https://bugs.kde.org/show_bug.cgi?id=472329
--- Comment #5 from Paul Floyd <pjfl...@wanadoo.fr> ---
For easy reference this is the commit that I need to look at. In particular the
FIXMEs.
Author: Julian Seward <jsew...@acm.org> 2019-10-21 11:19:59
Committer: Julian Seward <jsew...@acm.org> 2020-01-02 06:42:21
Parent: 740381f8ac28e2878de7b068611385ed01985478 (Update following recent
bug-fix commits.)
Child: eaf8385b5ec1f6bf0f884b07d58ddb3f25ee5245 (Clean up machinery to do with
conditionalising IRStmts:)
Branches: debuginfo_load_option, freebsd, master, remotes/github/freebsd,
remotes/origin/VALGRIND_3_16_BRANCH, remotes/origin/master,
remotes/origin/users/mark/try-out-of-memory-reason,
remotes/origin/users/mark/try-realloc-again,
remotes/origin/users/mark/try-vgdb-multi,
remotes/origin/users/njn/try-cachegrind-cl-reqs,
remotes/origin/users/njn/try-rm-I
Follows: VALGRIND_3_15_0
Precedes: VALGRIND_3_16_0, VALGRIND_3_17_0
Initial implementation of C-source-level &&-idiom recovery
This branch contains code which avoids Memcheck false positives resulting
from
gcc and clang creating branches on uninitialised data. For example:
bool isClosed;
if (src.isRect(..., &isClosed, ...) && isClosed) {
clang9 -O2 compiles this as:
callq 7e7cdc0 <_ZNK6SkPath6isRectEP6SkRectPbPNS_9DirectionE>
cmpb $0x0,-0x60(%rbp) // "if (isClosed) { .."
je 7ed9e08 // "je after"
test %al,%al // "if (return value of call is nonzero) { .."
je 7ed9e08 // "je after"
..
after:
That is, the && has been evaluated right-to-left. This is a correct
transformation if the compiler can prove that the call to |isRect| returns
|false| along any path on which it does not write its out-parameter
|&isClosed|.
In general, for the lazy-semantics (L->R) C-source-level && operator, we
have
|A && B| == |B && A| if you can prove that |B| is |false| whenever A is
undefined. I assume that clang has some kind of interprocedural analysis
that
tells it that. The compiler is further obliged to show that |B| won't
trap,
since it is now being evaluated speculatively, but that's no big deal to
prove.
A similar result holds, per de Morgan, for transformations involving the C
language ||.
Memcheck correctly handles bitwise &&/|| in the presence of undefined
inputs.
It has done so since the beginning. However, it assumes that every
conditional branch in the program is important -- any branch on
uninitialised
data is an error. However, this idiom demonstrates otherwise. It defeats
Memcheck's existing &&/|| handling because the &&/|| is spread across two
basic blocks, rather than being bitwise.
This initial commit contains a complete initial implementation to fix that.
The basic idea is to detect the && condition spread across two blocks, and
transform it into a single block using bitwise &&. Then Memcheck's
existing
accurate instrumentation of bitwise && will correctly handle it. The
transformation is
<contents of basic block A>
C1 = ...
if (!C1) goto after
.. falls through to ..
<contents of basic block B>
C2 = ...
if (!C2) goto after
.. falls through to ..
after:
===>
<contents of basic block A>
C1 = ...
<contents of basic block B, conditional on C1>
C2 = ...
if (!C1 && !C2) goto after
.. falls through to ..
after:
This assumes that <contents of basic block B> can be conditionalised, at
the
IR level, so that the guest state is not modified if C1 is |false|. That's
not possible for all IRStmt kinds, but it is possible for a large enough
subset to make this transformation feasible.
There is no corresponding transformation that recovers an || condition,
because, per de Morgan, that merely corresponds to swapping the side exits
vs
fallthoughs, and inverting the sense of the tests, and the
pattern-recogniser
as implemented checks all possible combinations already.
The analysis and block-building is performed on the IR returned by the
architecture specific front ends. So they are almost not modified at all:
in
fact they are simplified because all logic related to chasing through
unconditional and conditional branches has been removed from them, redone
at
the IR level, and centralised.
The only file with big changes is the IRSB constructor logic,
guest_generic_bb_to_IR.c (a.k.a the "trace builder"). This is a complete
rewrite.
There is some additional work for the IR optimiser (ir_opt.c), since that
needs to do a quick initial simplification pass of the basic blocks, in
order
to reduce the number of different IR variants that the trace-builder has to
pattern match on. An important followup task is to further reduce this
cost.
There are two new IROps to support this: And1 and Or1, which both operate
on
Ity_I1. They are regarded as evaluating both arguments, consistent with
AndXX
and OrXX for all other sizes. It is possible to synthesise at the IR level
by
widening the value to Ity_I8 or above, doing bitwise And/Or, and
re-narrowing
it, but this gives inefficient code, so I chose to represent them directly.
The transformation appears to work for amd64-linux. In principle --
because
it operates entirely at the IR level -- it should work for all targets,
providing the initial pre-simplification pass can normalise the block ends
into the required form. That will no doubt require some tuning. And1 and
Or1
will have to be implemented in all instruction selectors, but that's easy
enough.
Remaining FIXMEs in the code:
* Rename `expr_is_speculatable` et al to `expr_is_conditionalisable`.
These
functions merely conditionalise code; the speculation has already been
done
by gcc/clang.
* `expr_is_speculatable`: properly check that Iex_Unop/Binop don't contain
operatins that might trap (Div, Rem, etc).
* `analyse_block_end`: recognise all block ends, and abort on ones that
can't
be recognised. Needed to ensure we don't miss any cases.
* maybe: guest_amd64_toIR.c: generate better code for And1/Or1
* ir_opt.c, do_iropt_BB: remove the initial flattening pass since presimp
will already have done it
* ir_opt.c, do_minimal_initial_iropt_BB (a.k.a. presimp). Make this as
cheap as possible. In particular, calling `cprop_BB_wrk` is total
overkill
since we only need copy propagation.
* ir_opt.c: once the above is done, remove boolean parameter for
`cprop_BB_wrk`.
* ir_opt.c: concatenate_irsbs: maybe de-dup w.r.t. maybe_unroll_loop_BB.
* remove option `guest_chase_cond` from VexControl (?). It was never used.
* convert option `guest_chase_thresh` from VexControl (?) into a Bool,
since
the revised code here only cares about the 0-vs-nonzero distinction now.
--
You are receiving this mail because:
You are watching all bug changes.
