78 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_analyze_test_checks.py UTC_ARGS: --version 52; RUN: opt < %s -passes='print<delinearization>' -disable-output 2>&1 | FileCheck %s3 4target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"5 6; Function Attrs: noinline nounwind uwtable7define void @mat_mul(ptr %C, ptr %A, ptr %B, i64 %N) !kernel_arg_addr_space !2 !kernel_arg_access_qual !3 !kernel_arg_type !4 !kernel_arg_base_type !4 !kernel_arg_type_qual !5 {8; CHECK-LABEL: 'mat_mul'9; CHECK-NEXT: Inst: %tmp = load float, ptr %arrayidx, align 410; CHECK-NEXT: AccessFunction: {(4 * %N * %call),+,4}<%for.inc>11; CHECK-NEXT: Base offset: %A12; CHECK-NEXT: ArrayDecl[UnknownSize][%N] with elements of 4 bytes.13; CHECK-NEXT: ArrayRef[%call][{0,+,1}<nuw><nsw><%for.inc>]14; CHECK-NEXT: Delinearization validation: Succeeded15; CHECK-EMPTY:16; CHECK-NEXT: Inst: %tmp5 = load float, ptr %arrayidx4, align 417; CHECK-NEXT: AccessFunction: {(4 * %call1),+,(4 * %N)}<%for.inc>18; CHECK-NEXT: Base offset: %B19; CHECK-NEXT: ArrayDecl[UnknownSize][%N] with elements of 4 bytes.20; CHECK-NEXT: ArrayRef[{0,+,1}<nuw><nsw><%for.inc>][%call1]21; CHECK-NEXT: Delinearization validation: Failed22;23entry:24 br label %entry.split25 26entry.split: ; preds = %entry27 %call = tail call i64 @_Z13get_global_idj(i32 0)28 %call1 = tail call i64 @_Z13get_global_idj(i32 1)29 %cmp1 = icmp sgt i64 %N, 030 %mul = mul nsw i64 %call, %N31 br i1 %cmp1, label %for.inc.lr.ph, label %for.end32 33for.inc.lr.ph: ; preds = %entry.split34 br label %for.inc35 36for.inc: ; preds = %for.inc.lr.ph, %for.inc37 %acc.03 = phi float [ 0.000000e+00, %for.inc.lr.ph ], [ %tmp6, %for.inc ]38 %m.02 = phi i64 [ 0, %for.inc.lr.ph ], [ %inc, %for.inc ]39 %add = add nsw i64 %m.02, %mul40 %arrayidx = getelementptr inbounds float, ptr %A, i64 %add41 %tmp = load float, ptr %arrayidx, align 442 %mul2 = mul nsw i64 %m.02, %N43 %add3 = add nsw i64 %mul2, %call144 %arrayidx4 = getelementptr inbounds float, ptr %B, i64 %add345 %tmp5 = load float, ptr %arrayidx4, align 446 %tmp6 = tail call float @llvm.fmuladd.f32(float %tmp, float %tmp5, float %acc.03)47 %inc = add nuw nsw i64 %m.02, 148 %exitcond = icmp ne i64 %inc, %N49 br i1 %exitcond, label %for.inc, label %for.cond.for.end_crit_edge50 51for.cond.for.end_crit_edge: ; preds = %for.inc52 %.lcssa = phi float [ %tmp6, %for.inc ]53 br label %for.end54 55for.end: ; preds = %for.cond.for.end_crit_edge, %entry.split56 %acc.0.lcssa = phi float [ %.lcssa, %for.cond.for.end_crit_edge ], [ 0.000000e+00, %entry.split ]57 %add7 = add nsw i64 %mul, %call158 %arrayidx8 = getelementptr inbounds float, ptr %C, i64 %add759 store float %acc.0.lcssa, ptr %arrayidx8, align 460 ret void61}62 63; Function Attrs: nounwind readnone64declare i64 @_Z13get_global_idj(i32)65 66; Function Attrs: nounwind readnone speculatable67declare float @llvm.fmuladd.f32(float, float, float)68 69!llvm.module.flags = !{!0}70!llvm.ident = !{!1}71 72!0 = !{i32 1, !"wchar_size", i32 4}73!1 = !{!"clang version 5.0.0 (trunk 303846) (llvm/trunk 303834)"}74!2 = !{i32 1, i32 1, i32 1, i32 0}75!3 = !{!"none", !"none", !"none", !"none"}76!4 = !{!"float*", !"float*", !"float*", !"long"}77!5 = !{!"", !"", !"", !""}78