RFR(S): 8213419: C2 may hang in MulLNode::Ideal()/MulINode::Ideal() with gcc 8.2.1

[Vladimir Kozlov]

Good.

Thanks,
Vladimir

On 11/6/18 6:16 AM, Roland Westrelin wrote:
> 
> http://cr.openjdk.java.net/~roland/8213419/webrev.00/
> 
> Multiplications by constant Integer.MIN_VALUE or Long.MIN_VALUE cause C2
> to hang in MulLNode::Ideal()/MulINode::Ideal() when hotspot is compiled
> with gcc 8.2.1.
> 
> The hang itself happens in inlined function:
> 
> // Returns largest i such that 2^i <= x.
> // If x < 0, the function returns 31 on a 32-bit machine and 63 on a 64-bit machine.
> // If x == 0, the function returns -1.
> inline int log2_intptr(intptr_t x) {
>    int i = -1;
>    uintptr_t p = 1;
>    while (p != 0 && p <= (uintptr_t)x) {
>      // p = 2^(i+1) && p <= x (i.e., 2^(i+1) <= x)
>      i++; p *= 2;
>    }
>    // p = 2^(i+1) && x < p (i.e., 2^i <= x < 2^(i+1))
>    // If p = 0, overflow has occurred and i = 31 or i = 63 (depending on the machine word size).
>    return i;
> }
> 
> Looking at the assembly code p != 0 is optimized out.
> 
>     0x00007f6494aae685 <+1157>: mov $0xffffffff,%ecx
>     0x00007f6494aae68a <+1162>: mov (%rax),%rdi
>     0x00007f6494aae68d <+1165>: mov $0x1,%eax
>     0x00007f6494aae692 <+1170>: nopw 0x0(%rax,%rax,1)
>     0x00007f6494aae698 <+1176>: add %rax,%rax
>     0x00007f6494aae69b <+1179>: add $0x1,%ecx
>     0x00007f6494aae69e <+1182>: cmp %r14,%rax
>     0x00007f6494aae6a1 <+1185>: jbe 0x7f6494aae698 <MulINode::Ideal(PhaseGVN*, bool)+1176>
> 
> That call is preceded by:
> 
> if( con < 0 ) {
>    con = -con;
> }
> 
> which is undefined behavior for INT_MIN so the compiler is free to
> assume con != INT_MIN and then optimizing the p != 0 test is legal.
> 
> The fix uses some utiliy functions that were added for aarch64.
> 
> Roland.
>