296 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 42; RUN: llc < %s | FileCheck %s3 4; Check that the SCEVs produced from the multiple loops don't attempt to get5; combines in invalid ways. The LSR filtering could attempt to combine addrecs6; from different loops.7 8target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"9target triple = "x86_64-unknown-linux-gnu"10 11define void @in4dob_(ptr nocapture writeonly %0, ptr nocapture readonly %1, ptr nocapture readonly %2, i64 %3, i1 %min.iters.check840) "target-cpu"="icelake-server" {12; CHECK-LABEL: in4dob_:13; CHECK: # %bb.0: # %.preheader26314; CHECK-NEXT: leaq (,%rcx,4), %r915; CHECK-NEXT: movl $1, %r10d16; CHECK-NEXT: xorl %eax, %eax17; CHECK-NEXT: vxorps %xmm0, %xmm0, %xmm018; CHECK-NEXT: jmp .LBB0_119; CHECK-NEXT: .p2align 420; CHECK-NEXT: .LBB0_20: # in Loop: Header=BB0_1 Depth=121; CHECK-NEXT: incq %r1022; CHECK-NEXT: addq %r9, %rax23; CHECK-NEXT: cmpq %r10, %rcx24; CHECK-NEXT: je .LBB0_1825; CHECK-NEXT: .LBB0_1: # =>This Inner Loop Header: Depth=126; CHECK-NEXT: vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero27; CHECK-NEXT: vucomiss %xmm0, %xmm128; CHECK-NEXT: jne .LBB0_2029; CHECK-NEXT: jp .LBB0_2030; CHECK-NEXT: # %bb.2: # in Loop: Header=BB0_1 Depth=131; CHECK-NEXT: vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero32; CHECK-NEXT: vucomiss %xmm0, %xmm133; CHECK-NEXT: jne .LBB0_2034; CHECK-NEXT: jp .LBB0_2035; CHECK-NEXT: # %bb.3: # %vector.body807.preheader36; CHECK-NEXT: leaq 1(%rcx), %rdx37; CHECK-NEXT: movl %edx, %esi38; CHECK-NEXT: andl $7, %esi39; CHECK-NEXT: cmpq $7, %rcx40; CHECK-NEXT: jae .LBB0_541; CHECK-NEXT: # %bb.4:42; CHECK-NEXT: xorl %r9d, %r9d43; CHECK-NEXT: jmp .LBB0_744; CHECK-NEXT: .LBB0_5: # %vector.body807.preheader.new45; CHECK-NEXT: movq %rdx, %r1046; CHECK-NEXT: andq $-8, %r1047; CHECK-NEXT: xorl %r9d, %r9d48; CHECK-NEXT: vxorps %xmm0, %xmm0, %xmm049; CHECK-NEXT: .p2align 450; CHECK-NEXT: .LBB0_6: # %vector.body80751; CHECK-NEXT: # =>This Inner Loop Header: Depth=152; CHECK-NEXT: leaq (%rdi,%r9), %r1153; CHECK-NEXT: vmovups %ymm0, (%rax,%r11)54; CHECK-NEXT: vmovups %ymm0, 1(%rax,%r11)55; CHECK-NEXT: vmovups %ymm0, 2(%rax,%r11)56; CHECK-NEXT: vmovups %ymm0, 3(%rax,%r11)57; CHECK-NEXT: vmovups %ymm0, 4(%rax,%r11)58; CHECK-NEXT: vmovups %ymm0, 5(%rax,%r11)59; CHECK-NEXT: vmovups %ymm0, 6(%rax,%r11)60; CHECK-NEXT: vmovups %ymm0, 7(%rax,%r11)61; CHECK-NEXT: addq $8, %r962; CHECK-NEXT: cmpq %r9, %r1063; CHECK-NEXT: jne .LBB0_664; CHECK-NEXT: .LBB0_7: # %.lr.ph373.unr-lcssa65; CHECK-NEXT: testq %rsi, %rsi66; CHECK-NEXT: je .LBB0_1067; CHECK-NEXT: # %bb.8: # %vector.body807.epil.preheader68; CHECK-NEXT: addq %rdi, %r969; CHECK-NEXT: xorl %r10d, %r10d70; CHECK-NEXT: vxorps %xmm0, %xmm0, %xmm071; CHECK-NEXT: .p2align 