1; RUN: opt -mtriple=thumbv8.1m.main -mattr=+mve.fp -loop-unroll -S < %s -o - | FileCheck %s 2 3; CHECK-LABEL: @loopfn 4; CHECK: vector.body: 5; CHECK: br i1 %7, label %middle.block, label %vector.body, !llvm.loop !0 6; CHECK: middle.block: 7; CHECK: br i1 %cmp.n, label %for.cond.cleanup, label %for.body.preheader13 8; CHECK: for.body: 9; CHECK: br i1 %exitcond, label %for.cond.cleanup.loopexit, label %for.body, !llvm.loop !2 10 11define void @loopfn(float* %s1, float* %s2, float* %d, i32 %n) { 12entry: 13 %cmp10 = icmp sgt i32 %n, 0 14 br i1 %cmp10, label %for.body.preheader, label %for.cond.cleanup 15 16for.body.preheader: ; preds = %entry 17 %min.iters.check = icmp ult i32 %n, 4 18 br i1 %min.iters.check, label %for.body.preheader13, label %vector.ph 19 20for.body.preheader13: ; preds = %middle.block, %for.body.preheader 21 %i.011.ph = phi i32 [ 0, %for.body.preheader ], [ %n.vec, %middle.block ] 22 br label %for.body 23 24vector.ph: ; preds = %for.body.preheader 25 %n.vec = and i32 %n, -4 26 br label %vector.body 27 28vector.body: ; preds = %vector.body, %vector.ph 29 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ] 30 %0 = getelementptr inbounds float, float* %s1, i32 %index 31 %1 = bitcast float* %0 to <4 x float>* 32 %wide.load = load <4 x float>, <4 x float>* %1, align 4 33 %2 = getelementptr inbounds float, float* %s2, i32 %index 34 %3 = bitcast float* %2 to <4 x float>* 35 %wide.load12 = load <4 x float>, <4 x float>* %3, align 4 36 %4 = fadd fast <4 x float> %wide.load12, %wide.load 37 %5 = getelementptr inbounds float, float* %d, i32 %index 38 %6 = bitcast float* %5 to <4 x float>* 39 store <4 x float> %4, <4 x float>* %6, align 4 40 %index.next = add i32 %index, 4 41 %7 = icmp eq i32 %index.next, %n.vec 42 br i1 %7, label %middle.block, label %vector.body, !llvm.loop !0 43 44middle.block: ; preds = %vector.body 45 %cmp.n = icmp eq i32 %n.vec, %n 46 br i1 %cmp.n, label %for.cond.cleanup, label %for.body.preheader13 47 48for.cond.cleanup.loopexit: ; preds = %for.body 49 br label %for.cond.cleanup 50 51for.cond.cleanup: ; preds = %for.cond.cleanup.loopexit, %middle.block, %entry 52 ret void 53 54for.body: ; preds = %for.body.preheader13, %for.body 55 %i.011 = phi i32 [ %add3, %for.body ], [ %i.011.ph, %for.body.preheader13 ] 56 %arrayidx = getelementptr inbounds float, float* %s1, i32 %i.011 57 %8 = load float, float* %arrayidx, align 4 58 %arrayidx1 = getelementptr inbounds float, float* %s2, i32 %i.011 59 %9 = load float, float* %arrayidx1, align 4 60 %add = fadd fast float %9, %8 61 %arrayidx2 = getelementptr inbounds float, float* %d, i32 %i.011 62 store float %add, float* %arrayidx2, align 4 63 %add3 = add nuw nsw i32 %i.011, 1 64 %exitcond = icmp eq i32 %add3, %n 65 br i1 %exitcond, label %for.cond.cleanup.loopexit, label %for.body, !llvm.loop !2 66} 67 68 69; Same as above but without the nounroll on the remainder loop. Neither loop should be unrolled. 