259 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 22; RUN: llc -mtriple=thumbv8.1m.main-none-none-eabi -mattr=+mve.fp %s -o - | FileCheck %s3 4define arm_aapcs_vfpcc void @start12(ptr nocapture readonly %x, ptr nocapture readonly %y, ptr noalias nocapture %z, float %a, i32 %n) {5; CHECK-LABEL: start12:6; CHECK: @ %bb.0: @ %entry7; CHECK-NEXT: .save {r4, lr}8; CHECK-NEXT: push {r4, lr}9; CHECK-NEXT: cmp r3, #110; CHECK-NEXT: it lt11; CHECK-NEXT: poplt {r4, pc}12; CHECK-NEXT: .LBB0_1: @ %vector.ph13; CHECK-NEXT: vmov r12, s014; CHECK-NEXT: adds r0, #4815; CHECK-NEXT: adds r1, #4816; CHECK-NEXT: adds r2, #4817; CHECK-NEXT: subs r3, #1218; CHECK-NEXT: dlstp.32 lr, r319; CHECK-NEXT: .LBB0_2: @ %vector.body20; CHECK-NEXT: @ =>This Inner Loop Header: Depth=121; CHECK-NEXT: vldrw.u32 q0, [r1], #1622; CHECK-NEXT: vldrw.u32 q1, [r0], #1623; CHECK-NEXT: vfmas.f32 q1, q0, r1224; CHECK-NEXT: vstrw.32 q1, [r2], #1625; CHECK-NEXT: letp lr, .LBB0_226; CHECK-NEXT: @ %bb.3: @ %for.cond.cleanup27; CHECK-NEXT: pop {r4, pc}28entry:29 %cmp8 = icmp sgt i32 %n, 030 br i1 %cmp8, label %vector.ph, label %for.cond.cleanup31 32vector.ph: ; preds = %entry33 %n.rnd.up = add i32 %n, 334 %n.vec = and i32 %n.rnd.up, -435 %broadcast.splatinsert13 = insertelement <4 x float> undef, float %a, i32 036 %broadcast.splat14 = shufflevector <4 x float> %broadcast.splatinsert13, <4 x float> undef, <4 x i32> zeroinitializer37 br label %vector.body38 39vector.body: ; preds = %vector.body, %vector.ph40 %index = phi i32 [ 12, %vector.ph ], [ %index.next, %vector.body ]41 %0 = getelementptr inbounds float, ptr %x, i32 %index42 %1 = call <4 x i1> @llvm.get.active.lane.mask.v4i1.i32(i32 %index, i32 %n)43 %wide.masked.load = call <4 x float> @llvm.masked.load.v4f32.p0(ptr %0, i32 4, <4 x i1> %1, <4 x float> undef)44 %2 = getelementptr inbounds float, ptr %y, i32 %index45 %wide.masked.load12 = call <4 x float> @llvm.masked.load.v4f32.p0(ptr %2, i32 4, <4 x i1> %1, <4 x float> undef)46 %3 = call fast <4 x float> @llvm.fma.v4f32(<4 x float> %wide.masked.load, <4 x float> %wide.masked.load12, <4 x float> %broadcast.splat14)47 %4 = getelementptr inbounds float, ptr %z, i32 %index48 call void @llvm.masked.store.v4f32.p0(<4 x float> %3, ptr %4, i32 4, <4 x i1> %1)49 %index.next = add i32 %index, 450 %5 = icmp eq i32 %index.next, %n.vec51 br i1 %5, label %for.cond.cleanup, label %vector.body52 53for.cond.cleanup: ; preds = %vector.body, %entry54 ret void55}56 57 58define arm_aapcs_vfpcc void @start11(ptr nocapture readonly %x, ptr nocapture readonly %y, ptr noalias nocapture %z, float %a, i32 %n) {59; CHECK-LABEL: start11:60; CHECK: @ %bb.0: @ %entry61; CHECK-NEXT: cmp r3, #162; CHECK-NEXT: it lt63; CHECK-NEXT: bxlt lr64; CHECK-NEXT: .LBB1_1: @ %vector.ph65; CHECK-NEXT: .save {r4, r5, r7, lr}66; CHECK-NEXT: push {r4, r5, r7, lr}67; CHECK-NEXT: vmov r12, s068; CHECK-NEXT: adds r4, r3, #369; CHECK-NEXT: adr r5, .LCPI1_070; CHECK-NEXT: bic lr, r4, #371; CHECK-NEXT: adds r0, #4472; CHECK-NEXT: adds r1, #4473; CHECK-NEXT: adds r2, #4474; CHECK-NEXT: vldrw.u32 q0, [r5]75; CHECK-NEXT: movs r4, #1176; CHECK-NEXT: vdup.32 q1, r377; CHECK-NEXT: .LBB1_2: @ %vector.body78; CHECK-NEXT: @ =>This Inner Loop Header: Depth=179; CHECK-NEXT: vqadd.u32 q2, q0, r480; CHECK-NEXT: adds r4, #481; CHECK-NEXT: cmp lr, r482; CHECK-NEXT: vptt.u32 hi, q1, q283; CHECK-NEXT: vldrwt.u32 q2, [r1], #1684; CHECK-NEXT: vldrwt.u32 q3, [r0], #1685; CHECK-NEXT: vfmas.f32 q3, q2, r1286; CHECK-NEXT: vpst87; CHECK-NEXT: vstrwt.32 q3, [r2], #1688; CHECK-NEXT: bne .LBB1_289; CHECK-NEXT: @ %bb.3:90; CHECK-NEXT: pop.w {r4, r5, r7, lr}91; CHECK-NEXT: bx lr92; CHECK-NEXT: .p2align 493; CHECK-NEXT: @ %bb.4:94; CHECK-NEXT: .LCPI1_0:95; CHECK-NEXT: .long 0 @ 0x096; CHECK-NEXT: .long 1 @ 0x197; CHECK-NEXT: .long 2 @ 0x298; CHECK-NEXT: .long 3 @ 0x399entry:100 %cmp8 = icmp sgt i32 %n, 0101 br i1 %cmp8, label %vector.ph, label %for.cond.cleanup102 103vector.ph: ; preds = %entry104 %n.rnd.up = add i32 %n, 3105 %n.vec = and i32 %n.rnd.up, -4106 %broadcast.splatinsert13 = insertelement <4 x float> undef, float %a, i32 0107 %broadcast.splat14 = shufflevector <4 x float> %broadcast.splatinsert13, <4 x float> undef, <4 x i32> zeroinitializer108 br label %vector.body109 110vector.body: ; preds = %vector.body, %vector.ph111 %index = phi i32 [ 11, %vector.ph ], [ %index.next, %vector.body ]112 %0 = getelementptr inbounds float, ptr %x, i32 %index113 %1 = call <4 x i1> @llvm.get.active.lane.mask.v4i1.i32(i32 %index, i32 %n)114 %wide.masked.load = call <4 x float> @llvm.masked.load.v4f32.p0(ptr %0, i32 4, <4 x i1> %1, <4 x float> undef)115 %2 = getelementptr inbounds float, ptr %y, i32 %index116 %wide.masked.load12 = call <4 x float> @llvm.masked.load.v4f32.p0(ptr %2, i32 4, <4 x i1> %1, <4 x float> undef)117 %3 = call fast <4 x float> @llvm.fma.v4f32(<4 x float> %wide.masked.load, <4 x float> %wide.masked.load12, <4 x float> %broadcast.splat14)118 %4 = getelementptr inbounds float, ptr %z, i32 %index119 call void @llvm.masked.store.v4f32.p0(<4 x float> %3, ptr %4, i32 4, <4 x i1> %1)120 %index.next = add i32 %index, 4121 %5 = icmp eq i32 %index.next, %n.vec122 br i1 %5, label %for.cond.cleanup, label %vector.body123 124for.cond.cleanup: ; preds = %vector.body, %entry125 ret void126}127 128define arm_aapcs_vfpcc void @startS(i32 %S, ptr nocapture readonly %x, ptr nocapture readonly %y, ptr noalias nocapture %z, float %a, i32 %n) {129; CHECK-LABEL: startS:130; CHECK: @ %bb.0: @ %entry131; CHECK-NEXT: .save {r4, r5, r7, lr}132; CHECK-NEXT: push {r4, r5, r7, lr}133; CHECK-NEXT: ldr r5, [sp, #16]134; CHECK-NEXT: cmp r5, #1135; CHECK-NEXT: blt .LBB2_3136; CHECK-NEXT: @ %bb.1: @ %vector.ph137; CHECK-NEXT: vmov r12, s0138; CHECK-NEXT: adds r4, r5, #3139; CHECK-NEXT: bic lr, r4, #3140; CHECK-NEXT: adr r4, .LCPI2_0141; CHECK-NEXT: add.w r1, r1, r0, lsl #2142; CHECK-NEXT: add.w r2, r2, r0, lsl #2143; CHECK-NEXT: add.w r3, r3, r0, lsl #2144; CHECK-NEXT: vldrw.u32 q0, [r4]145; CHECK-NEXT: vdup.32 q1, r5146; CHECK-NEXT: .LBB2_2: @ %vector.body147; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1148; CHECK-NEXT: vqadd.u32 q2, q0, r0149; CHECK-NEXT: adds r0, #4150; CHECK-NEXT: cmp lr, r0151; CHECK-NEXT: vptt.u32 hi, q1, q2152; CHECK-NEXT: vldrwt.u32 q2, [r2], #16153; CHECK-NEXT: vldrwt.u32 q3, [r1], #16154; CHECK-NEXT: vfmas.f32 q3, q2, r12155; CHECK-NEXT: vpst156; CHECK-NEXT: vstrwt.32 q3, [r3], #16157; CHECK-NEXT: bne .LBB2_2158; CHECK-NEXT: .LBB2_3: @ %for.cond.cleanup159; CHECK-NEXT: pop {r4, r5, r7, pc}160; CHECK-NEXT: .p2align 4161; CHECK-NEXT: @ %bb.4:162; CHECK-NEXT: .LCPI2_0:163; CHECK-NEXT: .long 0 @ 0x0164; CHECK-NEXT: .long 1 @ 0x1165; CHECK-NEXT: .long 2 @ 0x2166; CHECK-NEXT: .long 3 @ 0x3167entry:168 %cmp8 = icmp sgt i32 %n, 0169 br i1 %cmp8, label %vector.ph, label %for.cond.cleanup170 171vector.ph: ; preds = %entry172 %n.rnd.up = add i32 %n, 3173 %n.vec = and i32 %n.rnd.up, -4174 %broadcast.splatinsert13 = insertelement <4 x float> undef, float %a, i32 0175 %broadcast.splat14 = shufflevector <4 x float> %broadcast.splatinsert13, <4 x float> undef, <4 x i32> zeroinitializer176 br label %vector.body177 178vector.body: ; preds = %vector.body, %vector.ph179 %index = phi i32 [ %S, %vector.ph ], [ %index.next, %vector.body ]180 %0 = getelementptr inbounds float, ptr %x, i32 %index181 %1 = call <4 x i1> @llvm.get.active.lane.mask.v4i1.i32(i32 %index, i32 %n)182 %wide.masked.load = call <4 x float> @llvm.masked.load.v4f32.p0(ptr %0, i32 4, <4 x i1> %1, <4 x float> undef)183 %2 = getelementptr inbounds float, ptr %y, i32 %index184 %wide.masked.load12 = call <4 x float> @llvm.masked.load.v4f32.p0(ptr %2, i32 4, <4 x i1> %1, <4 x float> undef)185 %3 = call fast <4 x float> @llvm.fma.v4f32(<4 x float> %wide.masked.load, <4 x float> %wide.masked.load12, <4 x float> %broadcast.splat14)186 %4 = getelementptr inbounds float, ptr %z, i32 %index187 call void @llvm.masked.store.v4f32.p0(<4 x float> %3, ptr %4, i32 4, <4 x i1> %1)188 %index.next = add i32 %index, 4189 %5 = icmp eq i32 %index.next, %n.vec190 br i1 %5, label %for.cond.cleanup, label %vector.body191 192for.cond.cleanup: ; preds = %vector.body, %entry193 ret void194}195 196define arm_aapcs_vfpcc void @startSmod4(i32 %S, ptr nocapture readonly %x, ptr