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

[Luke Lau via llvm-commits]

================
@@ -1116,9 +1116,51 @@ define void @br_under_switch_default_common_dest_with_case(ptr %start, ptr %end,
 ; COST-LABEL: define void @br_under_switch_default_common_dest_with_case(
 ; COST-SAME: ptr [[START:%.*]], ptr [[END:%.*]], i64 [[X:%.*]]) #[[ATTR0]] {
 ; COST-NEXT:  [[ENTRY:.*]]:
+; COST-NEXT:    [[START2:%.*]] = ptrtoint ptr [[START]] to i64
+; COST-NEXT:    [[END1:%.*]] = ptrtoint ptr [[END]] to i64
+; COST-NEXT:    [[TMP0:%.*]] = add i64 [[END1]], -8
+; COST-NEXT:    [[TMP1:%.*]] = sub i64 [[TMP0]], [[START2]]
+; COST-NEXT:    [[TMP2:%.*]] = lshr i64 [[TMP1]], 3
+; COST-NEXT:    [[TMP3:%.*]] = add nuw nsw i64 [[TMP2]], 1
+; COST-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP3]], 4
+; COST-NEXT:    br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
+; COST:       [[VECTOR_PH]]:
+; COST-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[TMP3]], 4
+; COST-NEXT:    [[N_VEC:%.*]] = sub i64 [[TMP3]], [[N_MOD_VF]]
+; COST-NEXT:    [[TMP4:%.*]] = mul i64 [[N_VEC]], 8
+; COST-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[START]], i64 [[TMP4]]
+; COST-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i64> poison, i64 [[X]], i64 0
+; COST-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i64> [[BROADCAST_SPLATINSERT]], <4 x i64> poison, <4 x i32> zeroinitializer
+; COST-NEXT:    br label %[[VECTOR_BODY:.*]]
+; COST:       [[VECTOR_BODY]]:
+; COST-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; COST-NEXT:    [[OFFSET_IDX:%.*]] = mul i64 [[INDEX]], 8
+; COST-NEXT:    [[NEXT_GEP:%.*]] = getelementptr i8, ptr [[START]], i64 [[OFFSET_IDX]]
+; COST-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i64>, ptr [[NEXT_GEP]], align 1
+; COST-NEXT:    [[TMP6:%.*]] = icmp eq <4 x i64> [[WIDE_LOAD]], splat (i64 -12)
+; COST-NEXT:    [[TMP7:%.*]] = icmp eq <4 x i64> [[WIDE_LOAD]], splat (i64 13)
+; COST-NEXT:    [[TMP8:%.*]] = or <4 x i1> [[TMP6]], [[TMP7]]
+; COST-NEXT:    [[TMP9:%.*]] = xor <4 x i1> [[TMP8]], splat (i1 true)
+; COST-NEXT:    [[TMP10:%.*]] = icmp ule <4 x i64> [[WIDE_LOAD]], [[BROADCAST_SPLAT]]
+; COST-NEXT:    [[TMP11:%.*]] = xor <4 x i1> [[TMP10]], splat (i1 true)
+; COST-NEXT:    [[TMP12:%.*]] = select <4 x i1> [[TMP6]], <4 x i1> [[TMP11]], <4 x i1> zeroinitializer
+; COST-NEXT:    [[TMP13:%.*]] = or <4 x i1> [[TMP12]], [[TMP7]]
+; COST-NEXT:    call void @llvm.masked.store.v4i64.p0(<4 x i64> zeroinitializer, ptr [[NEXT_GEP]], i32 1, <4 x i1> [[TMP13]])
+; COST-NEXT:    [[TMP14:%.*]] = select <4 x i1> [[TMP6]], <4 x i1> [[TMP10]], <4 x i1> zeroinitializer
+; COST-NEXT:    call void @llvm.masked.store.v4i64.p0(<4 x i64> splat (i64 42), ptr [[NEXT_GEP]], i32 1, <4 x i1> [[TMP14]])
+; COST-NEXT:    [[TMP15:%.*]] = or <4 x i1> [[TMP14]], [[TMP9]]
+; COST-NEXT:    call void @llvm.masked.store.v4i64.p0(<4 x i64> splat (i64 2), ptr [[NEXT_GEP]], i32 1, <4 x i1> [[TMP15]])
+; COST-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
+; COST-NEXT:    [[TMP16:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; COST-NEXT:    br i1 [[TMP16]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP10:![0-9]+]]
+; COST:       [[MIDDLE_BLOCK]]:
+; COST-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[TMP3]], [[N_VEC]]
+; COST-NEXT:    br i1 [[CMP_N]], label %[[EXIT:.*]], label %[[SCALAR_PH]]
+; COST:       [[SCALAR_PH]]:
+; COST-NEXT:    [[BC_RESUME_VAL:%.*]] = phi ptr [ [[TMP5]], %[[MIDDLE_BLOCK]] ], [ [[START]], %[[ENTRY]] ]
 ; COST-NEXT:    br label %[[LOOP_HEADER:.*]]
 ; COST:       [[LOOP_HEADER]]:
-; COST-NEXT:    [[PTR_IV:%.*]] = phi ptr [ [[START]], %[[ENTRY]] ], [ [[PTR_IV_NEXT:%.*]], %[[LOOP_LATCH:.*]] ]
+; COST-NEXT:    [[PTR_IV:%.*]] = phi ptr [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[PTR_IV_NEXT:%.*]], %[[LOOP_LATCH:.*]] ]
 ; COST-NEXT:    [[L:%.*]] = load i64, ptr [[PTR_IV]], align 1
----------------
lukel97 wrote:

The total probability of all these switch destinations being executed is 1. BFI correctly divides up the probabilities, but previously each destination was given a 50% chance which meant the VF=1 VPlan was overly discounted.

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