RFR: 8277850: C2: optimize mask checks in counted loops

[Roland Westrelin]

On Sat, 4 Dec 2021 00:33:04 GMT, Vladimir Kozlov <kvn at openjdk.org> wrote:

>> This is another fix that addresses a performance issue with panama and that was brought up by Maurizio. The pattern to optimize is:
>> if ((base + (offset << 2)) & 3) != 0) {
>> }
>> 
>> where base is loop independent but offset depends on a loop variable. This can be transformed to:
>> 
>> if ((base & 3) != 0) {
>> 
>> That check becomes loop independent and be optimized by loop predication (or I suppose loop unswitching but that wasn't the case of the micro benchmark I worked on).
>> 
>> This change also optimizes the pattern:
>> 
>> (offset << 2) & 3
>> 
>> to return 0.
>
> src/hotspot/share/opto/mulnode.cpp line 514:
> 
>> 512:   }
>> 513: 
>> 514:   return MulINode::Value(phase);
> 
> There is no `MulINode::value()` or `MulLNode::value()`, only based class `MulNode` have it.
> You missing initial check done by `MulNode::value()`.
> I suggest to move this code into `AndINode::mul_ring()` and `AndLNode::mul_ring()`.
> Also in `mul_ring()` you can pass `r1->get_con()` as `mask` value to `AndIL_shift_and_mask()`.

Thanks for reviewing this.
MulNode::AndIL_shift_and_mask() needs the shift node so it can test for shift->Opcode() == Op_ConvI2L. mult_ring() only gets the Type* as input. Are you suggesting to change mul_ring's signature?

> src/hotspot/share/opto/mulnode.cpp line 1719:
> 
>> 1717: bool MulNode::AndIL_shift_and_mask(PhaseGVN* phase, Node* mask, Node* shift, BasicType bt) const {
>> 1718:   if (mask == NULL || shift == NULL) {
>> 1719:     return false;
> 
> You need to check `shift` for `TOP`.

Code below:  

const TypeInteger* shift_t = phase->type(shift)->isa_integer(bt);
  if (mask_t == NULL || shift_t == NULL) {

catches the case where shift is top, I think.

-------------

PR: https://git.openjdk.java.net/jdk/pull/6697