472; CHECK-NEXT: .LBB0_9: # %vector.body807.epil73; CHECK-NEXT: # =>This Inner Loop Header: Depth=174; CHECK-NEXT: leaq (%r9,%r10), %r1175; CHECK-NEXT: vmovups %ymm0, (%rax,%r11)76; CHECK-NEXT: incq %r1077; CHECK-NEXT: cmpq %r10, %rsi78; CHECK-NEXT: jne .LBB0_979; CHECK-NEXT: .LBB0_10: # %.lr.ph37380; CHECK-NEXT: testb $1, %r8b81; CHECK-NEXT: je .LBB0_1182; CHECK-NEXT: # %bb.19: # %scalar.ph839.preheader83; CHECK-NEXT: movl $0, (%rdi)84; CHECK-NEXT: vzeroupper85; CHECK-NEXT: retq86; CHECK-NEXT: .LBB0_11: # %vector.body847.preheader87; CHECK-NEXT: movl %edx, %esi88; CHECK-NEXT: andl $7, %esi89; CHECK-NEXT: cmpq $7, %rcx90; CHECK-NEXT: jae .LBB0_1391; CHECK-NEXT: # %bb.12:92; CHECK-NEXT: xorl %ecx, %ecx93; CHECK-NEXT: jmp .LBB0_1594; CHECK-NEXT: .LBB0_13: # %vector.body847.preheader.new95; CHECK-NEXT: andq $-8, %rdx96; CHECK-NEXT: xorl %ecx, %ecx97; CHECK-NEXT: vxorps %xmm0, %xmm0, %xmm098; CHECK-NEXT: .p2align 499; CHECK-NEXT: .LBB0_14: # %vector.body847100; CHECK-NEXT: # =>This Inner Loop Header: Depth=1101; CHECK-NEXT: leaq (%rdi,%rcx), %r8102; CHECK-NEXT: vmovups %ymm0, 96(%rax,%r8)103; CHECK-NEXT: vmovups %ymm0, 97(%rax,%r8)104; CHECK-NEXT: vmovups %ymm0, 98(%rax,%r8)105; CHECK-NEXT: vmovups %ymm0, 99(%rax,%r8)106; CHECK-NEXT: vmovups %ymm0, 100(%rax,%r8)107; CHECK-NEXT: vmovups %ymm0, 101(%rax,%r8)108; CHECK-NEXT: vmovups %ymm0, 102(%rax,%r8)109; CHECK-NEXT: vmovups %ymm0, 103(%rax,%r8)110; CHECK-NEXT: addq $8, %rcx111; CHECK-NEXT: cmpq %rcx, %rdx112; CHECK-NEXT: jne .LBB0_14113; CHECK-NEXT: .LBB0_15: # %common.ret.loopexit.unr-lcssa114; CHECK-NEXT: testq %rsi, %rsi115; CHECK-NEXT: je .LBB0_18116; CHECK-NEXT: # %bb.16: # %vector.body847.epil.preheader117; CHECK-NEXT: leaq 96(%rcx,%rdi), %rcx118; CHECK-NEXT: xorl %edx, %edx119; CHECK-NEXT: vxorps %xmm0, %xmm0, %xmm0120; CHECK-NEXT: .p2align 4121; CHECK-NEXT: .LBB0_17: # %vector.body847.epil122; CHECK-NEXT: # =>This Inner Loop Header: Depth=1123; CHECK-NEXT: leaq (%rcx,%rdx), %rdi124; CHECK-NEXT: vmovups %ymm0, (%rax,%rdi)125; CHECK-NEXT: incq %rdx126; CHECK-NEXT: cmpq %rdx, %rsi127; CHECK-NEXT: jne .LBB0_17128; CHECK-NEXT: .LBB0_18: # %common.ret129; CHECK-NEXT: vzeroupper130; CHECK-NEXT: retq131.preheader263:132 %4 = shl i64 %3, 2133 br label %5134 1355: ; preds = %16, %.preheader263136 %lsr.iv1135 = phi ptr [ %0, %.preheader263 ], [ %uglygep1136, %16 ]137 %indvars.iv487 = phi i64 [ 1, %.preheader263 ], [ %indvars.iv.next488, %16 ]138 %6 = getelementptr float, ptr %1, i64 %indvars.iv487139 %7 = load float, ptr %6, align 4140 %8 = fcmp oeq float %7, 0.000000e+00141 %9 = getelementptr float, ptr %2, i64 %indvars.iv487142 %10 = load float, ptr %9, align 4143 %11 = fcmp oeq float %10, 0.000000e+00144 %12 = and i1 %8, %11145 br i1 %12, label %vector.body807.preheader, label %16146 