1; RUN: llc < %s -mcpu=cortex-a9 -verify-coalescing -verify-machineinstrs | FileCheck %s 2target datalayout = "e-p:32:32:32-i1:8:32-i8:8:32-i16:16:32-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:32:64-v128:32:128-a0:0:32-n32-S32" 3target triple = "thumbv7-apple-ios0.0.0" 4 5; CHECK: f 6; The vld2 and vst2 are not aligned wrt each other, the second Q loaded is the 7; first one stored. 8; The coalescer must find a super-register larger than QQ to eliminate the copy 9; setting up the vst2 data. 10; CHECK: vld2 11; CHECK-NOT: vorr 12; CHECK-NOT: vmov 13; CHECK: vst2 14define void @f(float* %p, i32 %c) nounwind ssp { 15entry: 16 %0 = bitcast float* %p to i8* 17 %vld2 = tail call { <4 x float>, <4 x float> } @llvm.arm.neon.vld2.v4f32.p0i8(i8* %0, i32 4) 18 %vld221 = extractvalue { <4 x float>, <4 x float> } %vld2, 1 19 %add.ptr = getelementptr inbounds float, float* %p, i32 8 20 %1 = bitcast float* %add.ptr to i8* 21 tail call void @llvm.arm.neon.vst2.p0i8.v4f32(i8* %1, <4 x float> %vld221, <4 x float> undef, i32 4) 22 ret void 23} 24 25; CHECK: f1 26; FIXME: This function still has copies. 27define void @f1(float* %p, i32 %c) nounwind ssp { 28entry: 29 %0 = bitcast float* %p to i8* 30 %vld2 = tail call { <4 x float>, <4 x float> } @llvm.arm.neon.vld2.v4f32.p0i8(i8* %0, i32 4) 31 %vld221 = extractvalue { <4 x float>, <4 x float> } %vld2, 1 32 %add.ptr = getelementptr inbounds float, float* %p, i32 8 33 %1 = bitcast float* %add.ptr to i8* 34 %vld22 = tail call { <4 x float>, <4 x float> } @llvm.arm.neon.vld2.v4f32.p0i8(i8* %1, i32 4) 35 %vld2215 = extractvalue { <4 x float>, <4 x float> } %vld22, 0 36 tail call void @llvm.arm.neon.vst2.p0i8.v4f32(i8* %1, <4 x float> %vld221, <4 x float> %vld2215, i32 4) 37 ret void 38} 39 40; CHECK: f2 41; FIXME: This function still has copies. 42define void @f2(float* %p, i32 %c) nounwind ssp { 43entry: 44 %0 = bitcast float* %p to i8* 45 %vld2 = tail call { <4 x float>, <4 x float> } @llvm.arm.neon.vld2.v4f32.p0i8(i8* %0, i32 4) 46 %vld224 = extractvalue { <4 x float>, <4 x float> } %vld2, 1 47 br label %do.body 48 49do.body: ; preds = %do.body, %entry 50 %qq0.0.1.0 = phi <4 x float> [ %vld224, %entry ], [ %vld2216, %do.body ] 51 %c.addr.0 = phi i32 [ %c, %entry ], [ %dec, %do.body ] 52 %p.addr.0 = phi float* [ %p, %entry ], [ %add.ptr, %do.body ] 53 %add.ptr = getelementptr inbounds float, float* %p.addr.0, i32 8 54 %1 = bitcast float* %add.ptr to i8* 55 %vld22 = tail call { <4 x float>, <4 x float> } @llvm.arm.neon.vld2.v4f32.p0i8(i8* %1, i32 4) 56 %vld2215 = extractvalue { <4 x float>, <4 x float> } %vld22, 0 57 %vld2216 = extractvalue { <4 x float>, <4 x float> } %vld22, 1 58 tail call void @llvm.arm.neon.vst2.p0i8.v4f32(i8* %1, <4 x float> %qq0.0.1.0, <4 x float> %vld2215, i32 4) 59 %dec = add nsw i32 %c.addr.0, -1 60 %tobool = icmp eq i32 %dec, 0 61 br i1 %tobool, label %do.end, label %do.body 62 63do.end: ; preds = %do.body 64 ret void 65} 66 67declare { <4 x float>, <4 x float> } @llvm.arm.neon.vld2.v4f32.p0i8(i8*, i32) nounwind readonly 68declare void @llvm.arm.neon.vst2.p0i8.v4f32(i8*, <4 x float>, <4 x float>, i32) nounwind 69 70; CHECK: f3 71; This function has lane insertions that span basic blocks. 