RFR: 8280123: C2: Infinite loop in CMoveINode::Ideal during IGVN

[Tobias Hartmann]

On Tue, 18 Jan 2022 18:57:02 GMT, Vladimir Ivanov <vlivanov at openjdk.org> wrote:

> There's a discrepancy between `CMoveNode::Ideal()` and `CMoveINode::Ideal()` in the way how constants are detected:
> the former looks for `Con` nodes (`in(IfFalse)->is_Con()`) while the latter checks node types (e.g., `phase->type(in(IfFalse)) == TypeInt::ZERO`). It leads to infinite loop during IGVN (where  `CMoveNode::Ideal()` and `CMoveINode::Ideal()` repeatedly rotate `CMoveI` node) because `CMoveNode::Ideal()` don't consider `CastII` to a constant as a constant while `CMoveINode::Ideal()` does.
> 
> The fix migrates `CMoveNode::Ideal()` away from `Node::is_Con()` to node types.
> 
> Testing: hs-tier1 - hs-tier4

Looks good. I added some minor cleanup suggestions.

src/hotspot/share/opto/movenode.cpp line 87:

> 85:   assert(in(Condition) != this &&
> 86:          in(IfFalse)   != this &&
> 87:          in(IfTrue)    != this, "dead loop in CMoveNode::Ideal" );

Suggestion:

         in(IfTrue)    != this, "dead loop in CMoveNode::Ideal");

src/hotspot/share/opto/movenode.cpp line 97:

> 95:     if (in(Condition)->is_Bool()) {
> 96:       BoolNode* b  = in(Condition)->as_Bool();
> 97:       BoolNode* b2 = b->negate(phase);

Suggestion:

      BoolNode* b  = in(Condition)->as_Bool()->negate(phase);

src/hotspot/share/opto/movenode.cpp line 98:

> 96:       BoolNode* b  = in(Condition)->as_Bool();
> 97:       BoolNode* b2 = b->negate(phase);
> 98:       return make(in(Control), phase->transform(b2), in(IfTrue), in(IfFalse), _type);

Suggestion:

      return make(in(Control), phase->transform(b), in(IfTrue), in(IfFalse), _type);

test/hotspot/jtreg/compiler/c2/TestCMoveInfiniteGVN.java line 25:

> 23: 
> 24: /*
> 25:  * @test

Suggestion:

 * @test
 * @key stress randomness

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

Marked as reviewed by thartmann (Reviewer).

PR: https://git.openjdk.java.net/jdk/pull/7133