[llvm] [LoopVectorize] Vectorize select-cmp reduction pattern for increasing integer induction variable (PR #67812)

Mel Chen via llvm-commits llvm-commits at lists.llvm.org
Wed Oct 4 00:10:39 PDT 2023


================
@@ -662,6 +664,96 @@ RecurrenceDescriptor::isAnyOfPattern(Loop *Loop, PHINode *OrigPhi,
                                                      : RecurKind::FAnyOf);
 }
 
+// We are looking for loops that do something like this:
+//   int r = 0;
+//   for (int i = 0; i < n; i++) {
+//     if (src[i] > 3)
+//       r = i;
+//   }
+// The reduction value (r) is derived from either the values of an increasing
+// induction variable (i) sequence, or from the start value (0).
+// The LLVM IR generated for such loops would be as follows:
+//   for.body:
+//     %r = phi i32 [ %spec.select, %for.body ], [ 0, %entry ]
+//     %i = phi i32 [ %inc, %for.body ], [ 0, %entry ]
+//     ...
+//     %cmp = icmp sgt i32 %5, 3
+//     %spec.select = select i1 %cmp, i32 %i, i32 %r
+//     %inc = add nsw i32 %i, 1
+//     ...
+// Since 'i' is an increasing induction variable, the reduction value after the
+// loop will be the maximum value of 'i' that the condition (src[i] > 3) is
+// satisfied, or the start value (0 in the example above). When the start value
+// of the increasing induction variable 'i' is greater than the minimum value of
+// the data type, we can use the minimum value of the data type as a sentinel
+// value to replace the start value. This allows us to perform a single
+// reduction max operation to obtain the final reduction result.
+// TODO: It is possible to solve the case where the start value is the minimum
+// value of the data type or a non-constant value by using mask and multiple
+// reduction operations.
+RecurrenceDescriptor::InstDesc
+RecurrenceDescriptor::isFindLastIVPattern(Loop *Loop, PHINode *OrigPhi,
+                                          Instruction *I, ScalarEvolution *SE) {
+  // Only match select with single use cmp condition.
+  // TODO: Only handle single use for now.
+  CmpInst::Predicate Pred;
+  if (!match(I, m_Select(m_OneUse(m_Cmp(Pred, m_Value(), m_Value())), m_Value(),
+                         m_Value())))
+    return InstDesc(false, I);
+
+  SelectInst *SI = cast<SelectInst>(I);
+  Value *NonRdxPhi = nullptr;
+
+  if (OrigPhi == dyn_cast<PHINode>(SI->getTrueValue()))
+    NonRdxPhi = SI->getFalseValue();
+  else if (OrigPhi == dyn_cast<PHINode>(SI->getFalseValue()))
+    NonRdxPhi = SI->getTrueValue();
+  else
+    return InstDesc(false, I);
+
+  auto IsIncreasingLoopInduction = [&SE, &Loop](Value *V) {
+    if (!SE)
+      return false;
+
+    Type *Ty = V->getType();
+    if (!SE->isSCEVable(Ty))
+      return false;
+
+    auto *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(V));
+    if (!AR)
+      return false;
+
+    const SCEV *Step = AR->getStepRecurrence(*SE);
+    if (!SE->isKnownPositive(Step))
+      return false;
+
+    const ConstantRange IVRange = SE->getSignedRange(AR);
+    unsigned NumBits = Ty->getIntegerBitWidth();
+    // Keep the minmum value of the recurrence type as the sentinel value.
+    // The maximum acceptable range for the increasing induction variable,
+    // called the valid range, will be defined as
+    //   [<sentinel value> + 1, SignedMin(<recurrence type>))
+    // TODO: This range restriction can be lifted by adding an additional
+    // virtual OR reduction.
+    const APInt Sentinel = APInt::getSignedMinValue(NumBits);
+    const ConstantRange ValidRange = ConstantRange::getNonEmpty(
+        Sentinel + 1, APInt::getSignedMinValue(NumBits));
----------------
Mel-Chen wrote:

Sure. 
This is because the ConstantRange::getNonEmpty(L, U) actually produces the range [L, U). If calling ConstantRange::getNonEmpty(SignedMin + 1, SignedMin), for integer type, [SignedMin + 1, SignedMin) is equivalent to [SignedMin + 1, SignedMax]. 
This part has been confusing for many people, including myself when I was initially learning about ConstantRange. I am currently considering change it to:
```
// Valid range = Full set - Sentinel 
const ConstantRange ValidRange = 
    ConstantRange::getFull(NumBits).difference(ConstantRange(Sentinel));
```
Please let me know if you think this representation would be better.


https://github.com/llvm/llvm-project/pull/67812


More information about the llvm-commits mailing list