nocapture readonly %y, ptr noalias nocapture %z, float %a, i32 %n) {197; CHECK-LABEL: startSmod4:198; CHECK: @ %bb.0: @ %entry199; CHECK-NEXT: .save {r4, lr}200; CHECK-NEXT: push {r4, lr}201; CHECK-NEXT: ldr.w lr, [sp, #8]202; CHECK-NEXT: cmp.w lr, #1203; CHECK-NEXT: it lt204; CHECK-NEXT: poplt {r4, pc}205; CHECK-NEXT: .LBB3_1: @ %vector.ph206; CHECK-NEXT: vmov r12, s0207; CHECK-NEXT: mvn r4, #12208; CHECK-NEXT: and.w r4, r4, r0, lsl #2209; CHECK-NEXT: add r1, r4210; CHECK-NEXT: add r2, r4211; CHECK-NEXT: add r3, r4212; CHECK-NEXT: sub.w r0, lr, #4213; CHECK-NEXT: dlstp.32 lr, r0214; CHECK-NEXT: .LBB3_2: @ %vector.body215; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1216; CHECK-NEXT: vldrw.u32 q0, [r2], #16217; CHECK-NEXT: vldrw.u32 q1, [r1], #16218; CHECK-NEXT: vfmas.f32 q1, q0, r12219; CHECK-NEXT: vstrw.32 q1, [r3], #16220; CHECK-NEXT: letp lr, .LBB3_2221; CHECK-NEXT: @ %bb.3: @ %for.cond.cleanup222; CHECK-NEXT: pop {r4, pc}223entry:224 %cmp8 = icmp sgt i32 %n, 0225 br i1 %cmp8, label %vector.ph, label %for.cond.cleanup226 227vector.ph: ; preds = %entry228 %Sm = and i32 %S, -4229 %n.rnd.up = add i32 %n, 3230 %n.vec = and i32 %n.rnd.up, -4231 %broadcast.splatinsert13 = insertelement <4 x float> undef, float %a, i32 0232 %broadcast.splat14 = shufflevector <4 x float> %broadcast.splatinsert13, <4 x float> undef, <4 x i32> zeroinitializer233 br label %vector.body234 235vector.body: ; preds = %vector.body, %vector.ph236 %index = phi i32 [ %Sm, %vector.ph ], [ %index.next, %vector.body ]237 %0 = getelementptr inbounds float, ptr %x, i32 %index238 %1 = call <4 x i1> @llvm.get.active.lane.mask.v4i1.i32(i32 %index, i32 %n)239 %wide.masked.load = call <4 x float> @llvm.masked.load.v4f32.p0(ptr %0, i32 4, <4 x i1> %1, <4 x float> undef)240 %2 = getelementptr inbounds float, ptr %y, i32 %index241 %wide.masked.load12 = call <4 x float> @llvm.masked.load.v4f32.p0(ptr %2, i32 4, <4 x i1> %1, <4 x float> undef)242 %3 = call fast <4 x float> @llvm.fma.v4f32(<4 x float> %wide.masked.load, <4 x float> %wide.masked.load12, <4 x float> %broadcast.splat14)243 %4 = getelementptr inbounds float, ptr %z, i32 %index244 call void @llvm.masked.store.v4f32.p0(<4 x float> %3, ptr %4, i32 4, <4 x i1> %1)245 %index.next = add i32 %index, 4246 %5 = icmp eq i32 %index.next, %n.vec247 br i1 %5, label %for.cond.cleanup, label %vector.body248 249for.cond.cleanup: ; preds = %vector.body, %entry250 ret void251}252 253 254declare <4 x float> @llvm.masked.load.v4f32.p0(ptr, i32 immarg, <4 x i1>, <4 x float>)255declare <4 x float> @llvm.fma.v4f32(<4 x float>, <4 x float>, <4 x float>)256declare void @llvm.masked.store.v4f32.p0(<4 x float>, ptr, i32 immarg, <4 x i1>)257declare <4 x i1> @llvm.get.active.lane.mask.v4i1.i32(i32, i32)258 259