70 71; CHECK-LABEL: @remainder 72; CHECK: vector.body: 73; CHECK: br i1 %7, label %middle.block, label %vector.body, !llvm.loop !0 74; CHECK: middle.block: 75; CHECK: br i1 %cmp.n, label %for.cond.cleanup, label %for.body.preheader13 76; CHECK: for.body: 77; CHECK: br i1 %exitcond, label %for.cond.cleanup.loopexit, label %for.body, !llvm.loop !0 78 79define void @remainder(float* %s1, float* %s2, float* %d, i32 %n) { 80entry: 81 %cmp10 = icmp sgt i32 %n, 0 82 br i1 %cmp10, label %for.body.preheader, label %for.cond.cleanup 83 84for.body.preheader: ; preds = %entry 85 %min.iters.check = icmp ult i32 %n, 4 86 br i1 %min.iters.check, label %for.body.preheader13, label %vector.ph 87 88for.body.preheader13: ; preds = %middle.block, %for.body.preheader 89 %i.011.ph = phi i32 [ 0, %for.body.preheader ], [ %n.vec, %middle.block ] 90 br label %for.body 91 92vector.ph: ; preds = %for.body.preheader 93 %n.vec = and i32 %n, -4 94 br label %vector.body 95 96vector.body: ; preds = %vector.body, %vector.ph 97 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ] 98 %0 = getelementptr inbounds float, float* %s1, i32 %index 99 %1 = bitcast float* %0 to <4 x float>* 100 %wide.load = load <4 x float>, <4 x float>* %1, align 4 101 %2 = getelementptr inbounds float, float* %s2, i32 %index 102 %3 = bitcast float* %2 to <4 x float>* 103 %wide.load12 = load <4 x float>, <4 x float>* %3, align 4 104 %4 = fadd fast <4 x float> %wide.load12, %wide.load 105 %5 = getelementptr inbounds float, float* %d, i32 %index 106 %6 = bitcast float* %5 to <4 x float>* 107 store <4 x float> %4, <4 x float>* %6, align 4 108 %index.next = add i32 %index, 4 109 %7 = icmp eq i32 %index.next, %n.vec 110 br i1 %7, label %middle.block, label %vector.body, !llvm.loop !0 111 112middle.block: ; preds = %vector.body 113 %cmp.n = icmp eq i32 %n.vec, %n 114 br i1 %cmp.n, label %for.cond.cleanup, label %for.body.preheader13 115 116for.cond.cleanup.loopexit: ; preds = %for.body 117 br label %for.cond.cleanup 118 119for.cond.cleanup: ; preds = %for.cond.cleanup.loopexit, %middle.block, %entry 120 ret void 121 122for.body: ; preds = %for.body.preheader13, %for.body 123 %i.011 = phi i32 [ %add3, %for.body ], [ %i.011.ph, %for.body.preheader13 ] 124 %arrayidx = getelementptr inbounds float, float* %s1, i32 %i.011 125 %8 = load float, float* %arrayidx, align 4 126 %arrayidx1 = getelementptr inbounds float, float* %s2, i32 %i.011 127 %9 = load float, float* %arrayidx1, align 4 128 %add = fadd fast float %9, %8 129 %arrayidx2 = getelementptr inbounds float, float* %d, i32 %i.011 130 store float %add, float* %arrayidx2, align 4 131 %add3 = add nuw nsw i32 %i.011, 1 132 %exitcond = icmp eq i32 %add3, %n 133 br i1 %exitcond, label %for.cond.cleanup.loopexit, label %for.body, !llvm.loop !0 134} 135 136 137 138; CHECK-LABEL: @nested 139; CHECK: for.outer: 140; CHECK: br label %vector.body 141; CHECK: vector.body: 142; CHECK: br i1 %8, label %for.latch, label %vector.body, !llvm.loop !0 143; CHECK: for.latch: 144; CHECK: br i1 %exitcond34, label %for.cond.cleanup.loopexit, label %for.outer 145 146define void @nested(float* %s1, float* %s2, float* %d, i32 %n) { 147entry: 148 %cmp31 = icmp eq i32 %n, 0 149 br i1 %cmp31, label %for.cond.cleanup, label %for.outer.preheader 150 151for.outer.preheader: ; preds = %entry 152 %min.iters.check = icmp ult i32 %n, 4 153 %n.vec = and i32 %n, -4 154 %cmp.n = icmp eq i32 %n.vec, %n 155 br label %for.outer 156 157for.outer: ; preds = %for.outer.preheader, %for.cond1.for.cond.cleanup3_crit_edge.us 158 %j.032.us = phi i32 [ %inc.us, %for.latch ], [ 0, %for.outer.preheader ] 