[llvm] [WIP][VPlan] Use BlockFrequencyInfo in getPredBlockCostDivisor (PR #158690)

[Luke Lau via llvm-commits]

================
@@ -1073,27 +1073,49 @@ for.end:                                          ; preds = %for.inc
 define void @foo6(ptr nocapture readonly %in, ptr nocapture %out, i32 %size, ptr nocapture readonly %trigger) local_unnamed_addr #0 {
 ; AVX1-LABEL: define void @foo6(
 ; AVX1-SAME: ptr readonly captures(none) [[IN:%.*]], ptr captures(none) [[OUT:%.*]], i32 [[SIZE:%.*]], ptr readonly captures(none) [[TRIGGER:%.*]]) local_unnamed_addr #[[ATTR0]] {
-; AVX1-NEXT:  [[ENTRY:.*]]:
-; AVX1-NEXT:    br label %[[FOR_BODY:.*]]
-; AVX1:       [[FOR_BODY]]:
-; AVX1-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ 4095, %[[ENTRY]] ], [ [[INDVARS_IV_NEXT:%.*]], %[[FOR_INC:.*]] ]
+; AVX1-NEXT:  [[ENTRY:.*:]]
+; AVX1-NEXT:    br label %[[VECTOR_MEMCHECK:.*]]
+; AVX1:       [[VECTOR_MEMCHECK]]:
+; AVX1-NEXT:    [[SCEVGEP:%.*]] = getelementptr i8, ptr [[OUT]], i64 32768
+; AVX1-NEXT:    [[SCEVGEP1:%.*]] = getelementptr i8, ptr [[TRIGGER]], i64 16384
+; AVX1-NEXT:    [[SCEVGEP2:%.*]] = getelementptr i8, ptr [[IN]], i64 32768
+; AVX1-NEXT:    [[BOUND0:%.*]] = icmp ult ptr [[OUT]], [[SCEVGEP1]]
+; AVX1-NEXT:    [[BOUND1:%.*]] = icmp ult ptr [[TRIGGER]], [[SCEVGEP]]
+; AVX1-NEXT:    [[FOUND_CONFLICT:%.*]] = and i1 [[BOUND0]], [[BOUND1]]
+; AVX1-NEXT:    [[BOUND03:%.*]] = icmp ult ptr [[OUT]], [[SCEVGEP2]]
+; AVX1-NEXT:    [[BOUND14:%.*]] = icmp ult ptr [[IN]], [[SCEVGEP]]
+; AVX1-NEXT:    [[FOUND_CONFLICT5:%.*]] = and i1 [[BOUND03]], [[BOUND14]]
+; AVX1-NEXT:    [[CONFLICT_RDX:%.*]] = or i1 [[FOUND_CONFLICT]], [[FOUND_CONFLICT5]]
+; AVX1-NEXT:    br i1 [[CONFLICT_RDX]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
+; AVX1:       [[VECTOR_PH]]:
+; AVX1-NEXT:    br label %[[VECTOR_BODY:.*]]
+; AVX1:       [[VECTOR_BODY]]:
+; AVX1-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; AVX1-NEXT:    [[INDVARS_IV:%.*]] = sub i64 4095, [[INDEX]]
 ; AVX1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i32, ptr [[TRIGGER]], i64 [[INDVARS_IV]]
-; AVX1-NEXT:    [[TMP0:%.*]] = load i32, ptr [[ARRAYIDX]], align 4
-; AVX1-NEXT:    [[CMP1:%.*]] = icmp sgt i32 [[TMP0]], 0
-; AVX1-NEXT:    br i1 [[CMP1]], label %[[IF_THEN:.*]], label %[[FOR_INC]]
-; AVX1:       [[IF_THEN]]:
-; AVX1-NEXT:    [[ARRAYIDX3:%.*]] = getelementptr inbounds double, ptr [[IN]], i64 [[INDVARS_IV]]
-; AVX1-NEXT:    [[TMP1:%.*]] = load double, ptr [[ARRAYIDX3]], align 8
-; AVX1-NEXT:    [[ADD:%.*]] = fadd double [[TMP1]], 5.000000e-01
-; AVX1-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds double, ptr [[OUT]], i64 [[INDVARS_IV]]
-; AVX1-NEXT:    store double [[ADD]], ptr [[ARRAYIDX5]], align 8
-; AVX1-NEXT:    br label %[[FOR_INC]]
-; AVX1:       [[FOR_INC]]:
-; AVX1-NEXT:    [[INDVARS_IV_NEXT]] = add nsw i64 [[INDVARS_IV]], -1
-; AVX1-NEXT:    [[CMP:%.*]] = icmp eq i64 [[INDVARS_IV]], 0
-; AVX1-NEXT:    br i1 [[CMP]], label %[[FOR_END:.*]], label %[[FOR_BODY]]
-; AVX1:       [[FOR_END]]:
-; AVX1-NEXT:    ret void
+; AVX1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds i32, ptr [[ARRAYIDX]], i32 0
+; AVX1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i32, ptr [[TMP1]], i32 -3
+; AVX1-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, ptr [[TMP2]], align 4, !alias.scope [[META18:![0-9]+]]
+; AVX1-NEXT:    [[REVERSE:%.*]] = shufflevector <4 x i32> [[WIDE_LOAD]], <4 x i32> poison, <4 x i32> <i32 3, i32 2, i32 1, i32 0>
+; AVX1-NEXT:    [[TMP3:%.*]] = icmp sgt <4 x i32> [[REVERSE]], zeroinitializer
----------------
lukel97 wrote:

BPI computes `>= 0` to be slightly >50%, which influences the cost for AVX1 here.

https://github.com/llvm/llvm-project/pull/158690