llvm/test/Analysis/LoopCacheAnalysis/partially-perfect-nest.ll

; NOTE: Assertions have been autogenerated by utils/update_analyze_test_checks.py UTC_ARGS: --version 6
; RUN: opt < %s -cache-line-size=64 -passes='print<loop-cache-cost>' -disable-output 2>&1 | FileCheck %s

;; FIXME: For a forked nest getInnerMostLoop() fails to
;; detect the sibling sub-loops and picks a single wrong innermost loop.

;; for (i) {
;;   for (j)
;;     for (jj) A[jj][j] = 1;
;;   for (k)
;;     for (kk) B[k][kk] = 1;
;; }
;; FIXME: only the 'k/kk' sub-nest is costed, 'a.i' (a fork) should not have a
;; cost, and 'a.j'/'a.jj' are wrongly costed as loop-invariant.

define void @fork_at_top(ptr %A, ptr %B) {
; CHECK-LABEL: 'fork_at_top'
; CHECK-NEXT:  Loop 'a.k' has cost = 10000000000
; CHECK-NEXT:  Loop 'a.kk' has cost = 1300000000
; CHECK-NEXT:  Loop 'a.i' has cost = 100000000
; CHECK-NEXT:  Loop 'a.j' has cost = 100000000
; CHECK-NEXT:  Loop 'a.jj' has cost = 100000000
;
entry:
  br label %a.i
a.i:
  %iv = phi i64 [ 0, %entry ], [ %iv.n, %a.i.inc ]
  br label %a.j
a.j:
  %jv = phi i64 [ 0, %a.i ], [ %jv.n, %a.j.inc ]
  br label %a.jj
a.jj:
  %jjv = phi i64 [ 0, %a.j ], [ %jjv.n, %a.jj ]
  %ga = getelementptr inbounds [100 x double], ptr %A, i64 %jjv, i64 %jv
  store double 1.0, ptr %ga
  %jjv.n = add nsw i64 %jjv, 1
  %jj.e = icmp eq i64 %jjv.n, 100
  br i1 %jj.e, label %a.j.inc, label %a.jj
a.j.inc:
  %jv.n = add nsw i64 %jv, 1
  %j.e = icmp eq i64 %jv.n, 100
  br i1 %j.e, label %a.k, label %a.j
a.k:
  %kv = phi i64 [ 0, %a.j.inc ], [ %kv.n, %a.k.inc ]
  br label %a.kk
a.kk:
  %kkv = phi i64 [ 0, %a.k ], [ %kkv.n, %a.kk ]
  %gb = getelementptr inbounds [100 x double], ptr %B, i64 %kv, i64 %kkv
  store double 1.0, ptr %gb
  %kkv.n = add nsw i64 %kkv, 1
  %kk.e = icmp eq i64 %kkv.n, 100
  br i1 %kk.e, label %a.k.inc, label %a.kk
a.k.inc:
  %kv.n = add nsw i64 %kv, 1
  %k.e = icmp eq i64 %kv.n, 100
  br i1 %k.e, label %a.i.inc, label %a.k
a.i.inc:
  %iv.n = add nsw i64 %iv, 1
  %i.e = icmp eq i64 %iv.n, 100
  br i1 %i.e, label %exit, label %a.i
exit:
  ret void
}

;; for (i)
;;   for (j) {
;;     for (k) C[k] = 1;
;;     for (l) D[l] = 1;
;;   }
;; FIXME: only the 'l' sub-nest is costed, and the fork loops 'b.i'/'b.j' should
;; not have a cost.

define void @fork_in_middle(ptr %C, ptr %D) {
; CHECK-LABEL: 'fork_in_middle'
; CHECK-NEXT:  Loop 'b.l' has cost = 13000000
; CHECK-NEXT:  Loop 'b.i' has cost = 1000000
; CHECK-NEXT:  Loop 'b.j' has cost = 1000000
; CHECK-NEXT:  Loop 'b.k' has cost = 1000000
;
entry:
  br label %b.i
b.i:
  %iv = phi i64 [ 0, %entry ], [ %iv.n, %b.i.inc ]
  br label %b.j
b.j:
  %jv = phi i64 [ 0, %b.i ], [ %jv.n, %b.j.inc ]
  br label %b.k
b.k:
  %kv = phi i64 [ 0, %b.j ], [ %kv.n, %b.k ]
  %gc = getelementptr inbounds double, ptr %C, i64 %kv
  store double 1.0, ptr %gc
  %kv.n = add nsw i64 %kv, 1
  %k.e = icmp eq i64 %kv.n, 100
  br i1 %k.e, label %b.l, label %b.k
b.l:
  %lv = phi i64 [ 0, %b.k ], [ %lv.n, %b.l ]
  %gd = getelementptr inbounds double, ptr %D, i64 %lv
  store double 1.0, ptr %gd
  %lv.n = add nsw i64 %lv, 1
  %l.e = icmp eq i64 %lv.n, 100
  br i1 %l.e, label %b.j.inc, label %b.l
b.j.inc:
  %jv.n = add nsw i64 %jv, 1
  %j.e = icmp eq i64 %jv.n, 100
  br i1 %j.e, label %b.i.inc, label %b.j
b.i.inc:
  %iv.n = add nsw i64 %iv, 1
  %i.e = icmp eq i64 %iv.n, 100
  br i1 %i.e, label %exit, label %b.i
exit:
  ret void
}