[PATCH] D31068: [SDAG] Expand MergedConsecutiveStores to better handle Giving up in Chain Analysis

[Aditya Nandakumar via llvm-commits]

The test case I sent earlier is what I’m using too. The DAG right before DAGCombiner::MergeConsecutiveStores is


SelectionDAG has 47 nodes:
  t0: ch = EntryToken
  t2: i32,ch = CopyFromReg t0, Register:i32 %vreg2
  t3: ch = ValueType:i32
  t5: i32,ch = CopyFromReg t2:1, Register:i32 %vreg3
  t6: i64 = build_pair t2, t5
  t40: i32 = Constant<0>
    t43: i32 = add t2, Constant:i32<46>
            t44: i32 = add t2, Constant:i32<46>
          t45: v2i16 = bitcast t44
        t46: i16 = add t45, Constant:i32<1>
      t47: i32 = zero_extend t46
    t48: i32 = add t5, t47
  t49: i64 = build_pair t43, t48
                  t8: i64 = add t6, Constant:i64<2>
                t12: ch = store<ST1[%dst.gep2.i105.2](align=2)> t0, Constant:i8<0>, t8, undef:i64
                t14: i64 = add t6, Constant:i64<3>
              t15: ch = store<ST1[%dst.gep2.i105.3]> t12, Constant:i8<0>, t14, undef:i64
              t17: i64 = add t6, Constant:i64<4>
            t18: ch = store<ST1[%dst.gep2.i105.4](align=2)> t15, Constant:i8<0>, t17, undef:i64
              t20: i64 = add t6, Constant:i64<5>
            t50: ch = store<ST1[%dst.gep2.i105.5]> t0, Constant:i8<0>, t20, undef:i64
          t51: ch = TokenFactor t18, t50
        t22: ch = store<ST1[%dst.gep2.i105.61](align=2)> t51, Constant:i8<0>, Constant:i64<0>, undef:i64
          t27: i32 = extract_element t49, Constant:i64<0>
        t29: ch = CopyToReg t0, Register:i32 %vreg4, t27
          t26: i32 = extract_element t49, Constant:i64<1>
        t31: ch = CopyToReg t0, Register:i32 %vreg5, t26
      t39: ch = TokenFactor t22, t29, t31
    t36: ch = brcond t39, Constant:i1<0>, BasicBlock:ch<loop.i108Exit 0x107801c58>
  t38: ch = br t36, BasicBlock:ch<loop.i108Split 0x1078018c8>


Please look at how t50 will not be merged with the current implementation.

> On Apr 11, 2017, at 5:41 AM, Nirav Dave via Phabricator <reviews at reviews.llvm.org> wrote:
> 
> niravd added a comment.
> 
> I'm a little concerned about the selection candidates having a chain loop dependence after mething. Can you get a llvm testcase I can look at myself. Alternatively I can probably cobble one myself if I can get a snapshot of the DAG right before the relevant merge
> 
> In https://reviews.llvm.org/D31068#723192, @aditya_nandakumar wrote:
> 
>> Nirav - I posted the reason for the regression and a quick fix for the issue I saw in https://reviews.llvm.org/D31914.
> 
> 
> 
> 
> 
> https://reviews.llvm.org/D31068
> 
> 
>