[LV] Pick correct BB as insert point when fixing PHI for FORs.

[fhahn]

Summary

Currently we fail to pick the right insertion point when
PreviousLastPart of a first-order-recurrence is a PHI node not in the
LoopVectorBody. This can happen when PreviousLastPart is produce in a
predicated block. In that case, we should pick the insertion point in
the BB the PHI is in.

Fixes PR44020.

[merge_guards_bot]

Build result: pass - 60515 tests passed, 0 failed and 726 were skipped.

Log files: console-log.txt, CMakeCache.txt

[Ayal]

Looks good to me with minor optional notes.
Consider adding a target-indepedent test, e.g., augmenting first-order-recurrence.ll; tests can possibly be committed before the fix, demonstrating current behavior and then the change brought by this fix.

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

3557

(Note that Previous that is loop invariant forms no recurrence ... they best be recognized early (if/why not LICM'd prior to LV), and optimized accordingly. A todo for/following D68831?)

3568

Can refactor to have a single Builder.SetInsertPoint(Point); following the different cases for setting Point.

llvm/test/Transforms/LoopVectorize/SystemZ/predicated-first-order-recurrence.ll

3

State what this test is designed to check?
("In the test case we predicate due to low trip count right? It might be worth calling that out." ;-)

[Ayal]

llvm/test/Transforms/LoopVectorize/SystemZ/predicated-first-order-recurrence.ll

3

Mention PR44020

[fhahn]

In D71071#1772329, @Ayal wrote:

Looks good to me with minor optional notes.
Consider adding a target-indepedent test, e.g., augmenting first-order-recurrence.ll; tests can possibly be committed before the fix, demonstrating current behavior and then the change brought by this fix.

I tried, but the test case requires tail folding and it seems to depend on an actual target. We make the decision to tail-fold in computeMaxVF (https://github.com/llvm/llvm-project/blob/c49194969430f0ee817498a7000a979a7a0ded03/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp#L4941) and use the max target VF, not the forced vector width. This seems a bit odd and we might want to change that, to make it easier to test tail-folding without any specific target. What do you think?

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

3557

I've added a FIXME in the committed version

llvm/test/Transforms/LoopVectorize/SystemZ/predicated-first-order-recurrence.ll

3

Done thanks :)

[Ayal]

In D71071#1774076, @fhahn wrote:

In D71071#1772329, @Ayal wrote:

Looks good to me with minor optional notes.
Consider adding a target-indepedent test, e.g., augmenting first-order-recurrence.ll; tests can possibly be committed before the fix, demonstrating current behavior and then the change brought by this fix.

I tried, but the test case requires tail folding and it seems to depend on an actual target. We make the decision to tail-fold in computeMaxVF (https://github.com/llvm/llvm-project/blob/c49194969430f0ee817498a7000a979a7a0ded03/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp#L4941) and use the max target VF, not the forced vector width. This seems a bit odd and we might want to change that, to make it easier to test tail-folding without any specific target. What do you think?

The -prefer-predicate-over-epilog flag should help test tail folding w/o a specific target. Admittedly it is currently used by target-dependent tests only.

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

void InnerLoopVectorizer::fixFirstOrderRecurrence(PHINode *Phi) {
   VecPhi->addIncoming(VectorInit, LoopVectorPreHeader);
 
   // Get the vectorized previous value of the last part UF - 1. It appears last
   // among all unrolled iterations, due to the order of their construction.
   Value *PreviousLastPart = getOrCreateVectorValue(Previous, UF - 1);
 
   // Set the insertion point after the previous value if it is an instruction.
   // Note that the previous value may have been constant-folded so it is not
-  // guaranteed to be an instruction in the vector loop. Also, if the previous
-  // value is a phi node, we should insert after all the phi nodes to avoid
-  // breaking basic block verification.
-  if (LI->getLoopFor(LoopVectorBody)->isLoopInvariant(PreviousLastPart) ||
-      isa<PHINode>(PreviousLastPart))
+  // guaranteed to be an instruction in the vector loop.
+  if (LI->getLoopFor(LoopVectorBody)->isLoopInvariant(PreviousLastPart))
     Builder.SetInsertPoint(&*LoopVectorBody->getFirstInsertionPt());
-  else
+  else if (isa<PHINode>(PreviousLastPart)) {
+    // If the previous value is a phi node, we should insert after all the phi
+    // nodes in the block containing the PHI to avoid breaking basic block
+    // verification. Note that the basic block may be different to
+    // LoopVectorBody, in case we predicate the loop.
+    BasicBlock *PhiBlock = cast<Instruction>(PreviousLastPart)->getParent();
+    Builder.SetInsertPoint(&*PhiBlock->getFirstInsertionPt());
+  } else
     Builder.SetInsertPoint(
         &*++BasicBlock::iterator(cast<Instruction>(PreviousLastPart)));
 
