238 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2;3; RUN: llc -mtriple=thumbv8.1m.main -mattr=+mve -tail-predication=enabled %s -o - | \4; RUN: FileCheck %s --check-prefix=ENABLED5;6; Forcing tail-predication should not be necessary here, so we check the same7; ENABLED label as the run above:8; RUN: llc -mtriple=thumbv8.1m.main -mattr=+mve -tail-predication=force-enabled %s -o - | \9; RUN: FileCheck %s --check-prefix=ENABLED10;11; RUN: llc -mtriple=thumbv8.1m.main -mattr=+mve -tail-predication=enabled-no-reductions %s -o - | \12; RUN: FileCheck %s --check-prefix=NOREDUCTIONS13;14; RUN: llc -mtriple=thumbv8.1m.main -mattr=+mve -tail-predication=force-enabled-no-reductions %s -o - | \15; RUN: FileCheck %s --check-prefix=NOREDUCTIONS16 17define dso_local void @varying_outer_2d_reduction(ptr nocapture readonly %Input, ptr nocapture %Output, i16 signext %Size, i16 signext %N, i16 signext %Scale) local_unnamed_addr {18; ENABLED-LABEL: varying_outer_2d_reduction:19; ENABLED: @ %bb.0: @ %entry20; ENABLED-NEXT: cmp r3, #121; ENABLED-NEXT: it lt22; ENABLED-NEXT: bxlt lr23; ENABLED-NEXT: .LBB0_1: @ %for.body.lr.ph24; ENABLED-NEXT: push.w {r4, r5, r6, r7, r9, r10, r11, lr}25; ENABLED-NEXT: mov r11, r026; ENABLED-NEXT: ldr r0, [sp, #32]27; ENABLED-NEXT: add.w r9, r2, #328; ENABLED-NEXT: mov.w r12, #029; ENABLED-NEXT: mov r10, r1130; ENABLED-NEXT: uxth r0, r031; ENABLED-NEXT: rsbs r5, r0, #032; ENABLED-NEXT: b .LBB0_433; ENABLED-NEXT: .LBB0_2: @ in Loop: Header=BB0_4 Depth=134; ENABLED-NEXT: movs r0, #035; ENABLED-NEXT: .LBB0_3: @ %for.end36; ENABLED-NEXT: @ in Loop: Header=BB0_4 Depth=137; ENABLED-NEXT: lsrs r0, r0, #1638; ENABLED-NEXT: sub.w r9, r9, #139; ENABLED-NEXT: strh.w r0, [r1, r12, lsl #1]40; ENABLED-NEXT: add.w r12, r12, #141; ENABLED-NEXT: add.w r10, r10, #242; ENABLED-NEXT: cmp r12, r343; ENABLED-NEXT: beq .LBB0_844; ENABLED-NEXT: .LBB0_4: @ %for.body45; ENABLED-NEXT: @ =>This Loop Header: Depth=146; ENABLED-NEXT: @ Child Loop BB0_6 Depth 247; ENABLED-NEXT: cmp r2, r1248; ENABLED-NEXT: ble .LBB0_249; ENABLED-NEXT: @ %bb.5: @ %vector.ph50; ENABLED-NEXT: @ in Loop: Header=BB0_4 Depth=151; ENABLED-NEXT: bic r0, r9, #352; ENABLED-NEXT: movs r7, #153; ENABLED-NEXT: subs r0, #454; ENABLED-NEXT: sub.w r4, r2, r1255; ENABLED-NEXT: vmov.i32 q1, #0x056; ENABLED-NEXT: add.w r6, r7, r0, lsr #257; ENABLED-NEXT: adds r0, r2, #358; ENABLED-NEXT: sub.w r0, r0, r1259; ENABLED-NEXT: bic r0, r0, #360; ENABLED-NEXT: subs