the overall
+  //     min/max value.
+  //  4. Find the first canonical index of overall min/max and scale it back to
+  //     the original IV using VPDerivedIVRecipe.
+  //  5. If the overall min/max is equal to the start value, the condition in
+  //  the
+  //     loop was always false, due to being strict; return the start value in
+  //     that case.
+  //
+  // The original reductions need adjusting:
+  // For example, this transforms
+  // vp<%min.result> = compute-reduction-result ir<%min.val>, ir<%min.val.next>
+  // vp<%find.iv.result> = compute-find-iv-result ir<%min.idx>, ir<0>,
+  //                                              ir<Sentinel>,
+  //                                              vp<%min.idx.next>
+  //
+  // into:
+  //  vp<%min.result> = compute-reduction-result ir<%min.val>, ir<%min.val.next>
+  //  vp<%final.min.cmp> = icmp eq ir<%min.val.next>, vp<%min.result>
----------------
ayalz wrote:

Getting the recurrence kind from a live-in operand would be better than hooking it up to the header phi recipe, especially when only the former needs to be updated.

https://github.com/llvm/llvm-project/pull/170223