426 lines · plain
1; RUN: opt -mtriple=thumbv8.1m.main-none-none-eabi -passes=hardware-loops %s -S -o - | \2; RUN: FileCheck %s3; RUN: opt -mtriple=thumbv8.1m.main -passes=loop-unroll -unroll-remainder=false -S < %s | \4; RUN: llc -mtriple=thumbv8.1m.main | FileCheck %s --check-prefix=CHECK-UNROLL5; RUN: opt -mtriple=thumbv8.1m.main-none-none-eabi -passes=hardware-loops \6; RUN: -pass-remarks-analysis=hardware-loops %s -S -o - 2>&1 | \7; RUN: FileCheck %s --check-prefix=CHECK-REMARKS8 9 10; CHECK-REMARKS: remark: <unknown>:0:0: hardware-loop not created: it's not profitable to create a hardware-loop11; CHECK-REMARKS: remark: <unknown>:0:0: hardware-loop not created: nested hardware-loops not supported12; CHECK-REMARKS: remark: <unknown>:0:0: hardware-loop not created: it's not profitable to create a hardware-loop13; CHECK-REMARKS: remark: <unknown>:0:0: hardware-loop not created: it's not profitable to create a hardware-loop14; CHECK-REMARKS: remark: <unknown>:0:0: hardware-loop not created: it's not profitable to create a hardware-loop15; CHECK-REMARKS: remark: <unknown>:0:0: hardware-loop not created: it's not profitable to create a hardware-loop16; CHECK-REMARKS: remark: <unknown>:0:0: hardware-loop not created: loop is not a candidate17; CHECK-REMARKS: remark: <unknown>:0:0: hardware-loop not created: loop is not a candidate18; CHECK-REMARKS: remark: <unknown>:0:0: hardware-loop not created: it's not profitable to create a hardware-loop19; CHECK-REMARKS: remark: <unknown>:0:0: hardware-loop not created: it's not profitable to create a hardware-loop20 21 22; CHECK-LABEL: early_exit23; CHECK-NOT: llvm.set.loop.iterations24; CHECK-NOT: llvm.loop.decrement25define i32 @early_exit(ptr nocapture readonly %a, i32 %max, i32 %n) {26entry:27 br label %do.body28 29do.body:30 %i.0 = phi i32 [ 0, %entry ], [ %inc, %if.end ]31 %arrayidx = getelementptr inbounds i32, ptr %a, i32 %i.032 %0 = load i32, ptr %arrayidx, align 433 %cmp = icmp sgt i32 %0, %max34 br i1 %cmp, label %do.end, label %if.end35 36if.end:37 %inc = add nuw i32 %i.0, 138 %cmp1 = icmp ult i32 %inc, %n39 br i1 %cmp1, label %do.body, label %if.end.do.end_crit_edge40 41if.end.do.end_crit_edge:42 %arrayidx2.phi.trans.insert = getelementptr inbounds i32, ptr %a, i32 %inc43 %.pre = load i32, ptr %arrayidx2.phi.trans.insert, align 444 br label %do.end45 46do.end:47 %1 = phi i32 [ %.pre, %if.end.do.end_crit_edge ], [ %0, %do.body ]48 ret i32 %149}50 51; CHECK-LABEL: nested52; CHECK-NOT: call i32 @llvm.start.loop.iterations.i32(i32 %N)53; CHECK: br i1 %cmp20, label %while.end7, label %while.cond1.preheader.us54 55; CHECK: [[START:%[^ ]+]] = call i32 @llvm.start.loop.iterations.i32(i32 %N)56; CHECK: br label %while.body3.us57 58; CHECK: [[REM:%[^ ]+]] = phi i32 [ [[START]], %while.cond1.preheader.us ], [ [[LOOP_DEC:%[^ ]+]], %while.body3.us ]59; CHECK: [[LOOP_DEC]] = call i32 @llvm.loop.decrement.reg.i32(i32 [[REM]], i32 1)60; CHECK: [[CMP:%[^ ]+]] = icmp ne i32 [[LOOP_DEC]], 061; CHECK: br i1 [[CMP]], label %while.body3.us, label %while.cond1.while.end_crit_edge.us62 63; CHECK-NOT: %{{[^ ]+}} = call i1 @llvm.loop.decrement.i32(i32 1)64 65define void @nested(ptr nocapture %A, i32 %N) {66entry:67 %cmp20 = icmp eq i32 %N, 068 br i1 %cmp20, label %while.end7, label %while.cond1.preheader.us69 70while.cond1.preheader.us:71 %i.021.us = phi i32 [ %inc6.us, %while.cond1.while.end_crit_edge.us ], [ 0, %entry ]72 %mul.us = mul i32 %i.021.us, %N73 br label %while.body3.us74 75while.body3.us:76 %j.019.us = phi i32 [ 0, %while.cond1.preheader.us ], [ %inc.us, %while.body3.us ]77 %add.us = add i32 %j.019.us, %mul.us78 %arrayidx.us = getelementptr inbounds i32, ptr %A, i32 %add.us79 store i32 %add.us, ptr %arrayidx.us, align 480 %inc.us = add nuw i32 %j.019.us, 181 %exitcond = icmp eq i32 %inc.us, %N82 br i1 %exitcond, label %while.cond1.while.end_crit_edge.us, label %while.body3.us83 84while.cond1.while.end_crit_edge.us:85 %inc6.us = add nuw i32 %i.021.us, 186 %exitcond23 = icmp eq i32 %inc6.us, %N87 br i1 %exitcond23, label %while.end7, label %while.cond1.preheader.us88 89while.end7:90 ret void91}92 93; CHECK-LABEL: pre_existing94; CHECK: llvm.start.loop.iterations95; CHECK-NOT: llvm.start.loop.iterations96; CHECK: call i32 @llvm.loop.decrement.reg.i32(i32 %0, i32 1)97; CHECK-NOT: call i32 @llvm.loop.decrement.reg98define i32 @pre_existing(i32 %n, ptr nocapture %p, ptr nocapture readonly %q) {99entry:100 %start = call i32 @llvm.start.loop.iterations.i32(i32 %n)101 br label %while.body102 103while.body: ; preds = %while.body, %entry104 %q.addr.05 = phi ptr [ %incdec.ptr, %while.body ], [ %q, %entry ]105 %p.addr.04 = phi ptr [ %incdec.ptr1, %while.body ], [ %p, %entry ]106 %0 = phi i32 [ %start, %entry ], [ %2, %while.body ]107 %incdec.ptr = getelementptr inbounds i32, ptr %q.addr.05, i32 1108 %1 = load i32, ptr %q.addr.05, align 4109 %incdec.ptr1 = getelementptr inbounds i32, ptr %p.addr.04, i32 1110 store i32 %1, ptr %p.addr.04, align 4111 %2 = call i32 @llvm.loop.decrement.reg.i32(i32 %0, i32 1)112 %3 = icmp ne i32 %2, 0113 br i1 %3, label %while.body, label %while.end114 115while.end: ; preds = %while.body116 ret i32 0117}118 119; CHECK-LABEL: pre_existing_test_set120; CHECK: call { i32, i1 } @llvm.test.start.loop.iterations121; CHECK-NOT: llvm.set{{.*}}.loop.iterations122; CHECK: call i32 @llvm.loop.decrement.reg.i32(i32 %0, i32 1)123; CHECK-NOT: call i32 @llvm.loop.decrement.reg124define i32 @pre_existing_test_set(i32 %n, ptr nocapture %p, ptr nocapture readonly %q) {125entry:126 %guard = call { i32, i1 } @llvm.test.start.loop.iterations.i32(i32 %n)127 %g0 = extractvalue { i32, i1 } %guard, 0128 %g1 = extractvalue { i32, i1 } %guard, 1129 br i1 %g1, label %while.preheader, label %while.end130 131while.preheader:132 br label %while.body133 134while.body: ; preds = %while.body, %entry135 %q.addr.05 = phi ptr [ %incdec.ptr, %while.body ], [ %q, %while.preheader ]136 %p.addr.04 = phi ptr [ %incdec.ptr1, %while.body ], [ %p, %while.preheader ]137 %0 = phi i32 [ %g0, %while.preheader ], [ %2, %while.body ]138 %incdec.ptr = getelementptr inbounds i32, ptr %q.addr.05, i32 1139 %1 = load i32, ptr %q.addr.05, align 4140 %incdec.ptr1 = getelementptr inbounds i32, ptr %p.addr.04, i32 1141 store i32 %1, ptr %p.addr.04, align 4142 %2 = call i32 @llvm.loop.decrement.reg.i32(i32 %0, i32 1)143 %3 = icmp