r0, #461; ENABLED-NEXT: add.w r0, r7, r0, lsr #262; ENABLED-NEXT: mov r7, r1063; ENABLED-NEXT: dls lr, r064; ENABLED-NEXT: mov r0, r1165; ENABLED-NEXT: .LBB0_6: @ %vector.body66; ENABLED-NEXT: @ Parent Loop BB0_4 Depth=167; ENABLED-NEXT: @ => This Inner Loop Header: Depth=268; ENABLED-NEXT: vctp.32 r469; ENABLED-NEXT: vmov q0, q170; ENABLED-NEXT: vpstt71; ENABLED-NEXT: vldrht.s32 q1, [r0], #872; ENABLED-NEXT: vldrht.s32 q2, [r7], #873; ENABLED-NEXT: mov lr, r674; ENABLED-NEXT: subs r6, #175; ENABLED-NEXT: vmul.i32 q1, q2, q176; ENABLED-NEXT: subs r4, #477; ENABLED-NEXT: vshl.s32 q1, r578; ENABLED-NEXT: vadd.i32 q1, q1, q079; ENABLED-NEXT: le lr, .LBB0_680; ENABLED-NEXT: @ %bb.7: @ %middle.block81; ENABLED-NEXT: @ in Loop: Header=BB0_4 Depth=182; ENABLED-NEXT: vpsel q0, q1, q083; ENABLED-NEXT: vaddv.u32 r0, q084; ENABLED-NEXT: b .LBB0_385; ENABLED-NEXT: .LBB0_8:86; ENABLED-NEXT: pop.w {r4, r5, r6, r7, r9, r10, r11, lr}87; ENABLED-NEXT: bx lr88;89; NOREDUCTIONS-LABEL: varying_outer_2d_reduction:90; NOREDUCTIONS: @ %bb.0: @ %entry91; NOREDUCTIONS-NEXT: cmp r3, #192; NOREDUCTIONS-NEXT: it lt93; NOREDUCTIONS-NEXT: bxlt lr94; NOREDUCTIONS-NEXT: .LBB0_1: @ %for.body.lr.ph95; NOREDUCTIONS-NEXT: push.w {r4, r5, r6, r7, r9, r10, r11, lr}96; NOREDUCTIONS-NEXT: mov r11, r097; NOREDUCTIONS-NEXT: ldr r0, [sp, #32]98; NOREDUCTIONS-NEXT: add.w r9, r2, #399; NOREDUCTIONS-NEXT: mov.w r12, #0100; NOREDUCTIONS-NEXT: mov r10, r11101; NOREDUCTIONS-NEXT: uxth r0, r0102; NOREDUCTIONS-NEXT: rsbs r5, r0, #0103; NOREDUCTIONS-NEXT: b .LBB0_4104; NOREDUCTIONS-NEXT: .LBB0_2: @ in Loop: Header=BB0_4 Depth=1105; NOREDUCTIONS-NEXT: movs r0, #0106; NOREDUCTIONS-NEXT: .LBB0_3: @ %for.end107; NOREDUCTIONS-NEXT: @ in Loop: Header=BB0_4 Depth=1108; NOREDUCTIONS-NEXT: lsrs r0, r0, #16109; NOREDUCTIONS-NEXT: sub.w r9, r9, #1110; NOREDUCTIONS-NEXT: strh.w r0, [r1, r12, lsl #1]111; NOREDUCTIONS-NEXT: add.w r12, r12, #1112; NOREDUCTIONS-NEXT: add.w r10, r10, #2113; NOREDUCTIONS-NEXT: cmp r12, r3114; NOREDUCTIONS-NEXT: beq .LBB0_8115; NOREDUCTIONS-NEXT: .LBB0_4: @ %for.body116; NOREDUCTIONS-NEXT: @ =>This Loop Header: Depth=1117; NOREDUCTIONS-NEXT: @ Child Loop BB0_6 Depth 2118; NOREDUCTIONS-NEXT: cmp r2, r12119; NOREDUCTIONS-NEXT: ble .LBB0_2120; NOREDUCTIONS-NEXT: @ %bb.5: @ %vector.ph121; NOREDUCTIONS-NEXT: @ in Loop: Header=BB0_4 Depth=1122; NOREDUCTIONS-NEXT: bic r0, r9, #3123; NOREDUCTIONS-NEXT: movs r7, #1124; NOREDUCTIONS-NEXT: subs r0, #4125; NOREDUCTIONS-NEXT: sub.w r4, r2, r12126; NOREDUCTIONS-NEXT: vmov.i32 q1, #0x0127; NOREDUCTIONS-NEXT: add.w r6, r7, r0, lsr #2128; NOREDUCTIONS-NEXT: adds r0, r2, #3129; NOREDUCTIONS-NEXT: sub.w r0, r0, r12130; NOREDUCTIONS-NEXT: bic r0, r0, #3131; NOREDUCTIONS-NEXT: subs r0, #4132; NOREDUCTIONS-NEXT: add.w r0, r7, r0, lsr #2133; NOREDUCTIONS-NEXT: mov r7, r10134; NOREDUCTIONS-NEXT: dls lr, r0135; NOREDUCTIONS-NEXT: mov r0, r11136; NOREDUCTIONS-NEXT: .LBB0_6: @ %vector.body137; NOREDUCTIONS-NEXT: @ Parent Loop BB0_4 Depth=1138; NOREDUCTIONS-NEXT: @ => This Inner Loop Header: Depth=2139; NOREDUCTIONS-NEXT: vctp.32 r4140; NOREDUCTIONS-NEXT: vmov q0, q1141; NOREDUCTIONS-NEXT: vpstt142; NOREDUCTIONS-NEXT: vldrht.s32 q1, [r0], #8143; NOREDUCTIONS-NEXT: vldrht.s32 q2, [r7], #8144; NOREDUCTIONS-NEXT: mov lr, r6145; NOREDUCTIONS-NEXT: subs r6, #1146; NOREDUCTIONS-NEXT: vmul.i32 q1, q2, q1147; NOREDUCTIONS-NEXT: subs r4, #4148; NOREDUCTIONS-NEXT: vshl.s32 q1, r5149; NOREDUCTIONS-NEXT: vadd.i32 q1, q1, q0150; NOREDUCTIONS-NEXT: le lr, .LBB0_6151; NOREDUCTIONS-NEXT: @ %bb.7: @ %middle.block152; NOREDUCTIONS-NEXT: @ in Loop: Header=BB0_4 Depth=1153; NOREDUCTIONS-NEXT: vpsel q0, q1, q0154; NOREDUCTIONS-NEXT: vaddv.u32 r0, q0155; NOREDUCTIONS-NEXT: b .LBB0_3156; NOREDUCTIONS-NEXT: .LBB0_8:157; NOREDUCTIONS-NEXT: pop.w {r4, r5, r6, r7, r9, r10, r11, lr}158; NOREDUCTIONS-NEXT: bx lr159entry:160 %conv = sext i16 %N to i32161 %cmp36 = icmp sgt i16 %N, 0162 br i1 %cmp36, label %for.body.lr.ph, label %for.end17163 164for.body.lr.ph: ; preds = %entry165 %conv2 = sext i16 %Size to i32166 %conv1032 = zext i16 %Scale to i32167 %i = add i32 %conv2, 3168 br label %for.body169 170for.body: ; preds = %for.end, %for.body.lr.ph171 %lsr.iv51 = phi i32 [ %lsr.iv.next, %for.end ], [ %i, %for.body.lr.ph ]172 %lsr.iv46 = phi ptr [ %scevgep47, %for.end ], [ %Input, %for.body.lr.ph ]173 %i.037 = phi i32 [ 0, %for.body.lr.ph ], [ %inc16, %for.end ]174 %i1 = mul nsw i32 %i.037, -1175 %i2 = add i32 %i, %i1176 %i3 = lshr i32 %i2, 2177 %i4 = shl nuw i32 %i3, 2178 %i5 = add i32 %i4, -4179 %i6 = lshr i32 %i5, 2180 %i7 = add nuw nsw i32 %i6, 1181 %i8 = sub i32 %conv2, %i.037182 %cmp433 = icmp slt i32 %i.037, %conv2183 br i1 %cmp433, label %vector.ph, label %for.end184 185vector.ph: ; preds = %for.body186 %start = call i32 @llvm.start.loop.iterations.i32(i32 %i7)187 br label %vector.body188 189vector.body: ; preds = %vector.body, %vector.ph190 %lsr.iv48 = phi ptr [ %scevgep49, %vector.body ], [ %lsr.iv46, %vector.ph ]191 %lsr.iv = phi ptr [ %scevgep, %vector.body ], [ %Input, %vector.ph ]192 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]193 %vec.phi = phi <4 x i32> [ zeroinitializer, %vector.ph ], [ %i16, %vector.body ]194 %i9 = phi i32 [ %start, %vector.ph ], [ %i17, %vector.body ]195 %active.lane.mask = call <4 x i1> @llvm.get.active.lane.mask.v4i1.i32(i32 %index, i32 %i8)196 %wide.masked.load = call <4 x i16> @llvm.masked.load.v4i16.p0(ptr %lsr.iv, i32 2, <4 x i1> %active.lane.mask, <4 x i16> undef)197 %i10 = sext <4 x i16> %wide.masked.load to <4 x i32>198 %wide.masked.load42 = call <4 x i16> @llvm.masked.load.v4i16.p0(ptr %lsr.iv48, i32 2, <4 x i1> %active.lane.mask, <4 x i16> undef)199 %i11 = sext <4 x i16> %wide.masked.load42 to <4 x i32>200 %i12 = mul nsw <4 x i32> %i11, %i10201 %i13 = insertelement <4 x i32> undef, i32 %conv1032, i32 0202 %i14 = shufflevector <4 x i32> %i13, <4 x i32> undef, <4 x i32> zeroinitializer203 %i15 = ashr <4 x i32> %i12, %i14204 %i16 = add <4 x i32> %i15, %vec.phi205 %index.next = add i32 %index, 4206 %scevgep = getelementptr i16, ptr %lsr.iv, i32 4207 %scevgep49 = getelementptr i16, ptr %lsr.iv48, i32 4208 %i17 = call i32 @llvm.loop.decrement.reg.i32(i32 %i9, i32 1)209 %i18 = icmp ne i32 %i17, 0210 br i1 %i18, label %vector.body, label %middle.block211 212middle.block: ; preds = %vector.body213 %i19 = select <4 x i1> %active.lane.mask, <4 x i32> %i16, <4 x i32> %vec.phi214 %i20 = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> %i19)215 br label %for.end216 217for.end: ; preds = %middle.block, %for.body218 %Sum.0.lcssa = phi i32 [ 0, %for.body ], [ %i20, %middle.block ]219 %i21 = lshr i32 %Sum.0.lcssa, 16220 %conv13 = trunc i32 %i21 to i16221 %arrayidx14 = getelementptr inbounds i16, ptr %Output, i32 %i.037222 store i16 %conv13, ptr %arrayidx14, align 2223 %inc16 = add nuw nsw i32 %i.037, 1224 %scevgep47 = getelementptr i16, ptr %lsr.iv46, i32 1225 %lsr.iv.next = add i32 %lsr.iv51, -1226 %exitcond39 = icmp eq i32 %inc16, %conv227 br i1 %exitcond39, label %for.end17, label %for.body228 229for.end17: ; preds = %for.end, %entry230 ret void231}232 233declare <4 x i1> @llvm.get.active.lane.mask.v4i1.i32(i32, i32)234declare <4 x i16> @llvm.masked.load.v4i16.p0(ptr, i32 immarg, <4 x i1>, <4 x i16>)235declare i32 @llvm.vector.reduce.add.v4i32(<4 x i32>)236declare i32 @llvm.loop.decrement.reg.i32(i32, i32)237declare i32 @llvm.start.loop.iterations.i32(i32)238