RFR: 8331575: C2: crash when ConvL2I is split thru phi at LongCountedLoop

[Christian Hagedorn]

On Tue, 7 May 2024 16:47:47 GMT, Emanuel Peter <epeter at openjdk.org> wrote:

>> In the test case:
>> 
>> 
>> long i;
>> for (; i > 0; i--) {
>>     res += 42 / ((int) i);
>> 
>> 
>> The long counted loop phi has type `[1..100]`. As a consequence, the
>> `ConvL2I` also has type `[1..100]`. The `DivI` node that follows can't
>> fault: it is not guarded by a zero check and has no control set.
>> 
>> The `ConvL2I` is split through phi and so is the `DiVI` node:
>> `PhaseIdealLoop::cannot_split_division()` returns true because the
>> value coming from the backedge into the `DivI` (when it is about to be
>> split thru phi) is the result of the `ConvL2I` which has type
>> `[1..100`] so is not zero as far as the compiler can tell.
>> 
>> On the last iteration of the loop, i is 1. Because the DivI was split
>> thru Phi, it computes the value for the following iteration, so for i
>> = 0. This causes a crash when the compiled code runs.
>> 
>> The same problem can't happen with an int counted loop because logic
>> in `PhaseIdealLoop::split_thru_phi()` prevents a `ConvI2L` from being
>> split thru phi. I propose to fix this the same way: in the test case,
>> it's not true that once the `ConvL2I` is split thru phi it keeps type
>> `[1..100]`. The fix is fairly conservative because it's base on the
>> existing logic for `ConvI2L`: we would want to not split a `ConvL2I`
>> only a counted loopd but. I suppose the same is true for the `ConvI2L`
>> and I thought it would be best to revisit both together.
>
> test/hotspot/jtreg/compiler/splitif/TestLongCountedLoopConvL2I.java line 31:
> 
>> 29:  *                   -XX:+StressGCM -XX:StressSeed=92643864 TestLongCountedLoopConvL2I
>> 30:  * @run main/othervm -XX:-BackgroundCompilation -XX:-TieredCompilation -XX:-UseOnStackReplacement
>> 31:  *                   -XX:+StressGCM TestLongCountedLoopConvL2I
> 
> Would it make sense to have a run that allows OSR?

You should also add `-XX:+UnlockDiagnosticVMOptions` for the stress flag.

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

PR Review Comment: https://git.openjdk.org/jdk/pull/19086#discussion_r1593501246