Summary:
Show that the SSA restrictions imply that the blocks can be
topologically sorted so that any use of a variable follows its
definition in the list of blocks. This showed a slight flaw in my
definition of SSA form. It used to treat program counters up to
equality. However, PCs that point to a block header contain the location
that control came from (so that the Phi instructions can be executed),
but the SSA restrictions shouldn't pay attention to that, so now the
definition of SSA introduces a equivalence on PCs that ignores that
additional information.
Reviewed By: jberdine
Differential Revision: D21066873
fbshipit-source-id: 735575a1f
Summary:
The LLVM->LLAIR translation keeps a mapping of variables to expressions.
Previously, the invariants related to that mapping were kept in the
state relation, and so the proof needed show that they were preserved
along execution traces. This wasn't obvious as the state changes in
non-SSA ways during evaluation, but the correctness of the mappings is
heavily based on the program being in SSA form. This change separates
out the invariants, and the proof uses the final mapping that the
compiler builds, which contains all of the relevant bindings that might
be needed during execution.
Reviewed By: jberdine
Differential Revision: D20625109
fbshipit-source-id: d4c2dfe19
Summary:
Add a field to LLAIR variables to indicate whether they are global or
local. Update the LLVM semantics for constant expression evaluation to
be relational, so that it doesn't have to have an answer for references
to undefined globals.
Reviewed By: jberdine
Differential Revision: D19446312
fbshipit-source-id: 9bbfd180e
Summary:
This commit adds truncation, sign extension and zero extension to LLVM
and the Convert instruction to LLAIR.
The LLVM instructions use HOL's build-in word/int and word/num
conversions. Sanity-checking theorems prove that zero-extending leaves
the value of the word unchanged when considered as an unsigned value,
and that sign-extending leaves the value unchanged when considered as a
signed value.
The llair semantics for Convert uses the truncate_2comp function which
converts an integer to another integer as though they were represented
in 2's complement. e.g. truncate_2comp 255 16 = 255, truncate_2comp
255 8 = -1, truncate_2comp -3 2 = 1
Reviewed By: jberdine
Differential Revision: D18058833
fbshipit-source-id: df9de480c
Summary:
The old syntactic invariant in prog_ok was in the wrong direction,
saying that all labels in a phi instruction have to exist, rather than
saying that when we jump to a new block, the label of the block we came
from must be in all of the phi instructions.
Reviewed By: jberdine
Differential Revision: D18058832
fbshipit-source-id: d2ad33b04
Summary:
This required some minor tweaks to how the semantics encode values into
and out of byte lists. The remaining problems have to do with how LLVM
globals are translated into llair. At the moment, llair semantic's state
keeps a mapping for globals to their addresses, following the LLVM
semantics. However, it is not used because the translation (following
the code in frontend.ml) translates LLVM globals into llair locals,
which the llair semantics isn't set up to handle.
Reviewed By: jberdine
Differential Revision: D17930787
fbshipit-source-id: 06c6368e0
Summary:
Previously, the LLVM semantics could be stuck where the LLAIR semantics
was not yet stuck, but would become stuck (at the same place) after
taking a step. This was due to LLVM using the traditional definition of
stuck states: any state from which there are no transitions. However,
LLAIR cannot do that because it might get stuck in the middle of a block
that contains several visible stores. We don't want to consider the
whole block stuck, nor can we finish it. Thus, the LLAIR definition of
stuckness is when the state has the stuck flag set which happens when
stopping in the middle of a block after encountering a stuck
instruction. Now LLVM takes the same approach.
Reviewed By: jberdine
Differential Revision: D17855085
fbshipit-source-id: a094d25d5
Summary:
Add an argument to the Exit instruction. Update the LLVM semantics to
execute the Exit instruction and store the result in an "exited"
component of the state. (Previously it just noticed that it was stuck
about to do an Exit.)
With exiting treated uniformly, now in the proof that for every LLVM
trace, there is a llair trace that simulates it, all of the cheats
except for 1 are just cases that I haven't got to yet. However, the last
cheat is for the situation where the LLVM program gets stuck and the
llair program doesn't. For example, the following two line LLVM program
gets stuck because r2 is not assigned (ignoring for the moment the static
restriction that LLVM is in SSA form).
r1 := r2
Exit(0)
The compilation to llair omits the assignment and so we get a llair
program that doesn't get stuck:
Exit(0)
The key question is whether the static restrictions are sufficient to
ensure that no expression that might be omitted can get stuck.