159 %mul.us = mul i32 %j.032.us, %n 160 br label %vector.body 161 162vector.body: ; preds = %for.outer, %vector.body 163 %index = phi i32 [ %index.next, %vector.body ], [ 0, %for.outer ] 164 %0 = add i32 %index, %mul.us 165 %1 = getelementptr inbounds float, float* %s1, i32 %0 166 %2 = bitcast float* %1 to <4 x float>* 167 %wide.load = load <4 x float>, <4 x float>* %2, align 4 168 %3 = getelementptr inbounds float, float* %s2, i32 %0 169 %4 = bitcast float* %3 to <4 x float>* 170 %wide.load35 = load <4 x float>, <4 x float>* %4, align 4 171 %5 = fadd fast <4 x float> %wide.load35, %wide.load 172 %6 = getelementptr inbounds float, float* %d, i32 %0 173 %7 = bitcast float* %6 to <4 x float>* 174 store <4 x float> %5, <4 x float>* %7, align 4 175 %index.next = add i32 %index, 4 176 %8 = icmp eq i32 %index.next, %n.vec 177 br i1 %8, label %for.latch, label %vector.body, !llvm.loop !0 178 179for.latch: ; preds = %vector.body, %for.outer 180 %i.030.us.ph = phi i32 [ %n.vec, %vector.body ] 181 %inc.us = add nuw i32 %j.032.us, 1 182 %exitcond34 = icmp eq i32 %inc.us, %n 183 br i1 %exitcond34, label %for.cond.cleanup.loopexit, label %for.outer 184 185for.cond.cleanup.loopexit: 186 br label %for.cond.cleanup 187 188for.cond.cleanup: ; preds = %for.cond1.for.cond.cleanup3_crit_edge.us, %entry 189 ret void 190} 191 192; Test that we don't unroll loops that only contain vector intrinsics. 193; CHECK-LABEL: test_intrinsics 194; CHECK: call <16 x i8> @llvm.arm.mve.sub 195; CHECK-NOT: call <16 x i8> @llvm.arm.mve.sub 196define dso_local arm_aapcs_vfpcc void @test_intrinsics(i8* noalias nocapture readonly %a, i8* noalias nocapture readonly %b, i8* noalias nocapture %c, i32 %N) { 197entry: 198 %cmp8 = icmp eq i32 %N, 0 199 %tmp8 = add i32 %N, 15 200 %tmp9 = lshr i32 %tmp8, 4 201 %tmp10 = shl nuw i32 %tmp9, 4 202 %tmp11 = add i32 %tmp10, -16 203 %tmp12 = lshr i32 %tmp11, 4 204 %tmp13 = add nuw nsw i32 %tmp12, 1 205 br i1 %cmp8, label %for.cond.cleanup, label %vector.ph 206 207vector.ph: 208 br label %vector.body 209 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, i8* %a, i32 %index 215 %1 = call <16 x i1> @llvm.arm.mve.vctp8(i32 %0) 216 %2 = sub i32 %0, 16 217 %tmp2 = bitcast i8* %tmp to <16 x i8>* 218 %wide.masked.load = tail call <16 x i8> @llvm.masked.load.v16i8.p0v16i8(<16 x i8>* %tmp2, i32 4, <16 x i1> %1, <16 x i8> undef) 219 %tmp3 = getelementptr inbounds i8, i8* %b, i32 %index 220 %tmp4 = bitcast i8* %tmp3 to <16 x i8>* 221 %wide.masked.load2 = tail call <16 x i8> @llvm.masked.load.v16i8.p0v16i8(<16 x i8>* %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, i8* %c, i32 %index 224 %tmp7 = bitcast i8* %tmp6 to <16 x i8>* 225 tail call void @llvm.masked.store.v16i8.p0v16i8(<16 x i8> %sub, <16 x i8>* %tmp7, i32 4, <16 x i1> %1) 226 %index.next = add i32 %index, 16 227 %tmp15 = sub i32 %tmp14, 1 228 %tmp16 = icmp ne i32 %tmp15, 0 229 br i1 %tmp16, label %vector.body, label %for.cond.cleanup 230 231for.cond.cleanup: ; preds = %vector.body, %entry 232 ret void 233} 234 235declare <16 x i1> @llvm.arm.mve.vctp8(i32) 236declare <16 x i8> @llvm.masked.load.v16i8.p0v16i8(<16 x i8>*, 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.p0v16i8(<16 x i8>, <16 x i8>*, 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