72; The trivial REG_SEQUENCE lowering can't handle that, but the coalescer can. 73; 74; void f3(float *p, float *q) { 75; float32x2_t x; 76; x[1] = p[3]; 77; if (q) 78; x[0] = q[0] + q[1]; 79; else 80; x[0] = p[2]; 81; vst1_f32(p+4, x); 82; } 83; 84; CHECK-NOT: vmov 85; CHECK-NOT: vorr 86define void @f3(float* %p, float* %q) nounwind ssp { 87entry: 88 %arrayidx = getelementptr inbounds float, float* %p, i32 3 89 %0 = load float, float* %arrayidx, align 4 90 %vecins = insertelement <2 x float> undef, float %0, i32 1 91 %tobool = icmp eq float* %q, null 92 br i1 %tobool, label %if.else, label %if.then 93 94if.then: ; preds = %entry 95 %1 = load float, float* %q, align 4 96 %arrayidx2 = getelementptr inbounds float, float* %q, i32 1 97 %2 = load float, float* %arrayidx2, align 4 98 %add = fadd float %1, %2 99 %vecins3 = insertelement <2 x float> %vecins, float %add, i32 0 100 br label %if.end 101 102if.else: ; preds = %entry 103 %arrayidx4 = getelementptr inbounds float, float* %p, i32 2 104 %3 = load float, float* %arrayidx4, align 4 105 %vecins5 = insertelement <2 x float> %vecins, float %3, i32 0 106 br label %if.end 107 108if.end: ; preds = %if.else, %if.then 109 %x.0 = phi <2 x float> [ %vecins3, %if.then ], [ %vecins5, %if.else ] 110 %add.ptr = getelementptr inbounds float, float* %p, i32 4 111 %4 = bitcast float* %add.ptr to i8* 112 tail call void @llvm.arm.neon.vst1.p0i8.v2f32(i8* %4, <2 x float> %x.0, i32 4) 113 ret void 114} 115 116declare void @llvm.arm.neon.vst1.p0i8.v2f32(i8*, <2 x float>, i32) nounwind 117declare <2 x float> @llvm.arm.neon.vld1.v2f32.p0i8(i8*, i32) nounwind readonly 118 119; CHECK: f4 120; This function inserts a lane into a fully defined vector. 121; The destination lane isn't read, so the subregs can coalesce. 122; CHECK-NOT: vmov 123; CHECK-NOT: vorr 124define void @f4(float* %p, float* %q) nounwind ssp { 125entry: 126 %0 = bitcast float* %p to i8* 127 %vld1 = tail call <2 x float> @llvm.arm.neon.vld1.v2f32.p0i8(i8* %0, i32 4) 128 %tobool = icmp eq float* %q, null 129 br i1 %tobool, label %if.end, label %if.then 130 131if.then: ; preds = %entry 132 %1 = load float, float* %q, align 4 133 %arrayidx1 = getelementptr inbounds float, float* %q, i32 1 134 %2 = load float, float* %arrayidx1, align 4 135 %add = fadd float %1, %2 136 %vecins = insertelement <2 x float> %vld1, float %add, i32 1 137 br label %if.end 138 139if.end: ; preds = %entry, %if.then 140 %x.0 = phi <2 x float> [ %vecins, %if.then ], [ %vld1, %entry ] 141 tail call void @llvm.arm.neon.vst1.p0i8.v2f32(i8* %0, <2 x float> %x.0, i32 4) 142 ret void 143} 144 145; CHECK: f5 146; Coalesce vector lanes through phis. 147; CHECK: vmov.f32 {{.*}}, #1.0 148; CHECK-NOT: vmov 149; CHECK-NOT: vorr 150; CHECK: bx 151; We may leave the last insertelement in the if.end block. 152; It is inserting the %add value into a dead lane, but %add causes interference 153; in the entry block, and we don't do dead lane checks across basic blocks. 