147vector.body807.preheader: ; preds = %5148 %13 = add i64 %3, 1149 %xtraiter = and i64 %13, 7150 %14 = icmp ult i64 %3, 7151 br i1 %14, label %.lr.ph373.unr-lcssa, label %vector.body807.preheader.new152 153vector.body807.preheader.new: ; preds = %vector.body807.preheader154 %unroll_iter = and i64 %13, -8155 br label %vector.body807156 157vector.body807: ; preds = %vector.body807, %vector.body807.preheader.new158 %lsr.iv1194 = phi i64 [ 0, %vector.body807.preheader.new ], [ %lsr.iv.next1195.7, %vector.body807 ]159 %niter = phi i64 [ 0, %vector.body807.preheader.new ], [ %niter.next.7, %vector.body807 ]160 %uglygep1197 = getelementptr i8, ptr %lsr.iv1135, i64 %lsr.iv1194161 store <8 x float> zeroinitializer, ptr %uglygep1197, align 4162 %lsr.iv.next1195 = or disjoint i64 %lsr.iv1194, 1163 %uglygep1197.1 = getelementptr i8, ptr %lsr.iv1135, i64 %lsr.iv.next1195164 store <8 x float> zeroinitializer, ptr %uglygep1197.1, align 4165 %lsr.iv.next1195.1 = or disjoint i64 %lsr.iv1194, 2166 %uglygep1197.2 = getelementptr i8, ptr %lsr.iv1135, i64 %lsr.iv.next1195.1167 store <8 x float> zeroinitializer, ptr %uglygep1197.2, align 4168 %lsr.iv.next1195.2 = or disjoint i64 %lsr.iv1194, 3169 %uglygep1197.3 = getelementptr i8, ptr %lsr.iv1135, i64 %lsr.iv.next1195.2170 store <8 x float> zeroinitializer, ptr %uglygep1197.3, align 4171 %lsr.iv.next1195.3 = or disjoint i64 %lsr.iv1194, 4172 %uglygep1197.4 = getelementptr i8, ptr %lsr.iv1135, i64 %lsr.iv.next1195.3173 store <8 x float> zeroinitializer, ptr %uglygep1197.4, align 4174 %lsr.iv.next1195.4 = or disjoint i64 %lsr.iv1194, 5175 %uglygep1197.5 = getelementptr i8, ptr %lsr.iv1135, i64 %lsr.iv.next1195.4176 store <8 x float> zeroinitializer, ptr %uglygep1197.5, align 4177 %lsr.iv.next1195.5 = or disjoint i64 %lsr.iv1194, 6178 %uglygep1197.6 = getelementptr i8, ptr %lsr.iv1135, i64 %lsr.iv.next1195.5179 store <8 x float> zeroinitializer, ptr %uglygep1197.6, align 4180 %lsr.iv.next1195.6 = or disjoint i64 %lsr.iv1194, 7181 %uglygep1197.7 = getelementptr i8, ptr %lsr.iv1135, i64 %lsr.iv.next1195.6182 store <8 x float> zeroinitializer, ptr %uglygep1197.7, align 4183 %lsr.iv.next1195.7 = add i64 %lsr.iv1194, 8184 %niter.next.7 = add i64 %niter, 8185 %niter.ncmp.7 = icmp eq i64 %niter.next.7, %unroll_iter186 br i1 %niter.ncmp.7, label %.lr.ph373.unr-lcssa.loopexit, label %vector.body807187 188.lr.ph373.unr-lcssa.loopexit: ; preds = %vector.body807189 br label %.lr.ph373.unr-lcssa190 191.lr.ph373.unr-lcssa: ; preds = %.lr.ph373.unr-lcssa.loopexit, %vector.body807.preheader192 %lsr.iv1194.unr = phi i64 [ 0, %vector.body807.preheader ], [ %lsr.iv.next1195.7, %.lr.ph373.unr-lcssa.loopexit ]193 %lcmp.mod.not = icmp eq i64 %xtraiter, 0194 br i1 %lcmp.mod.not, label %.lr.ph373, label %vector.body807.epil.preheader195 196vector.body807.epil.preheader: ; preds = %.lr.ph373.unr-lcssa197 br label %vector.body807.epil198 199vector.body807.epil: ; preds = %vector.body807.epil.preheader, %vector.body807.epil200 %lsr.iv1194.epil = phi i64 [ %lsr.iv.next1195.epil, %vector.body807.epil ], [ %lsr.iv1194.unr, %vector.body807.epil.preheader ]201 %epil.iter = phi i64 [ %epil.iter.next, %vector.body807.epil ], [ 0, %vector.body807.epil.preheader ]202 %uglygep1197.epil = getelementptr i8, ptr %lsr.iv1135, i64 %lsr.iv1194.epil203 store <8 x float> zeroinitializer, ptr %uglygep1197.epil, align 4204 %lsr.iv.next1195.epil = add i64 %lsr.iv1194.epil, 1205 %epil.iter.next = add i64 %epil.iter, 1206 %epil.iter.cmp.not = icmp eq i64 %epil.iter.next, %xtraiter207 br i1 %epil.iter.cmp.not, label %.lr.ph373.loopexit, label %vector.body807.epil208 209.lr.ph373.loopexit: ; preds = %vector.body807.epil210 br label %.lr.ph373211 212.lr.ph373: ; preds = %.lr.ph373.loopexit, %.lr.ph373.unr-lcssa213 br i1 %min.iters.check840, label %scalar.ph839.preheader, label %vector.body847.preheader214 215vector.body847.preheader: ; preds = %.lr.ph373216 %uglygep11551 = getelementptr i8, ptr %lsr.iv1135, i64 96217 %xtraiter12 = and i64 %13, 7218 %15 = icmp ult i64 %3, 7219 br i1 %15, label %common.ret.loopexit.unr-lcssa, label %vector.body847.preheader.new220 221vector.body847.preheader.new: ; preds = %vector.body847.preheader222 %unroll_iter15 = and i64 %13, -8223 br label %vector.body847224 225vector.body847: ; preds = %vector.body847, %vector.body847.preheader.new226 %lsr.iv1152 = phi i64 [ 0, %vector.body847.preheader.new ], [ %lsr.iv.next1153.7, %vector.body847 ]227 %niter16 = phi i64 [ 0, %vector.body847.preheader.new ], [ %niter16.next.7, %vector.body847 ]228 %uglygep1156 = getelementptr i8, ptr %uglygep11551, i64 %lsr.iv1152229 store <8 x float> zeroinitializer, ptr %uglygep1156, align 4230 %lsr.iv.next1153 = or disjoint i64 %lsr.iv1152, 1231 %uglygep1156.1 = getelementptr i8, ptr %uglygep11551, i64 %lsr.iv.next1153232 store <8 x float> zeroinitializer, ptr %uglygep1156.1, align 4233 %lsr.iv.next1153.1 = or disjoint i64 %lsr.iv1152, 2234 %uglygep1156.2 = getelementptr i8, ptr %uglygep11551, i64 %lsr.iv.next1153.1235 store <8 x float> zeroinitializer, ptr %uglygep1156.2, align 4236 %lsr.iv.next1153.2 = or disjoint i64 %lsr.iv1152, 3237 %uglygep1156.3 = getelementptr i8, ptr %uglygep11551, i64 %lsr.iv.next1153.2238 store <8 x float> zeroinitializer, ptr %uglygep1156.3, align 4239 %lsr.iv.next1153.3 = or disjoint i64 %lsr.iv1152, 4240 %uglygep1156.4 = getelementptr i8, ptr %uglygep11551, i64 %lsr.iv.next1153.3241 store <8 x float> zeroinitializer, ptr %uglygep1156.4, align 4242 %lsr.iv.next1153.4 = or disjoint i64 %lsr.iv1152, 5243 %uglygep1156.5 = getelementptr