245 lines · plain
1; RUN: opt -mtriple=thumbv8.1m.main -mattr=+mve.fp -passes=loop-unroll -S < %s -o - | FileCheck %s2 3; CHECK-LABEL: @loopfn4; CHECK: vector.body:5; CHECK: br i1 %7, label %middle.block, label %vector.body, !llvm.loop !06; CHECK: middle.block:7; CHECK: br i1 %cmp.n, label %for.cond.cleanup, label %for.body.preheader138; CHECK: for.body:9; CHECK: br i1 %exitcond, label %for.cond.cleanup.loopexit, label %for.body, !llvm.loop !210 11define void @loopfn(ptr %s1, ptr %s2, ptr %d, i32 %n) {12entry:13 %cmp10 = icmp sgt i32 %n, 014 br i1 %cmp10, label %for.body.preheader, label %for.cond.cleanup15 16for.body.preheader: ; preds = %entry17 %min.iters.check = icmp ult i32 %n, 418 br i1 %min.iters.check, label %for.body.preheader13, label %vector.ph19 20for.body.preheader13: ; preds = %middle.block, %for.body.preheader21 %i.011.ph = phi i32 [ 0, %for.body.preheader ], [ %n.vec, %middle.block ]22 br label %for.body23 24vector.ph: ; preds = %for.body.preheader25 %n.vec = and i32 %n, -426 br label %vector.body27 28vector.body: ; preds = %vector.body, %vector.ph29 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]30 %0 = getelementptr inbounds float, ptr %s1, i32 %index31 %1 = bitcast ptr %0 to ptr32 %wide.load = load <4 x float>, ptr %1, align 433 %2 = getelementptr inbounds float, ptr %s2, i32 %index34 %3 = bitcast ptr %2 to ptr35 %wide.load12 = load <4 x float>, ptr %3, align 436 %4 = fadd fast <4 x float> %wide.load12, %wide.load37 %5 = getelementptr inbounds float, ptr %d, i32 %index38 %6 = bitcast ptr %5 to ptr39 store <4 x float> %4, ptr %6, align 440 %index.next = add i32 %index, 441 %7 = icmp eq i32 %index.next, %n.vec42 br i1 %7, label %middle.block, label %vector.body, !llvm.loop !043 44middle.block: ; preds = %vector.body45 %cmp.n = icmp eq i32 %n.vec, %n46 br i1 %cmp.n, label %for.cond.cleanup, label %for.body.preheader1347 48for.cond.cleanup.loopexit: ; preds = %for.body49 br label %for.cond.cleanup50 51for.cond.cleanup: ; preds = %for.cond.cleanup.loopexit, %middle.block, %entry52 ret void53 54for.body: ; preds = %for.body.preheader13, %for.body55 %i.011 = phi i32 [ %add3, %for.body ], [ %i.011.ph, %for.body.preheader13 ]56 %arrayidx = getelementptr inbounds float, ptr %s1, i32 %i.01157 %8 = load float, ptr %arrayidx, align 458 %arrayidx1 = getelementptr inbounds float, ptr %s2, i32 %i.01159 %9 = load float, ptr %arrayidx1, align 460 %add = fadd fast float %9, %861 %arrayidx2 = getelementptr inbounds float, ptr %d, i32 %i.01162 store float %add, ptr %arrayidx2, align 463 %add3 = add nuw nsw i32 %i.011, 164 %exitcond = icmp eq i32 %add3, %n65 br i1 %exitcond, label %for.cond.cleanup.loopexit, label %for.body, !llvm.loop !266}67 68 69; Same as above but without the nounroll on the remainder loop. Neither loop should be unrolled.70 71; CHECK-LABEL: @remainder72; CHECK: vector.body:73; CHECK: br i1 %7, label %middle.block, label %vector.body, !llvm.loop !074; CHECK: middle.block:75; CHECK: br i1 %cmp.n, label %for.cond.cleanup, label %for.body.preheader1376; CHECK: for.body:77; CHECK: br i1 %exitcond, label %for.cond.cleanup.loopexit, label %for.body, !llvm.loop !078 79define void @remainder(ptr %s1, ptr %s2, ptr %d, i32 %n) {80entry:81 %cmp10 = icmp sgt i32 %n, 082 br i1 %cmp10, label %for.body.preheader, label %for.cond.cleanup83 84for.body.preheader: ; preds = %entry85 %min.iters.check = icmp ult i32 %n, 486 br i1 %min.iters.check, label %for.body.preheader13, label %vector.ph87 88for.body.preheader13: ; preds = %middle.block, %for.body.preheader89 %i.011.ph = phi i32 [ 0, %for.body.preheader ], [ %n.vec, %middle.block ]90 br label %for.body91 92vector.ph: ; preds = %for.body.preheader93 %n.vec = and i32 %n, -494 br label %vector.body95 96vector.body: ; preds = %vector.body, %vector.ph97 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]98 %0 = getelementptr inbounds float, ptr %s1, i32 %index99 %1 = bitcast ptr %0 to ptr100 %wide.load = load <4 x float>, ptr %1, align 4101 %2 = getelementptr inbounds float, ptr %s2, i32 %index102 %3 = bitcast ptr %2 to ptr103 %wide.load12 = load <4 x float>, ptr %3, align 4104 %4 = fadd fast <4 x float> %wide.load12, %wide.load105 %5 = getelementptr inbounds float, ptr %d, i32 %index106 %6 = bitcast ptr %5 to ptr107 store <4 x float> %4, ptr %6, align 4108 %index.next = add i32 %index, 4109 %7 = icmp eq i32 %index.next, %n.vec110 br i1 %7, label %middle.block, label %vector.body, !llvm.loop !0111 112middle.block: ; preds = %vector.body113 %cmp.n = icmp eq i32 %n.vec, %n114 br i1 %cmp.n, label %for.cond.cleanup, label %for.body.preheader13115 116for.cond.cleanup.loopexit: ; preds = %for.body117 br label %for.cond.cleanup118 119for.cond.cleanup: ; preds = %for.cond.cleanup.loopexit, %middle.block, %entry120 ret void121 122for.body: ; preds = %for.body.preheader13, %for.body123 %i.011 = phi i32 [ %add3, %for.body ], [ %i.011.ph, %for.body.preheader13 ]124 %arrayidx = getelementptr inbounds float, ptr %s1, i32 %i.011125 %8 = load float, ptr %arrayidx, align 4126 %arrayidx1 = getelementptr inbounds float, ptr %s2, i32 %i.011127 %9 = load float, ptr %arrayidx1, align 4128 %add = fadd fast float %9, %8129 %arrayidx2 = getelementptr inbounds float, ptr %d, i32 %i.011130 store float %add, ptr %arrayidx2, align 4131 %add3 = add nuw nsw i32 %i.011, 1132 %exitcond = icmp eq i32 %add3, %n133 br i1 %exitcond, label %for.cond.cleanup.loopexit, label %for.body, !llvm.loop !0134}135 136 137 138; CHECK-LABEL: @nested139; CHECK: for.outer:140; CHECK: br label %vector.body141; CHECK: vector.body:142; CHECK: br i1 %8, label %for.latch, label %vector.body, !llvm.loop !0143; CHECK: for.latch:144; CHECK: br i1 %exitcond34, label %for.cond.cleanup.loopexit, label %for.outer145 146define void @nested(ptr %s1, ptr %s2, ptr %d, i32 %n) {147entry:148 %cmp31 = icmp eq i32 %n, 0149 br i1 %cmp31, label %for.cond.cleanup, label %for.outer.preheader150 151for.outer.preheader: ; preds = %entry152 %min.iters.check = icmp ult i32 %n, 4153 %n.vec = and i32 %n, -4154 %cmp.n = icmp eq i32 %n.vec, %n155 br label %for.outer156 157for.outer: ; preds = %for.outer.preheader, %for.cond1.for.cond.cleanup3_crit_edge.us158 %j.032.us = phi i32 [ %inc.us, %for.latch ], [ 0, %for.outer.preheader ]159 %mul.us = mul i32 %j.032.us, %n160 br label %vector.body161 162vector.body: ; preds = %for.outer, %vector.body163 %index = phi i32 [ %index.next, %vector.body ], [ 0, %for.outer ]164 %0 = add i32 %index, %mul.us165 %1 = getelementptr inbounds float, ptr %s1, i32 %0166 %2 = bitcast ptr %1 to ptr167 %wide.load = load <4 x float>, ptr %2, align 4168 %3 = getelementptr inbounds float, ptr %s2, i32 %0169 %4 = bitcast ptr %3 to ptr170 %wide.load35 = load <4 x float>, ptr %4, align 4171 %5 = fadd fast <4 x float> %wide.load35, %wide.load172 %6 = getelementptr inbounds float, ptr %d, i32 %0173 %7 = bitcast ptr %6 to ptr174 store <4 x float> %5, ptr %7, align 4175 %index.next = add i32 %index, 4176 %8 = icmp eq i32 %index.next, %n.vec177 br i1 %8, label %for.latch, label %vector.body, !llvm.loop !0178 179for.latch: ; preds = %vector.body, %for.outer180 %i.030.us.ph = phi i32 [ %n.vec, %vector.body ]181 %inc.us = add nuw i32 %j.032.us, 1182 %exitcond34 = icmp eq i32 %inc.us, %n183 br i1 %exitcond34, label %for.cond.cleanup.loopexit, label %for.outer184 185for.cond.cleanup.loopexit:186 br label %for.cond.cleanup187 188for.cond.cleanup: ; preds = %for.cond1.for.cond.cleanup3_crit_edge.us, %entry189 ret void190}191 192; Test that we don't unroll loops that only contain vector intrinsics.193; CHECK-LABEL: test_intrinsics194; CHECK: call <16 x i8> @llvm.arm.mve.sub195; CHECK-NOT: call <16 x i8> @llvm.arm.mve.sub196define dso_local arm_aapcs_vfpcc void @test_intrinsics(ptr noalias nocapture readonly %a, ptr noalias nocapture readonly %b, ptr noalias nocapture %c, i32 %N) {197entry:198 %cmp8 = icmp eq i32 %N, 0199 %tmp8 = add i32 %N, 15200 %tmp9 = lshr i32 %tmp8, 4201 %tmp10 = shl nuw i32 %tmp9, 4202 %tmp11 = add i32 %tmp10, -16203 %tmp12 = lshr i32 %tmp11, 4204 %tmp13 = add nuw nsw i32 %tmp12, 1205 br i1 %cmp8, label %for.cond.cleanup, label %vector.ph206 207vector.ph:208 br label %vector.body209 210vector.body:211 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]212 %tmp14 = phi i32 [ %tmp13, %vector.ph ], [ %tmp15, %vector.body ]213 %0 = phi i32 [ %N, %vector.ph ], [ %2, %vector.body ]214 %tmp = getelementptr inbounds i8, ptr %a, i32 %index215 %1 = call <16 x i1> @llvm.arm.mve.vctp8(i32 %0)216 %2 = sub i32 %0, 16217 %tmp2 = bitcast ptr %tmp to ptr218 %wide.masked.load = tail call <16 x i8> @llvm.masked.load.v16i8.p0(ptr %tmp2, i32 4, <16 x i1> %1, <16 x i8> undef)219 %tmp3 = getelementptr inbounds i8, ptr %b, i32 %index220 %tmp4 = bitcast ptr %tmp3 to ptr221 %wide.masked.load2 = tail call <16 x i8> @llvm.masked.load.v16i8.p0(ptr %tmp4, i32 4, <16 x i1> %1, <16 x i8> undef)222 %sub = call <16 x i8> @llvm.arm.mve.sub.predicated.v16i8.v16i1(<16 x i8> %wide.masked.load2, <16 x i8> %wide.masked.load, <16 x i1> %1, <16 x i8> undef)223 %tmp6 = getelementptr inbounds i8, ptr %c, i32 %index224 %tmp7 = bitcast ptr %tmp6 to ptr225 tail call void @llvm.masked.store.v16i8.p0(<16 x i8> %sub, ptr %tmp7, i32 4, <16 x i1> %1)226 %index.next = add i32 %index, 16227 %tmp15 = sub i32 %tmp14, 1228 %tmp16 = icmp ne i32 %tmp15, 0229 br i1 %tmp16, label %vector.body, label %for.cond.cleanup230 231for.cond.cleanup: ; preds = %vector.body, %entry232 ret void233}234 235declare <16 x i1> @llvm.arm.mve.vctp8(i32)236declare <16 x i8> @llvm.masked.load.v16i8.p0(ptr, i32, <16 x i1>, <16 x i8>)237declare <16 x i8> @llvm.arm.mve.sub.predicated.v16i8.v16i1(<16 x i8>, <16 x i8>, <16 x i1>, <16 x i8>)238declare void @llvm.masked.store.v16i8.p0(<16 x i8>, ptr, i32, <16 x i1>)239 240 241!0 = distinct !{!0, !1}242!1 = !{!"llvm.loop.isvectorized", i32 1}243!2 = distinct !{!2, !3, !1}244!3 = !{!"llvm.loop.unroll.runtime.disable"}245