tree 62fec55af0263e7c5fa165df7ba853c6065873b2
parent f8a98901b3b1093163fb1c7438b62bfcb3b27ee1
author David Gross <dgross@google.com> 1490820855 +0000
committer David Gross <dgross@google.com> 1490824317 -0700

Fix for crashes and failures due to 32-bit x86 struct layout.

(Revert "Revert "Fix for crashes and failures due to 32-bit x86 struct layout.""

This reverts commit 871eb011c1b5e9474ea021a4096824669b765b32.)

Add explicit padding to structure types, including invokable function
parameter structure types.  The padding does not change field offset
or structure size -- it makes explicit any padding that was implicit
due to the ABI.  This ensures that if the frontend compiles for an ABI
with stricter alignment requirements than the backend compiles for,
the frontend and backend will still agree on structure layout (field
offset and structure size).  This is important for 32-bit x86: The
frontend compiles for 32-bit ARM ABI, in which 64-bit scalars are
64-bit aligned; but the 32-bit x86 ABI says that 64-bit scalars are
only 32-bit aligned.

(Ideally, we would pad only exported structure types; but the most
convenient time to insert the padding is as soon as a structure type
definition is complete, so that we don't have to modify the AST to
update references to the structure's fields.  Unfortunately, this is
long before we can tell whether or not a structure type is exported.)

We had partial fixes for the 32-bit x86 problem in the backend (bcc),
but they were incomplete.  They compute field offsets according to the
ARM layout (thereby compatible with the frontend, including reflected
code and Allocation cell size) rather than the x86 layout; but:
- A stack-based local variable was allocated according to the
  (potentially smaller) x86 size rather than the ARM size, whereas
  field accesses occurred at ARM offsets, potentially spilling off
  the end of the local variable.
- Despite the old fixes, certain analyses/transformations (for
  example, certain loop optimizations) look at structure sizes and field
  offsets according to x86 rules rather than ARM rules

Also, for the benefit of libbcc change
  https://android-review.googlesource.com/#/c/358954/
make slang header files available as a module "slang_headers".

Bug: http://b/29154200
Bug: http://b/28070272

Test: (aosp_x86_64-eng emulator, full_fugu-eng, aosp_angler-eng) x
      (RsTest 32-bit, RsTest 64-bit, cts -m RenderscriptTest)
      tests/slang_test.py
      lit-tests/run-lit-tests.sh
  RsTest includes forthcoming additional regression tests:
    https://android-review.googlesource.com/#/c/299370/
  Tried (unmodified slang,   modified bcc) and
        (  modified slang, unmodified bcc) and
        (  modified slang,   modified bcc)
  By instrumenting modified bcc, confimed that:
  - Special x8632 layout transformations only run with unmodified slang,
    and only when test is compiled for x8632.
  "Modified bcc" is a forthcoming bcc change to turn off the "partial
  fixes" (x8632 layout transformations) mentioned above, and to verify
  that front end (Module) and back end (TargetMachine) agree on
  the layout of every exported struct type:
    https://android-review.googlesource.com/#/c/299531/

Change-Id: I25aac8e88812b5d3198e99f1929d4908ce663c46