i8, ptr %uglygep11551, i64 %lsr.iv.next1153.4244 store <8 x float> zeroinitializer, ptr %uglygep1156.5, align 4245 %lsr.iv.next1153.5 = or disjoint i64 %lsr.iv1152, 6246 %uglygep1156.6 = getelementptr i8, ptr %uglygep11551, i64 %lsr.iv.next1153.5247 store <8 x float> zeroinitializer, ptr %uglygep1156.6, align 4248 %lsr.iv.next1153.6 = or disjoint i64 %lsr.iv1152, 7249 %uglygep1156.7 = getelementptr i8, ptr %uglygep11551, i64 %lsr.iv.next1153.6250 store <8 x float> zeroinitializer, ptr %uglygep1156.7, align 4251 %lsr.iv.next1153.7 = add i64 %lsr.iv1152, 8252 %niter16.next.7 = add i64 %niter16, 8253 %niter16.ncmp.7 = icmp eq i64 %niter16.next.7, %unroll_iter15254 br i1 %niter16.ncmp.7, label %common.ret.loopexit.unr-lcssa.loopexit, label %vector.body847255 256common.ret.loopexit.unr-lcssa.loopexit: ; preds = %vector.body847257 br label %common.ret.loopexit.unr-lcssa258 259common.ret.loopexit.unr-lcssa: ; preds = %common.ret.loopexit.unr-lcssa.loopexit, %vector.body847.preheader260 %lsr.iv1152.unr = phi i64 [ 0, %vector.body847.preheader ], [ %lsr.iv.next1153.7, %common.ret.loopexit.unr-lcssa.loopexit ]261 %lcmp.mod14.not = icmp eq i64 %xtraiter12, 0262 br i1 %lcmp.mod14.not, label %common.ret, label %vector.body847.epil.preheader263 264vector.body847.epil.preheader: ; preds = %common.ret.loopexit.unr-lcssa265 br label %vector.body847.epil266 267vector.body847.epil: ; preds = %vector.body847.epil.preheader, %vector.body847.epil268 %lsr.iv1152.epil = phi i64 [ %lsr.iv.next1153.epil, %vector.body847.epil ], [ %lsr.iv1152.unr, %vector.body847.epil.preheader ]269 %epil.iter13 = phi i64 [ %epil.iter13.next, %vector.body847.epil ], [ 0, %vector.body847.epil.preheader ]270 %uglygep1156.epil = getelementptr i8, ptr %uglygep11551, i64 %lsr.iv1152.epil271 store <8 x float> zeroinitializer, ptr %uglygep1156.epil, align 4272 %lsr.iv.next1153.epil = add i64 %lsr.iv1152.epil, 1273 %epil.iter13.next = add i64 %epil.iter13, 1274 %epil.iter13.cmp.not = icmp eq i64 %epil.iter13.next, %xtraiter12275 br i1 %epil.iter13.cmp.not, label %common.ret.loopexit, label %vector.body847.epil276 277common.ret.loopexit: ; preds = %vector.body847.epil278 br label %common.ret279 280common.ret.loopexit1: ; preds = %16281 br label %common.ret282 283common.ret: ; preds = %common.ret.loopexit1, %common.ret.loopexit, %scalar.ph839.preheader, %common.ret.loopexit.unr-lcssa284 ret void285 286scalar.ph839.preheader: ; preds = %.lr.ph373287 store float 0.000000e+00, ptr %0, align 4288 br label %common.ret289 29016: ; preds = %5291 %indvars.iv.next488 = add i64 %indvars.iv487, 1292 %exitcond492.not = icmp eq i64 %indvars.iv.next488, %3293 %uglygep1136 = getelementptr i8, ptr %lsr.iv1135, i64 %4294 br i1 %exitcond492.not, label %common.ret.loopexit1, label %5295}296