[PATCH] D21449: Target independent codesize heuristics for Loop Idiom Recognition

[Sunita_Marathe via llvm-commits]

Sunita_Marathe added inline comments.

================
Comment at: test/Transforms/LoopIdiom/lir-heurs-multi-block-loop.ll:14
@@ +13,3 @@
+; CHECK-LABEL: LoopMemset
+; CHECK: memset
+;
----------------
mcrosier wrote:
> I don't understand what this (or the next test) is testing.  I believe you need two memsets that can be merged together in both cases to test this properly.
LoopMemset test:
============

   void LoopMemset(char (*__restrict D) [2048], unsigned int N)
   {
     unsigned int i, j;
     for (i=0; i<N; ++i) {
       memset(D[i], -1, 2048);
       if (i>15)  GF();
     }
   }

processLoopMemSet() optimizes the above loop memset to a non-loop memset which acts on the contiguous memory formed by concatenating the memory segments handled by the memset of each iteration.


NestedMemset_LoopMemset  test:
=======================

void NestedMemset_LoopMemset(char (*__restrict D) [2046], unsigned int N)
{
  unsigned int i, j;
  for (i=0; i<N; ++i) {
    for (j=0; j<2046; ++j) {
      D[i][j] = 0; // NestedMemset becomes LoopMemset
      if (j>15)  GF();
    }
  }
}

The LIR optimization for this test case happens in two steps:
- loop memset created when processLoopStridedStore() recognizes the D[I][j] store as a memset idiom.
- non-loop memset created when processLoopMemSet() recognizes the loop memset.



https://reviews.llvm.org/D21449