Huffman Benchmark has opportunities to fold addressing modes

[sivarv]

Huffman benchmark uses an array of structs

For e.g. the innermost triply nested while loop

while(hufftree[i].left!=-1)
{
    if(GetCompBit(comparray,bitoffset)==0)
        i=hufftree[i].left;
    else
       i=hufftree[i].right;
    bitoffset++;
}

CSE recognizes bounds checked &hufftree[i] as CSE candidate and replaces with a LclVar.

bounds check on hufftree[i]
tmp = LEA(&hufftree[i])

Now field accesses will turn into [reg+fieldoffset]. If we could fold tmp into field access, it will turn into
[huffTree + indexReg*scale + (16+fieldoffset)]. This avoids the need for a reg to tmp. In fact, for tmp there were no callee saved regs available and its gets spilled and reloaded across the call.

There are many such opportunities in Huffman benchmark to fold addr modes as it heavily accesses array of structs.

category:cq
theme:addressing-modes
skill-level:intermediate
cost:medium

[russellhadley]

@sivarv can we look at MusicStore or Techempower to see if we have the same problem there? It would help prioritize this if we knew this hit in those workloads.