223 lines · plain
1; This test verifies that the loop vectorizer will not vectorizes low trip count2; loops that require runtime checks (Trip count is computed with profile info).3; REQUIRES: asserts4; RUN: opt < %s -passes=loop-vectorize -loop-vectorize-with-block-frequency -S | FileCheck %s5 6target datalayout = "E-m:e-p:32:32-i64:32-f64:32:64-a:0:32-n32-S128"7 8@tab = common global [32 x i8] zeroinitializer, align 19 10define i32 @foo_low_trip_count1(i32 %bound) {11; Simple loop with low tripcount. Should not be vectorized.12 13; CHECK-LABEL: @foo_low_trip_count1(14; CHECK-NOT: <{{[0-9]+}} x i8>15 16entry:17 br label %for.body18 19for.body: ; preds = %for.body, %entry20 %i.08 = phi i32 [ 0, %entry ], [ %inc, %for.body ]21 %arrayidx = getelementptr inbounds [32 x i8], ptr @tab, i32 0, i32 %i.0822 %0 = load i8, ptr %arrayidx, align 123 %cmp1 = icmp eq i8 %0, 024 %. = select i1 %cmp1, i8 2, i8 125 store i8 %., ptr %arrayidx, align 126 %inc = add nsw i32 %i.08, 127 %exitcond = icmp eq i32 %i.08, %bound28 br i1 %exitcond, label %for.end, label %for.body, !prof !129 30for.end: ; preds = %for.body31 ret i32 032}33 34define i32 @foo_low_trip_count2(i32 %bound) !prof !0 {35; The loop has a same invocation count with the function, but has a low36; trip_count per invocation and not worth to vectorize.37 38; CHECK-LABEL: @foo_low_trip_count2(39; CHECK-NOT: <{{[0-9]+}} x i8>40 41entry:42 br label %for.body43 44for.body: ; preds = %for.body, %entry45 %i.08 = phi i32 [ 0, %entry ], [ %inc, %for.body ]46 %arrayidx = getelementptr inbounds [32 x i8], ptr @tab, i32 0, i32 %i.0847 %0 = load i8, ptr %arrayidx, align 148 %cmp1 = icmp eq i8 %0, 049 %. = select i1 %cmp1, i8 2, i8 150 store i8 %., ptr %arrayidx, align 151 %inc = add nsw i32 %i.08, 152 %exitcond = icmp eq i32 %i.08, %bound53 br i1 %exitcond, label %for.end, label %for.body, !prof !154 55for.end: ; preds = %for.body56 ret i32 057}58 59define i32 @foo_low_trip_count3(i1 %cond, i32 %bound) !prof !0 {60; The loop has low invocation count compare to the function invocation count,61; but has a high trip count per invocation. Vectorize it.62 63; CHECK-LABEL: @foo_low_trip_count3(64; CHECK: [[VECTOR_BODY:vector\.body]]:65; CHECK: br i1 [[TMP9:%.*]], label [[MIDDLE_BLOCK:%.*]], label %[[VECTOR_BODY]], !prof [[LP3:\!.*]],66; CHECK: [[FOR_BODY:for\.body]]:67; CHECK: br i1 [[EXITCOND:%.*]], label [[FOR_END_LOOPEXIT:%.*]], label %[[FOR_BODY]], !prof [[LP6:\!.*]],68entry:69 br i1 %cond, label %for.preheader, label %for.end, !prof !270 71for.preheader:72 br label %for.body73 74for.body: ; preds = %for.body, %entry75 %i.08 = phi i32 [ 0, %for.preheader ], [ %inc, %for.body ]76 %arrayidx = getelementptr inbounds [32 x i8], ptr @tab, i32 0, i32 %i.0877 %0 = load i8, ptr %arrayidx, align 178 %cmp1 = icmp eq i8 %0, 079 %. = select i1 %cmp1, i8 2, i8 180 store i8 %., ptr %arrayidx, align 181 %inc = add nsw i32 %i.08, 182 %exitcond = icmp eq i32 %i.08, %bound83 br i1 %exitcond, label %for.end, label %for.body, !prof !384 85for.end: ; preds = %for.body86 ret i32 087}88 89define i32 @foo_low_trip_count_icmp_sgt(i32 %bound) {90; Simple loop with low tripcount and inequality test for exit.91; Should not be vectorized.92 93; CHECK-LABEL: @foo_low_trip_count_icmp_sgt(94; CHECK-NOT: <{{[0-9]+}} x i8>95 96entry:97 br label %for.body98 99for.body: ; preds = %for.body, %entry100 %i.08 = phi i32 [ 0, %entry ], [ %inc, %for.body ]101 %arrayidx = getelementptr inbounds [32 x i8], ptr @tab, i32 0, i32 %i.08102 %0 = load i8, ptr %arrayidx, align 1103 %cmp1 = icmp eq i8 %0, 0104 %. = select i1 %cmp1, i8 2, i8 1105 store i8 %., ptr %arrayidx, align 1106 %inc = add nsw i32 %i.08, 1107 %exitcond = icmp sgt i32 %i.08, %bound108 br i1 %exitcond, label %for.end, label %for.body, !prof !1109 110for.end: ; preds = %for.body111 ret i32 0112}113 114define i32 @const_low_trip_count() {115; Simple loop with constant, small trip count and no profiling info.116 117; CHECK-LABEL: @const_low_trip_count118; CHECK-NOT: <{{[0-9]+}} x i8>119 120entry:121 br label %for.body122 123for.body: ; preds = %for.body, %entry124 %i.08 = phi i32 [ 0, %entry ], [ %inc, %for.body ]125 %arrayidx = getelementptr inbounds [32 x i8], ptr @tab, i32 0, i32 %i.08126 %0 = load i8, ptr %arrayidx, align 1127 %cmp1 = icmp eq i8 %0, 0128 %. = select i1 %cmp1, i8 2, i8 1129 store i8 %., ptr %arrayidx, align 1130 %inc = add nsw i32 %i.08, 1131 %exitcond = icmp slt i32 %i.08, 2132 br i1 %exitcond, label %for.body, label %for.end133 134for.end: ; preds = %for.body135 ret i32 0136}137 138define i32 @const_large_trip_count() {139; Simple loop with constant large trip count and no profiling info.140 141; CHECK-LABEL: @const_large_trip_count142; CHECK: <{{[0-9]+}} x i8>143 144entry:145 br label %for.body146 147for.body: ; preds = %for.body, %entry148 %i.08 = phi i32 [ 0, %entry ], [ %inc, %for.body ]149 %arrayidx = getelementptr inbounds [32 x i8], ptr @tab, i32 0, i32 %i.08150 %0 = load i8, ptr %arrayidx, align 1151 %cmp1 = icmp eq i8 %0, 0152 %. = select i1 %cmp1, i8 2, i8 1153 store i8 %., ptr %arrayidx, align 1154 %inc = add nsw i32 %i.08, 1155 %exitcond = icmp slt i32 %i.08, 1000156 br i1 %exitcond, label %for.body, label %for.end157 158for.end: ; preds = %for.body159 ret i32 0160}161 162define i32 @const_small_trip_count_step() {163; Simple loop with static, small trip count and no profiling info.164 165; CHECK-LABEL: @const_small_trip_count_step166; CHECK-NOT: <{{[0-9]+}} x i8>167 168entry:169 br label %for.body170 171for.body: ; preds = %for.body, %entry172 %i.08 = phi i32 [ 0, %entry ], [ %inc, %for.body ]173 %arrayidx = getelementptr inbounds [32 x i8], ptr @tab, i32 0, i32 %i.08174 %0 = load i8, ptr %arrayidx, align 1175 %cmp1 = icmp eq i8 %0, 0176 %. = select i1 %cmp1, i8 2, i8 1177 store i8 %., ptr %arrayidx, align 1178 %inc = add nsw i32 %i.08, 5179 %exitcond = icmp slt i32 %i.08, 10180 br i1 %exitcond, label %for.body, label %for.end181 182for.end: ; preds = %for.body183 ret i32 0184}185 186define i32 @const_trip_over_profile() {187; constant trip count takes precedence over profile data188 189; CHECK-LABEL: @const_trip_over_profile190; CHECK: <{{[0-9]+}} x i8>191 192entry:193 br label %for.body194 195for.body: ; preds = %for.body, %entry196 %i.08 = phi i32 [ 0, %entry ], [ %inc, %for.body ]197 %arrayidx = getelementptr inbounds [32 x i8], ptr @tab, i32 0, i32 %i.08198 %0 = load i8, ptr %arrayidx, align 1199 %cmp1 = icmp eq i8 %0, 0200 %. = select i1 %cmp1, i8 2, i8 1201 store i8 %., ptr %arrayidx, align 1202 %inc = add nsw i32 %i.08, 1203 %exitcond = icmp slt i32 %i.08, 1000204 br i1 %exitcond, label %for.body, label %for.end, !prof !1205 206for.end: ; preds = %for.body207 ret i32 0208}209 210; CHECK: [[LP3]] = !{!"branch_weights", i32 10, i32 2490}211; CHECK: [[LP6]] = !{!"branch_weights", i32 10, i32 0}212; original loop has latchExitWeight=10 and backedgeTakenWeight=10,000,213; therefore estimatedBackedgeTakenCount=1,000 and estimatedTripCount=1,001.214; Vectorizing by 4 produces estimatedTripCounts of 1,001/4=250 and 1,001%4=1215; for vectorized and remainder loops, respectively, therefore their216; estimatedBackedgeTakenCounts are 249 and 0, and so the weights recorded with217; loop invocation weights of 10 are the above {10, 2490} and {10, 0}.218 219!0 = !{!"function_entry_count", i64 100}220!1 = !{!"branch_weights", i32 100, i32 0}221!2 = !{!"branch_weights", i32 10, i32 90}222!3 = !{!"branch_weights", i32 10, i32 10000}223