154define void @f5(float* %p, float* %q) nounwind ssp { 155entry: 156 %0 = bitcast float* %p to i8* 157 %vld1 = tail call <4 x float> @llvm.arm.neon.vld1.v4f32.p0i8(i8* %0, i32 4) 158 %vecext = extractelement <4 x float> %vld1, i32 0 159 %vecext1 = extractelement <4 x float> %vld1, i32 1 160 %vecext2 = extractelement <4 x float> %vld1, i32 2 161 %vecext3 = extractelement <4 x float> %vld1, i32 3 162 %add = fadd float %vecext3, 1.000000e+00 163 %tobool = icmp eq float* %q, null 164 br i1 %tobool, label %if.end, label %if.then 165 166if.then: ; preds = %entry 167 %arrayidx = getelementptr inbounds float, float* %q, i32 1 168 %1 = load float, float* %arrayidx, align 4 169 %add4 = fadd float %vecext, %1 170 %2 = load float, float* %q, align 4 171 %add6 = fadd float %vecext1, %2 172 %arrayidx7 = getelementptr inbounds float, float* %q, i32 2 173 %3 = load float, float* %arrayidx7, align 4 174 %add8 = fadd float %vecext2, %3 175 br label %if.end 176 177if.end: ; preds = %entry, %if.then 178 %a.0 = phi float [ %add4, %if.then ], [ %vecext, %entry ] 179 %b.0 = phi float [ %add6, %if.then ], [ %vecext1, %entry ] 180 %c.0 = phi float [ %add8, %if.then ], [ %vecext2, %entry ] 181 %vecinit = insertelement <4 x float> undef, float %a.0, i32 0 182 %vecinit9 = insertelement <4 x float> %vecinit, float %b.0, i32 1 183 %vecinit10 = insertelement <4 x float> %vecinit9, float %c.0, i32 2 184 %vecinit11 = insertelement <4 x float> %vecinit10, float %add, i32 3 185 tail call void @llvm.arm.neon.vst1.p0i8.v4f32(i8* %0, <4 x float> %vecinit11, i32 4) 186 ret void 187} 188 189declare <4 x float> @llvm.arm.neon.vld1.v4f32.p0i8(i8*, i32) nounwind readonly 190 191declare void @llvm.arm.neon.vst1.p0i8.v4f32(i8*, <4 x float>, i32) nounwind 192 193; CHECK: pr13999 194define void @pr13999() nounwind readonly { 195entry: 196 br i1 true, label %outer_loop, label %loop.end 197 198outer_loop: 199 %d = phi double [ 0.0, %entry ], [ %add, %after_inner_loop ] 200 %0 = insertelement <2 x double> <double 0.0, double 0.0>, double %d, i32 0 201 br i1 undef, label %after_inner_loop, label %inner_loop 202 203inner_loop: 204 br i1 true, label %after_inner_loop, label %inner_loop 205 206after_inner_loop: 207 %1 = phi <2 x double> [ %0, %outer_loop ], [ <double 0.0, double 0.0>, 208%inner_loop ] 209 %2 = extractelement <2 x double> %1, i32 1 210 %add = fadd double 1.0, %2 211 br i1 false, label %loop.end, label %outer_loop 212 213loop.end: 214 %d.end = phi double [ 0.0, %entry ], [ %add, %after_inner_loop ] 215 ret void 216} 217 218; CHECK: pr14078 219define arm_aapcs_vfpcc i32 @pr14078(i8* nocapture %arg, i8* nocapture %arg1, i32 %arg2) nounwind uwtable readonly { 220bb: 221 br i1 undef, label %bb31, label %bb3 222 223bb3: ; preds = %bb12, %bb 224 %tmp = shufflevector <2 x i64> undef, <2 x i64> undef, <1 x i32> zeroinitializer 225 %tmp4 = bitcast <1 x i64> %tmp to <2 x float> 226 %tmp5 = shufflevector <2 x float> %tmp4, <2 x float> undef, <4 x i32> zeroinitializer 227 %tmp6 = bitcast <4 x float> %tmp5 to <2 x i64> 228 %tmp7 = shufflevector <2 x i64> %tmp6, <2 x i64> undef, <1 x i32> zeroinitializer 229 %tmp8 = bitcast <1 x i64> %tmp7 to <2 x float> 230 %tmp9 = tail call <2 x float> @baz(<2 x float> <float 0xFFFFFFFFE0000000, float 0.000000e+00>, <2 x float> %tmp8, <2 