[Bug 3533] [IcedTea8] HotSpot generates code with unaligned stack, crashes on SSE operations

[bugzilla-daemon at icedtea.classpath.org]

https://icedtea.classpath.org/bugzilla/show_bug.cgi?id=3533

--- Comment #32 from Maciej S. Szmigiero <mail at maciej.szmigiero.name> ---
Does the compiler version you are using add the stack realigning instruction
when compiling the example from comment 23 with "-mstackrealign" only?
If yes, then this might explain the difference we are seeing.

> My testing shows a reliable crash without -mstackrealign but a full
> bootstrap with it.

In the Gentoo bug I wrote that preliminary fixes stopped crashes during
bootstrap but the resulting IcedTea was still pretty unstable
(see "issue 1" from comment 24 in that Gentoo bug).

> Comment 23 shows an isolated example.

Think of "external" from that example as some function in glibc (using SSE)
where "internal" is an entry point from the generated code, like
"SafepointSynchronize::handle_polling_page_exception".
I see no reason why with "-mstackrealign" only the compiler would guarantee
realigning stack in the IcedTea case but not in the example from comment 23.

We might be lucky that every such entry point function by coincidence compiles
to a code that itself relies on proper stack alignement and so gets stack
realignment instruction even with just "-mstackrealign", but that:
1) Would require auditing every such entry function with every supported
compiler version with every reasonable optimizer flag (like different "-march"
settings) - that the stack realignment instruction gets added in every such
case as some optimization flag might result in a code that does not rely on
stack alignement and so without a stack realignment being generated,
2) Would result in a very fragile code that might generate subtle crash points
from seemingly innocent code changes.

As I wrote in comment 23 adding both "-mstackrealign" and
"-mincoming-stack-boundary=2" works fine, too - this might be the right
solution here.

Unfortunately, I don't have an easy access to the setup that had crashing
IcedTea, so I can't test it easily, but due to the two points above a
successful test now might not result in lack of future problems.

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.openjdk.java.net/pipermail/distro-pkg-dev/attachments/20180527/d6ae9775/attachment.html>