92 lines · plain
1; RUN: llc -O3 -mtriple=hexagon -mcpu=hexagonv5 < %s | FileCheck %s2;3; Generate hardware loop when unknown trip count loop is vectorized.4 5; CHECK: loop0(.LBB{{[0-9]*}}_{{[0-9]*}},r{{[0-9]+}})6; CHECK: endloop07; CHECK: loop0(.LBB{{[0-9]*}}_{{[0-9]*}},r{{[0-9]+}})8; CHECK: endloop09 10@A = common global [1000 x i32] zeroinitializer, align 811@B = common global [1000 x i32] zeroinitializer, align 812 13define i32 @dotprod2(i32 %count) #0 {14entry.split:15 %cmp6 = icmp sgt i32 %count, 016 br i1 %cmp6, label %polly.cond, label %for.end17 18for.end.loopexit:19 br label %for.end20 21for.end:22 %sum.0.lcssa.reg2mem.0.load37 = phi i32 [ 0, %entry.split ], [ %p_add34, %polly.loop_if13 ], [ %p_add, %for.end.loopexit ]23 ret i32 %sum.0.lcssa.reg2mem.0.load3724 25polly.cond:26 %0 = icmp sgt i32 %count, 127 br i1 %0, label %polly.loop_if, label %polly.loop_if1328 29polly.loop_exit.loopexit:30 br label %polly.loop_exit31 32polly.loop_exit:33 %1 = phi <2 x i32> [ zeroinitializer, %polly.loop_if ], [ %addp_vec, %polly.loop_exit.loopexit ]34 %2 = extractelement <2 x i32> %1, i32 035 %3 = extractelement <2 x i32> %1, i32 136 %add_sum = add i32 %2, %337 br label %polly.loop_if1338 39polly.loop_if:40 %4 = add i32 %count, -141 %leftover_lb = and i32 %4, -242 %polly.loop_guard = icmp eq i32 %leftover_lb, 043 br i1 %polly.loop_guard, label %polly.loop_exit, label %polly.loop_preheader44 45polly.stmt.for.body:46 %addp_vec28 = phi <2 x i32> [ zeroinitializer, %polly.loop_preheader ], [ %addp_vec, %polly.stmt.for.body ]47 %scevgep.phi = phi ptr [ @A, %polly.loop_preheader ], [ %scevgep.inc, %polly.stmt.for.body ]48 %scevgep9.phi = phi ptr [ @B, %polly.loop_preheader ], [ %scevgep9.inc, %polly.stmt.for.body ]49 %polly.indvar = phi i32 [ 0, %polly.loop_preheader ], [ %polly.indvar_next, %polly.stmt.for.body ]50 %_p_vec_full = load <2 x i32>, ptr %scevgep.phi, align 851 %_p_vec_full11 = load <2 x i32>, ptr %scevgep9.phi, align 852 %mulp_vec = mul <2 x i32> %_p_vec_full11, %_p_vec_full53 %addp_vec = add <2 x i32> %mulp_vec, %addp_vec2854 %polly.indvar_next = add nsw i32 %polly.indvar, 255 %polly.loop_cond = icmp eq i32 %polly.indvar, %polly.adjust_ub56 %scevgep.inc = getelementptr i32, ptr %scevgep.phi, i32 257 %scevgep9.inc = getelementptr i32, ptr %scevgep9.phi, i32 258 br i1 %polly.loop_cond, label %polly.loop_exit.loopexit, label %polly.stmt.for.body59 60polly.loop_preheader:61 %polly.adjust_ub = add i32 %leftover_lb, -262 br label %polly.stmt.for.body63 64polly.loop_if13:65 %p_add34 = phi i32 [ 0, %polly.cond ], [ %add_sum, %polly.loop_exit ]66 %merge.lb = phi i32 [ 0, %polly.cond ], [ %leftover_lb, %polly.loop_exit ]67 %polly.loop_guard17 = icmp slt i32 %merge.lb, %count68 br i1 %polly.loop_guard17, label %polly.loop_preheader15, label %for.end69 70polly.stmt.for.body22:71 %p_add30 = phi i32 [ %p_add34, %polly.loop_preheader15 ], [ %p_add, %polly.stmt.for.body22 ]72 %polly.indvar18 = phi i32 [ %merge.lb, %polly.loop_preheader15 ], [ %polly.indvar_next19, %polly.stmt.for.body22 ]73 %5 = tail call i32 @llvm.annotation.i32(i32 %polly.indvar18, ptr null, ptr null, i32 0), !polly.loop.smallTripCount !074 %scevgep23 = getelementptr [1000 x i32], ptr @A, i32 0, i32 %polly.indvar1875 %_p_scalar_ = load i32, ptr %scevgep23, align 476 %scevgep24 = getelementptr [1000 x i32], ptr @B, i32 0, i32 %polly.indvar1877 %_p_scalar_25 = load i32, ptr %scevgep24, align 478 %p_mul = mul nsw i32 %_p_scalar_25, %_p_scalar_79 %p_add = add nsw i32 %p_mul, %p_add3080 %polly.indvar_next19 = add nsw i32 %polly.indvar18, 181 %polly.loop_cond21 = icmp slt i32 %polly.indvar18, %polly.adjust_ub2082 br i1 %polly.loop_cond21, label %polly.stmt.for.body22, label %for.end.loopexit83 84polly.loop_preheader15:85 %polly.adjust_ub20 = add i32 %count, -186 br label %polly.stmt.for.body2287}88 89declare i32 @llvm.annotation.i32(i32, ptr, ptr, i32) #190 91!0 = !{}92