   // We will construct a vector for the recurrence by combining the values for
   // the current and previous iterations. This is the required shuffle mask.
   SmallVector<Constant *, 8> ShuffleMask(VF);
   ShuffleMask[0] = Builder.getInt32(VF - 1);
   for (unsigned I = 1; I < VF; ++I)
     ShuffleMask[I] = Builder.getInt32(I + VF - 1);
 

llvm/test/Transforms/LoopVectorize/SystemZ/predicated-first-order-recurrence.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -loop-vectorize -mtriple=s390x-ibm-linux -mcpu=z13 -force-vector-width=2 -S %s | FileCheck %s
+
+@A = external dso_local global [5 x i32], align 4
+@B = external dso_local global [5 x i32], align 4
+
+define void @func_21() {
+; CHECK-LABEL: @func_21(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
+; CHECK:       vector.ph:
+; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
+; CHECK:       vector.body:
+; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[PRED_STORE_CONTINUE4:%.*]] ]
+; CHECK-NEXT:    [[VECTOR_RECUR:%.*]] = phi <2 x i32> [ <i32 undef, i32 0>, [[VECTOR_PH]] ], [ [[TMP12:%.*]], [[PRED_STORE_CONTINUE4]] ]
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <2 x i64> undef, i64 [[INDEX]], i32 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <2 x i64> [[BROADCAST_SPLATINSERT]], <2 x i64> undef, <2 x i32> zeroinitializer
+; CHECK-NEXT:    [[INDUCTION:%.*]] = add <2 x i64> [[BROADCAST_SPLAT]], <i64 0, i64 1>
+; CHECK-NEXT:    [[TMP0:%.*]] = add i64 [[INDEX]], 0
+; CHECK-NEXT:    [[TMP1:%.*]] = add i64 [[INDEX]], 1
+; CHECK-NEXT:    [[TMP2:%.*]] = icmp ule <2 x i64> [[INDUCTION]], <i64 4, i64 4>
+; CHECK-NEXT:    [[TMP3:%.*]] = extractelement <2 x i1> [[TMP2]], i32 0
+; CHECK-NEXT:    br i1 [[TMP3]], label [[PRED_LOAD_IF:%.*]], label [[PRED_LOAD_CONTINUE:%.*]]
+; CHECK:       pred.load.if:
+; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [5 x i32], [5 x i32]* @A, i64 0, i64 [[TMP0]]
+; CHECK-NEXT:    [[TMP5:%.*]] = load i32, i32* [[TMP4]], align 4
+; CHECK-NEXT:    [[TMP6:%.*]] = insertelement <2 x i32> undef, i32 [[TMP5]], i32 0
+; CHECK-NEXT:    br label [[PRED_LOAD_CONTINUE]]
+; CHECK:       pred.load.continue:
+; CHECK-NEXT:    [[TMP7:%.*]] = phi <2 x i32> [ undef, [[VECTOR_BODY]] ], [ [[TMP6]], [[PRED_LOAD_IF]] ]
+; CHECK-NEXT:    [[TMP8:%.*]] = extractelement <2 x i1> [[TMP2]], i32 1
+; CHECK-NEXT:    br i1 [[TMP8]], label [[PRED_LOAD_IF1:%.*]], label [[PRED_LOAD_CONTINUE2:%.*]]
+; CHECK:       pred.load.if1:
+; CHECK-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [5 x i32], [5 x i32]* @A, i64 0, i64 [[TMP1]]
+; CHECK-NEXT:    [[TMP10:%.*]] = load i32, i32* [[TMP9]], align 4
+; CHECK-NEXT:    [[TMP11:%.*]] = insertelement <2 x i32> [[TMP7]], i32 [[TMP10]], i32 1
+; CHECK-NEXT:    br label [[PRED_LOAD_CONTINUE2]]
+; CHECK:       pred.load.continue2:
+; CHECK-NEXT:    [[TMP12]] = phi <2 x i32> [ [[TMP7]], [[PRED_LOAD_CONTINUE]] ], [ [[TMP11]], [[PRED_LOAD_IF1]] ]
+; CHECK-NEXT:    [[TMP13:%.*]] = shufflevector <2 x i32> [[VECTOR_RECUR]], <2 x i32> [[TMP12]], <2 x i32> <i32 1, i32 2>
+; CHECK-NEXT:    [[TMP14:%.*]] = extractelement <2 x i1> [[TMP2]], i32 0
+; CHECK-NEXT:    br i1 [[TMP14]], label [[PRED_STORE_IF:%.*]], label [[PRED_STORE_CONTINUE:%.*]]
+; CHECK:       pred.store.if:
+; CHECK-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [5 x i32], [5 x i32]* @B, i64 0, i64 [[TMP0]]
+; CHECK-NEXT:    [[TMP16:%.*]] = extractelement <2 x i32> [[TMP13]], i32 0
+; CHECK-NEXT:    store i32 [[TMP16]], i32* [[TMP15]], align 4
+; CHECK-NEXT:    br label [[PRED_STORE_CONTINUE]]
+; CHECK:       pred.store.continue:
+; CHECK-NEXT:    [[TMP17:%.*]] = extractelement <2 x i1> [[TMP2]], i32 1
+; CHECK-NEXT:    br i1 [[TMP17]], label [[PRED_STORE_IF3:%.*]], label [[PRED_STORE_CONTINUE4]]
+; CHECK:       pred.store.if3:
+; CHECK-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [5 x i32], [5 x i32]* @B, i64 0, i64 [[TMP1]]
+; CHECK-NEXT:    [[TMP19:%.*]] = extractelement <2 x i32> [[TMP13]], i32 1
+; CHECK-NEXT:    store i32 [[TMP19]], i32* [[TMP18]], align 4
+; CHECK-NEXT:    br label [[PRED_STORE_CONTINUE4]]
+; CHECK:       pred.store.continue4:
+; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 2
+; CHECK-NEXT:    [[TMP20:%.*]] = icmp eq i64 [[INDEX_NEXT]], 6
+; CHECK-NEXT:    br i1 [[TMP20]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !0
+; CHECK:       middle.block:
+; CHECK-NEXT:    [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <2 x i32> [[TMP12]], i32 1
+; CHECK-NEXT:    [[VECTOR_RECUR_EXTRACT_FOR_PHI:%.*]] = extractelement <2 x i32> [[TMP12]], i32 0
+; CHECK-NEXT:    br i1 true, label [[EXIT:%.*]], label [[SCALAR_PH]]
+; CHECK:       scalar.ph:
+; CHECK-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i32 [ [[VECTOR_RECUR_EXTRACT]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
+; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ 6, [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
+; CHECK-NEXT:    br label [[LOOP:%.*]]
+; CHECK:       loop:
+; CHECK-NEXT:    [[SCALAR_RECUR:%.*]] = phi i32 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[LV:%.*]], [[LOOP]] ]
+; CHECK-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[LOOP]] ]
+; CHECK-NEXT:    [[A_PTR:%.*]] = getelementptr inbounds [5 x i32], [5 x i32]* @A, i64 0, i64 [[INDVARS_IV]]
+; CHECK-NEXT:    [[LV]] = load i32, i32* [[A_PTR]], align 4
+; CHECK-NEXT:    [[B_PTR:%.*]] = getelementptr inbounds [5 x i32], [5 x i32]* @B, i64 0, i64 [[INDVARS_IV]]
+; CHECK-NEXT:    store i32 [[SCALAR_RECUR]], i32* [[B_PTR]], align 4
+; CHECK-NEXT:    [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
+; CHECK-NEXT:    [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], 5
+; CHECK-NEXT:    br i1 [[EXITCOND]], label [[EXIT]], label [[LOOP]], !llvm.loop !2
+; CHECK:       exit:
+; CHECK-NEXT:    ret void
+;
+entry:
+  br label %loop
+
+loop:                                    ; preds = %loop, %entry
+  %rec = phi i32 [ 0, %entry], [ %lv, %loop ]
+  %indvars.iv = phi i64 [ 0, %entry], [ %indvars.iv.next, %loop ]
+  %A.ptr= getelementptr inbounds [5 x i32], [5 x i32]* @A, i64 0, i64 %indvars.iv
+  %lv = load i32, i32* %A.ptr, align 4
+  %B.ptr = getelementptr inbounds [5 x i32], [5 x i32]* @B, i64 0, i64 %indvars.iv
+  store i32 %rec, i32* %B.ptr, align 4
+  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
+  %exitcond = icmp eq i64 %indvars.iv.next, 5
+  br i1 %exitcond, label %exit, label %loop
+
+exit:                                             ; preds = %loop
+  ret void
+}