ne i32 %2, 0144 br i1 %3, label %while.body, label %while.end145 146while.end: ; preds = %while.body147 ret i32 0148}149 150; CHECK-LABEL: pre_existing_inner151; CHECK-NOT: llvm.start.loop.iterations152; CHECK: while.cond1.preheader.us:153; CHECK: call i32 @llvm.start.loop.iterations.i32(i32 %N)154; CHECK: call i32 @llvm.loop.decrement.reg.i32(i32 %0, i32 1)155; CHECK: br i1156; CHECK-NOT: call i32 @llvm.loop.decrement157define void @pre_existing_inner(ptr nocapture %A, i32 %N) {158entry:159 %cmp20 = icmp eq i32 %N, 0160 br i1 %cmp20, label %while.end7, label %while.cond1.preheader.us161 162while.cond1.preheader.us:163 %i.021.us = phi i32 [ %inc6.us, %while.cond1.while.end_crit_edge.us ], [ 0, %entry ]164 %mul.us = mul i32 %i.021.us, %N165 %start = call i32 @llvm.start.loop.iterations.i32(i32 %N)166 br label %while.body3.us167 168while.body3.us:169 %j.019.us = phi i32 [ 0, %while.cond1.preheader.us ], [ %inc.us, %while.body3.us ]170 %0 = phi i32 [ %start, %while.cond1.preheader.us ], [ %1, %while.body3.us ]171 %add.us = add i32 %j.019.us, %mul.us172 %arrayidx.us = getelementptr inbounds i32, ptr %A, i32 %add.us173 store i32 %add.us, ptr %arrayidx.us, align 4174 %inc.us = add nuw i32 %j.019.us, 1175 %1 = call i32 @llvm.loop.decrement.reg.i32(i32 %0, i32 1)176 %2 = icmp ne i32 %1, 0177 br i1 %2, label %while.body3.us, label %while.cond1.while.end_crit_edge.us178 179while.cond1.while.end_crit_edge.us:180 %inc6.us = add nuw i32 %i.021.us, 1181 %exitcond23 = icmp eq i32 %inc6.us, %N182 br i1 %exitcond23, label %while.end7, label %while.cond1.preheader.us183 184while.end7:185 ret void186}187 188; CHECK-LABEL: not_rotated189; CHECK-NOT: call i32 @llvm.start.loop.iterations190; CHECK-NOT: call i32 @llvm.loop.decrement.i32191define void @not_rotated(i32, ptr nocapture, i16 signext) {192 br label %4193 1944:195 %5 = phi i32 [ 0, %3 ], [ %19, %18 ]196 %6 = icmp eq i32 %5, %0197 br i1 %6, label %20, label %7198 1997:200 %8 = mul i32 %5, %0201 br label %9202 2039:204 %10 = phi i32 [ %17, %12 ], [ 0, %7 ]205 %11 = icmp eq i32 %10, %0206 br i1 %11, label %18, label %12207 20812:209 %13 = add i32 %10, %8210 %14 = getelementptr inbounds i16, ptr %1, i32 %13211 %15 = load i16, ptr %14, align 2212 %16 = add i16 %15, %2213 store i16 %16, ptr %14, align 2214 %17 = add i32 %10, 1215 br label %9216 21718:218 %19 = add i32 %5, 1219 br label %4220 22120:222 ret void223}224 225; CHECK-LABEL: multi_latch226; CHECK-NOT: call i32 @llvm.start.loop.iterations227; CHECK-NOT: call i32 @llvm.loop.decrement228define void @multi_latch(ptr %a, ptr %b, i32 %N) {229entry:230 %half = lshr i32 %N, 1231 br label %header232 233header:234 %iv = phi i32 [ 0, %entry ], [ %count.next, %latch.0 ], [ %count.next, %latch.1 ]235 %cmp = icmp ult i32 %iv, %half236 %addr.a = getelementptr i32, ptr %a, i32 %iv237 %addr.b = getelementptr i32, ptr %b, i32 %iv238 br i1 %cmp, label %if.then, label %if.else239 240if.then:241 store i32 %iv, ptr %addr.a242 br label %latch.0243 244if.else:245 store i32 %iv, ptr %addr.b246 br label %latch.0247 248latch.0:249 %count.next = add nuw i32 %iv, 1250 %cmp.1 = icmp ult i32 %count.next, %half251 br i1 %cmp.1, label %header, label %latch.1252 253latch.1:254 %ld = load i32, ptr %addr.a255 store i32 %ld, ptr %addr.b256 %cmp.2 = icmp ult i32 %count.next, %N257 br i1 %cmp.2, label %header, label %latch.1258 259exit:260 ret void261}262 263; CHECK-LABEL: search264; CHECK: entry:265; CHECK: [[TEST1:%[^ ]+]] = call { i32, i1 } @llvm.test.start.loop.iterations.i32(i32 %N)266; CHECK: [[TEST:%[^ ]+]] = extractvalue { i32, i1 } [[TEST1]], 1267; CHECK: br i1 [[TEST]], label %for.body.preheader, label %for.cond.cleanup268; CHECK: for.body.preheader:269; CHECK: br label %for.body270; CHECK: for.body:271; CHECK: for.inc:272; CHECK: [[LOOP_DEC:%[^ ]+]] = call i32 @llvm.loop.decrement.reg.i32(273; CHECK: [[CMP:%[^ ]+]] = icmp ne i32 [[LOOP_DEC]], 0274; CHECK: br i1 [[CMP]], label %for.body, label %for.cond.cleanup275define i32 @search(ptr nocapture readonly %c, i32 %N) {276entry:277 %cmp11 = icmp eq i32 %N, 0278 br i1 %cmp11, label %for.cond.cleanup, label %for.body279 280for.cond.cleanup:281 %found.0.lcssa = phi i32 [ 0, %entry ], [ %found.1, %for.inc ]282 %spaces.0.lcssa = phi i32 [ 0, %entry ], [ %spaces.1, %for.inc ]283 %sub = sub nsw i32 %found.0.lcssa, %spaces.0.lcssa284 ret i32 %sub285 286for.body:287 %i.014 = phi i32 [ %inc3, %for.inc ], [ 0, %entry ]288 %spaces.013 = phi i32 [ %spaces.1, %for.inc ], [ 0, %entry ]289 %found.012 = phi i32 [ %found.1, %for.inc ], [ 0, %entry ]290 %arrayidx = getelementptr inbounds i8, ptr %c, i32 %i.014291 %0 = load i8, ptr %arrayidx, align 1292 switch i8 %0, label %for.inc [293 i8 108, label %sw.bb294 i8 111, label %sw.bb295 i8 112, label %sw.bb296 i8 32, label %sw.bb1297 ]298 299sw.bb: ; preds = %for.body, %for.body, %for.body300 %inc = add nsw i32 %found.012, 1301 br label %for.inc302 303sw.bb1: ; preds = %for.body304 %inc2 = add nsw i32 %spaces.013, 1305 br label %for.inc306 307for.inc: ; preds = %sw.bb, %sw.bb1, %for.body308 %found.1 = phi i32 [ %found.012, %for.body ], [ %found.012, %sw.bb1 ], [ %inc, %sw.bb ]309 %spaces.1 = phi i32 [ %spaces.013, %for.body ], [ %inc2, %sw.bb1 ], [ %spaces.013, %sw.bb ]310 %inc3 = add nuw i32 %i.014, 1311 %exitcond = icmp eq i32 %inc3, %N312 br i1 %exitcond, label %for.cond.cleanup, label %for.body313}314 315; CHECK-LABEL: unroll_inc_int316; CHECK: call i32 @llvm.start.loop.iterations.i32(i32 %N)317; CHECK: call i32 @llvm.loop.decrement.reg.i32(318 319; CHECK-UNROLL-LABEL: unroll_inc_int320; CHECK-UNROLL: [[PREHEADER:.LBB[0-9_]+]]: @ %for.body.preheader321; CHECK-UNROLL-NOT: dls322; CHECK-UNROLL: [[LOOP:.LBB[0-9_]+]]: @ %for.body323; CHECK-UNROLL: le lr, [[LOOP]]324; CHECK-UNROLL: dls lr, r12325; CHECK-UNROLL: [[EPIL:.LBB[0-9_]+]]:326; CHECK-UNROLL: le lr, [[EPIL]]327; CHECK-UNROLL-NEXT: {{\.LBB[0-9_]+}}: @ %for.cond.cleanup328 329define void @unroll_inc_int(ptr nocapture %a, ptr nocapture readonly %b, ptr nocapture readonly %c, i32 %N) {330entry:331 %cmp8 = icmp sgt i32 %N, 0332 br i1 %cmp8, label %for.body, label %for.cond.cleanup333 334for.cond.cleanup:335 ret void336 337for.body:338 %i.09 = phi i32 [ %inc, %for.body ], [ 0, %entry ]339 %arrayidx = getelementptr inbounds i32, ptr %b, i32 %i.09340 %0 = load i32, ptr %arrayidx, align 4341 %arrayidx1 = getelementptr inbounds i32, ptr %c, i32 %i.09342 %1 = load i32, ptr %arrayidx1, align 4343 %mul = mul nsw i32 %1, %0344 %arrayidx2 = getelementptr inbounds i32, ptr %a, i32 %i.09345 store i32 %mul, ptr %arrayidx2, align 4346 %inc = add nuw nsw i32 %i.09, 1347 %exitcond = icmp eq i32 %inc, %N348 br i1 %exitcond, label %for.cond.cleanup, label %for.body349}350 351; CHECK-LABEL: unroll_inc_unsigned352; CHECK: call { i32, i1 } @llvm.test.start.loop.iterations.i32(i32 %N)353; CHECK: call i32 @llvm.loop.decrement.reg.i32(354 355; CHECK-UNROLL-LABEL: unroll_inc_unsigned356; CHECK-UNROLL: [[PREHEADER:.LBB[0-9_]+]]: @ %for.body.preheader357; CHECK-UNROLL-NOT: dls358; CHECK-UNROLL: [[LOOP:.LBB[0-9_]+]]: @ %for.body359; CHECK-UNROLL: le lr, [[LOOP]]360; CHECK-UNROLL: dls lr, r12361; CHECK-UNROLL: [[EPIL:.LBB[0-9_]+]]:362; CHECK-UNROLL: le lr, [[EPIL]]363; CHECK-UNROLL: {{\.LBB[0-9_]+}}: @ %for.cond.cleanup364; CHECK-UNROLL: pop365define void @unroll_inc_unsigned(ptr nocapture %a, ptr nocapture readonly %b, ptr nocapture readonly %c, i32 %N) {366entry:367 %cmp8 = icmp eq i32 %N, 0368 br i1 %cmp8, label %for.cond.cleanup, label %for.body369 370for.cond.cleanup:371 ret void372 373for.body:374 %i.09 = phi i32 [ %inc, %for.body ], [ 0, %entry ]375 %arrayidx = getelementptr inbounds i32, ptr %b, i32 %i.09376 %0 = load i32, ptr %arrayidx, align 4377 %arrayidx1 = getelementptr inbounds i32, ptr %c, i32 %i.09378 %1 = load i32, ptr %arrayidx1, align 4379 %mul = mul nsw i32 %1, %0380 %arrayidx2 = getelementptr inbounds i32, ptr %a, i32 %i.09381 store i32 %mul, ptr %arrayidx2, align 4382 %inc = add nuw i32 %i.09, 1383 %exitcond = icmp eq i32 %inc, %N384 br i1 %exitcond, label %for.cond.cleanup, label %for.body385}386 387; CHECK-LABEL: unroll_dec_int388; CHECK: call i32 @llvm.start.loop.iterations.i32(i32 %N)389; CHECK: call i32 @llvm.loop.decrement.reg.i32(390 391; CHECK-UNROLL-LABEL: unroll_dec_int:392; CHECK-UNROLL: wls lr, {{.*}}, [[PROLOGUE_EXIT:.LBB[0-9_]+]]393; CHECK-UNROLL-NEXT: [[PROLOGUE:.LBB[0-9_]+]]:394; CHECK-UNROLL: le lr, [[PROLOGUE]]395; CHECK-UNROLL-NEXT: [[PROLOGUE_EXIT:.LBB[0-9_]+]]:396; CHECK-UNROLL: [[BODY:.LBB[0-9_]+]]:397; CHECK-UNROLL: le lr, [[BODY]]398; CHECK-UNROLL-NOT: b399; CHECK-UNROLL: pop400define void @unroll_dec_int(ptr nocapture %a, ptr nocapture readonly %b, ptr nocapture readonly %c, i32 %N) {401entry:402 %cmp8 = icmp sgt i32 %N, 0403 br i1 %cmp8, label %for.body, label %for.cond.cleanup404 405for.cond.cleanup:406 ret void407 408for.body:409 %i.09 = phi i32 [ %dec, %for.body ], [ %N, %entry ]410 %arrayidx = getelementptr inbounds i32, ptr %b, i32 %i.09411 %0 = load i32, ptr %arrayidx, align 4412 %arrayidx1 = getelementptr inbounds i32, ptr %c, i32 %i.09413 %1 = load i32, ptr %arrayidx1, align 4414 %mul = mul nsw i32 %1, %0415 %arrayidx2 = getelementptr inbounds i32, ptr %a, i32 %i.09416 store i32 %mul, ptr %arrayidx2, align 4417 %dec = add nsw i32 %i.09, -1418 %cmp = icmp sgt i32 %dec, 0419 br i1 %cmp, label %for.body, label %for.cond.cleanup420}421 422declare i32 @llvm.start.loop.iterations.i32(i32) #0423declare { i32, i1 } @llvm.test.start.loop.iterations.i32(i32) #0424declare i32 @llvm.loop.decrement.reg.i32(i32, i32) #0425 426