x float> zeroinitializer) nounwind 231 br i1 undef, label %bb10, label %bb12 232 233bb10: ; preds = %bb3 234 %tmp11 = load <4 x float>, <4 x float>* undef, align 8 235 br label %bb12 236 237bb12: ; preds = %bb10, %bb3 238 %tmp13 = shufflevector <2 x float> %tmp9, <2 x float> zeroinitializer, <2 x i32> <i32 0, i32 2> 239 %tmp14 = bitcast <2 x float> %tmp13 to <1 x i64> 240 %tmp15 = shufflevector <1 x i64> %tmp14, <1 x i64> zeroinitializer, <2 x i32> <i32 0, i32 1> 241 %tmp16 = bitcast <2 x i64> %tmp15 to <4 x float> 242 %tmp17 = fmul <4 x float> zeroinitializer, %tmp16 243 %tmp18 = bitcast <4 x float> %tmp17 to <2 x i64> 244 %tmp19 = shufflevector <2 x i64> %tmp18, <2 x i64> undef, <1 x i32> zeroinitializer 245 %tmp20 = bitcast <1 x i64> %tmp19 to <2 x float> 246 %tmp21 = tail call <2 x float> @baz67(<2 x float> %tmp20, <2 x float> undef) nounwind 247 %tmp22 = tail call <2 x float> @baz67(<2 x float> %tmp21, <2 x float> %tmp21) nounwind 248 %tmp23 = shufflevector <2 x float> %tmp22, <2 x float> undef, <4 x i32> zeroinitializer 249 %tmp24 = bitcast <4 x float> %tmp23 to <2 x i64> 250 %tmp25 = shufflevector <2 x i64> %tmp24, <2 x i64> undef, <1 x i32> zeroinitializer 251 %tmp26 = bitcast <1 x i64> %tmp25 to <2 x float> 252 %tmp27 = extractelement <2 x float> %tmp26, i32 0 253 %tmp28 = fcmp olt float %tmp27, 0.000000e+00 254 %tmp29 = select i1 %tmp28, i32 0, i32 undef 255 %tmp30 = icmp ult i32 undef, %arg2 256 br i1 %tmp30, label %bb3, label %bb31 257 258bb31: ; preds = %bb12, %bb 259 %tmp32 = phi i32 [ 1, %bb ], [ %tmp29, %bb12 ] 260 ret i32 %tmp32 261} 262 263declare <2 x float> @baz(<2 x float>, <2 x float>, <2 x float>) nounwind readnone 264 265declare <2 x float> @baz67(<2 x float>, <2 x float>) nounwind readnone 266 267%struct.wombat.5 = type { %struct.quux, %struct.quux, %struct.quux, %struct.quux } 268%struct.quux = type { <4 x float> } 269 270; CHECK: pr14079 271define linkonce_odr arm_aapcs_vfpcc %struct.wombat.5 @pr14079(i8* nocapture %arg, i8* nocapture %arg1, i8* nocapture %arg2) nounwind uwtable inlinehint { 272bb: 273 %tmp = shufflevector <2 x i64> zeroinitializer, <2 x i64> undef, <1 x i32> zeroinitializer 274 %tmp3 = bitcast <1 x i64> %tmp to <2 x float> 275 %tmp4 = shufflevector <2 x float> %tmp3, <2 x float> zeroinitializer, <2 x i32> <i32 1, i32 3> 276 %tmp5 = shufflevector <2 x float> %tmp4, <2 x float> undef, <2 x i32> <i32 1, i32 3> 277 %tmp6 = bitcast <2 x float> %tmp5 to <1 x i64> 278 %tmp7 = shufflevector <1 x i64> undef, <1 x i64> %tmp6, <2 x i32> <i32 0, i32 1> 279 %tmp8 = bitcast <2 x i64> %tmp7 to <4 x float> 280 %tmp9 = shufflevector <2 x i64> zeroinitializer, <2 x i64> undef, <1 x i32> <i32 1> 281 %tmp10 = bitcast <1 x i64> %tmp9 to <2 x float> 282 %tmp11 = shufflevector <2 x float> %tmp10, <2 x float> undef, <2 x i32> <i32 0, i32 2> 283 %tmp12 = shufflevector <2 x float> %tmp11, <2 x float> undef, <2 x i32> <i32 0, i32 2> 284 %tmp13 = bitcast <2 x float> %tmp12 to <1 x i64> 285 %tmp14 = shufflevector <1 x i64> %tmp13, <1 x i64> undef, <2 x i32> <i32 0, i32 1> 286 %tmp15 = bitcast <2 x i64> %tmp14 to <4 x float> 287 %tmp16 = insertvalue %struct.wombat.5 