The compiler was failing to freeze the source site containing the
value of the second word in a multi-word move, leading to mis-compiled
code in some cases.
The first problem was that, on x86, we failed to properly keep track
of whether to expect the return address to be on the stack or not when
unwinding through a frame. We were relying on a "stackLimit" pointer
to tell us whether we were looking at the most recently-called frame
by comparing it with the stack pointer for that frame. That was
inaccurate in the case of a thread executing at the beginning of a
method before a new frame is allocated, in which case the most recent
two frames share a stack pointer, confusing the unwinder. The
solution involves keeping track of how many frames we've looked at
while walking the stack.
The other problem was that compareIpToMethodBounds assumed every
method was followed by at least one byte of padding before the next
method started. That assumption was usually valid because we were
storing the size following method code prior to the code itself.
However, the last method of an AOT-compiled code image is not followed
by any such method header and may instead be followed directly by
native code with no intervening padding. In that case, we risk
interpreting that native code as part of the preceding method, with
potentially bizarre results.
The reason for the compareIpToMethodBounds assumption was that methods
which throw exceptions as their last instruction generate a
non-returning call, which nonetheless push a return address on the
stack which points past the end of the method, and the unwinder needs
to know that return address belongs to that method. A better solution
is to add an extra trap instruction to the end of such methods, which
is what this patch does.
This monster commit is the first step towards supporting
cross-architecture bootimage builds. The challenge is to build a heap
and code image for the target platform where the word size and
endianess may differ from those of the build architecture. That means
the memory layout of objects may differ due to alignment and size
differences, so we can't just copy objects into the heap image
unchanged; we must copy field by field, resizing values, reversing
endianess and shifting offsets as necessary.
This commit also removes POD (plain old data) type support from the
type generator because it added a lot of complication and little
value.
We can't reduce a conditional branch to an unconditional jump unless
both arguments to the comparison are constants *and* those constants
have been resolved. The latter may not be true in the case of a
bootimage build.
Code including subroutines and conditionals can result in frame and
register resources being held by values which aren't in scope when
resetFrame is called, so we need to clean them up after cleaning the
in-scope values.
Due to encoding limitations, the immediate operand of conditional
branches can be no more than 32KB forward or backward. Since the
JIT-compiled form of some methods can be larger than 32KB, and we also
do conditional jumps to code outside the current method in some cases,
we must work around this limitation.
The strategy of this commit is to provide inline, intermediate jump
tables where necessary. A given conditional branch whose target is
too far for a direct jump will instead point to an unconditional
branch in the nearest jump table which points to the actual target.
Unconditional immediate branches are also limited on PowerPC, but this
limit is 32MB, which is not an impediment in practice. If it does
become a problem, we'll need to encode such branches using multiple
instructions.
Previously, we unwound the stack by following the chain of frame
pointers for normal returns, stack trace creation, and exception
unwinding. On x86, this required reserving EBP/RBP for frame pointer
duties, making it unavailable for general computation and requiring
that it be explicitly saved and restored on entry and exit,
respectively.
On PowerPC, we use an ABI that makes the stack pointer double as a
frame pointer, so it doesn't cost us anything. We've been using the
same convention on ARM, but it doesn't match the native calling
convention, which makes it unusable when we want to call native code
from Java and pass arguments on the stack.
So far, the ARM calling convention mismatch hasn't been an issue
because we've never passed more arguments from Java to native code
than would fit in registers. However, we must now pass an extra
argument (the thread pointer) to e.g. divideLong so it can throw an
exception on divide by zero, which means the last argument must be
passed on the stack. This will clobber the linkage area we've been
using to hold the frame pointer, so we need to stop using it.
One solution would be to use the same convention on ARM as we do on
x86, but this would introduce the same overhead of making a register
unavailable for general use and extra code at method entry and exit.
Instead, this commit removes the need for a frame pointer. Unwinding
involves consulting a map of instruction offsets to frame sizes which
is generated at compile time. This is necessary because stack trace
creation can happen at any time due to Thread.getStackTrace being
called by another thread, and the frame size varies during the
execution of a method.
So far, only x86(_64) is working, and continuations and tail call
optimization are probably broken. More to come.
We now check for stack overflow in the JIT build as well as the
interpreted build, throwing a StackOverflowError if the limit
(currently hard-coded to 64KB, but should be easy to make
configurable) is exceeded.
We have to be careful about how we calculate return addresses on ARM
due to padding introduced by constant pools interspersed with code.
When calculating the offset of code where we're inserting a constant
pool, we want the offset of the end of the pool for jump targets, but
we want the offset just prior to the beginning of the pool (i.e. the
offset of the instruction responsible for jumping past the pool) when
calculating a return address.
Compiling the entire OpenJDK class library into a bootimage revealed
some corner cases which broke the compiler, including synchronization
in a finally block and gotos targeting the first instruction of an
unsynchronized method.
Previously, loading an arbitrary 32-bit constant required up to four
instructions (128 bytes), since we did so one byte at a time via
immediate-mode operations.
The preferred way to load constants on ARM is via PC-relative
addressing, but this is challenging because immediate memory offsets
are limited to 4096 bytes in either direction. We frequently need to
compile methods which are larger than 4096, or even 8192, bytes, so we
must intersperse code and data if we want to use PC-relative loads
everywhere.
This commit enables pervasive PC-relative loads by handling the
following cases:
1. Method is shorter than 4096 bytes: append data table to end
2. Method is longer than 4096 bytes, but no basic block is longer
than 4096 bytes: insert data tables as necessary after blocks, taking
care to minimize the total number of tables
3. Method is longer than 4096 bytes, and some blocks are longer than
4096 bytes: split large basic blocks and insert data tables as above
Previously, we only visited frame locations containing values, but
this invited the possibility of reusing the same site for two
locations in some cases.