undef, <4 x float> %tmp8, 1, 0 288 %tmp17 = insertvalue %struct.wombat.5 %tmp16, <4 x float> %tmp15, 2, 0 289 %tmp18 = insertvalue %struct.wombat.5 %tmp17, <4 x float> undef, 3, 0 290 ret %struct.wombat.5 %tmp18 291} 292 293; CHECK: adjustCopiesBackFrom 294; The shuffle in if.else3 must be preserved even though adjustCopiesBackFrom 295; is tempted to remove it. 296; CHECK: vorr d 297define internal void @adjustCopiesBackFrom(<2 x i64>* noalias nocapture sret %agg.result, <2 x i64> %in) { 298entry: 299 %0 = extractelement <2 x i64> %in, i32 0 300 %cmp = icmp slt i64 %0, 1 301 %.in = select i1 %cmp, <2 x i64> <i64 0, i64 undef>, <2 x i64> %in 302 %1 = extractelement <2 x i64> %in, i32 1 303 %cmp1 = icmp slt i64 %1, 1 304 br i1 %cmp1, label %if.then2, label %if.else3 305 306if.then2: ; preds = %entry 307 %2 = insertelement <2 x i64> %.in, i64 0, i32 1 308 br label %if.end4 309 310if.else3: ; preds = %entry 311 %3 = shufflevector <2 x i64> %.in, <2 x i64> %in, <2 x i32> <i32 0, i32 3> 312 br label %if.end4 313 314if.end4: ; preds = %if.else3, %if.then2 315 %result.2 = phi <2 x i64> [ %2, %if.then2 ], [ %3, %if.else3 ] 316 store <2 x i64> %result.2, <2 x i64>* %agg.result, align 128 317 ret void 318} 319 320; <rdar://problem/12758887> 321; RegisterCoalescer::updateRegDefsUses() could visit an instruction more than 322; once under rare circumstances. When widening a register from QPR to DTriple 323; with the original virtual register in dsub_1_dsub_2, the double rewrite would 324; produce an invalid sub-register. 325; 326; This is because dsub_1_dsub_2 is not an idempotent sub-register index. 327; It will translate %vr:dsub_0 -> %vr:dsub_1. 328define hidden fastcc void @radar12758887() nounwind optsize ssp { 329entry: 330 br i1 undef, label %for.body, label %for.end70 331 332for.body: ; preds = %for.end, %entry 333 br i1 undef, label %for.body29, label %for.end 334 335for.body29: ; preds = %for.body29, %for.body 336 %0 = load <2 x double>, <2 x double>* null, align 1 337 %splat40 = shufflevector <2 x double> %0, <2 x double> undef, <2 x i32> zeroinitializer 338 %mul41 = fmul <2 x double> undef, %splat40 339 %add42 = fadd <2 x double> undef, %mul41 340 %splat44 = shufflevector <2 x double> %0, <2 x double> undef, <2 x i32> <i32 1, i32 1> 341 %mul45 = fmul <2 x double> undef, %splat44 342 %add46 = fadd <2 x double> undef, %mul45 343 br i1 undef, label %for.end, label %for.body29 344 345for.end: ; preds = %for.body29, %for.body 346 %accumR2.0.lcssa = phi <2 x double> [ zeroinitializer, %for.body ], [ %add42, %for.body29 ] 347 %accumI2.0.lcssa = phi <2 x double> [ zeroinitializer, %for.body ], [ %add46, %for.body29 ] 348 %1 = shufflevector <2 x double> %accumI2.0.lcssa, <2 x double> undef, <2 x i32> <i32 1, i32 0> 349 %add58 = fadd <2 x double> undef, %1 350 %mul61 = fmul <2 x double> %add58, undef 351 %add63 = fadd <2 x double> undef, %mul61 352 %add64 = fadd <2 x double> undef, %add63 353 %add67 = fadd <2 x double> undef, %add64 354 store <2 x double> %add67, <2 x double>* undef, align 1 355 br i1 undef, label %for.end70, label %for.body 356 357for.end70: ; preds = %for.end, %entry 358 ret void 359} 360