Reviewed By: jberdine
Differential Revision: D17737589
fbshipit-source-id: bc6c01a1b
Summary:
If the LLVM to llair translation keeps a mapping from register r to
expression e, then for each register r' mentioned in e, there must be an
assignment to r' that dominates the entire live range of r. Thus, where
ever r might be replaced by e, the value of r' will be the same as it
was when the initial assignment to r occurred. Maintaining this
invariant relies on the LLVM being in SSA form.
Reviewed By: jberdine
Differential Revision: D17710288
fbshipit-source-id: fd3eaa57d
Summary:
The LLVM semantics and translation was not consistently treating the
1-bit word value condition as signed or unsigned.
Reviewed By: jberdine
Differential Revision: D17605766
fbshipit-source-id: 77edf63b7
Summary:
Previously the LLVM semantics did the phi instructions at the head of a
block as part of executing the branch into that block. This looked a bit
weird, but had the advantage that the semantics knew which block was
being jumped from, which is necessary to run the phi instructions.
However, it meant that the rules for doing phi instructions would need
to show up with each branching construct. It was also annoying for the
LLVM->llair proof, since the phis are removed and their effect happens as
a distinct step from the branch.
Here we add a distinct Phi_ip instruction pointer to indicate that the
phi instructions at the start of the block should execute next, and then
be incremented to the usual numeric instruction pointer that points to
the non-phi instructions. The Phi_ip contains the identity of the
previous block.
Reviewed By: jberdine
Differential Revision: D17452416
fbshipit-source-id: 78fef7cca
Summary:
Give the llair semantics observable side effects (writes to global
variables) and a semantic function mirroring the LLVM semantics. Start
sketching out the LLVM/llair translation equivalence proof in a top-down
way from the obvious statement of equality of the semantics.
Reviewed By: jberdine
Differential Revision: D17399654
fbshipit-source-id: 2170678a8
Summary:
The simple LLVM semantics steps one instruction at a time, but the
generated llair does whole blocks at a time, since many individual LLVM
instructions can become a single llair expression. We add a bigger-step
LLVM semantics that does whole blocks at a time (except that it also
stops at function calls, since those end blocks in llair). The steps in
this bigger-step semantics should be at the same granularity as the
llair steps, making it easier to verify the translation.
We add a notion of observation to the LLVM semantics (right now, just
global variable writes) and use that to define two top-level semantic
functions, which we prove to be equivalent.
Reviewed By: jberdine
Differential Revision: D17396016
fbshipit-source-id: ee632fb92
Summary:
This includes a few changes and corrections to the semantics, to support
the translation. This initial attempt to reason about LLVM -> llair
showed three things that needed repair in the semantics, in addition to
various bugs. We address them as follows.
Refactor llair semantics to have only a single kind of flat value:
integers that fit into specified bit widths. Operations on size values
(e.g., offsets, indices and the like) can just take an integer and
ignore its number of bits. Pointers can just be considered integers that
fit into a certain size given by the constant pointer_size. Later on we
can consider making this a parameter to the model.
Change the generic memory model interface to use numbers rather than
words as the generic encoding of a large value. This makes it more
useful for llair where words are not used.
Pay more careful attention to signed/unsigned issues. Neither LLVM nor
llair have a concept of signed vs unsigned value. Instead individual
operations interpret bit patterns in various ways, some of which are
ambiguous in the LLVM manual. For example, since getelementpointer's
indices are explicitly said to be interpreted as signed 2's complement,
we should probably do the same for insertvalue and extractvalue. However
it is not clear how the argument to alloca is to be interpreted. For now
we assume signed.
Reviewed By: jberdine
Differential Revision: D17164133
fbshipit-source-id: 31a8af635
Summary:
Not everything is here yet, and there is some confusion on what to do
about the size values. However, the semantics has the right general
shape and will be a nice starting point for thinking about the details.
Reviewed By: jberdine
Differential Revision: D17111041
fbshipit-source-id: cc75651c6
Summary:
LLVM and llair have similar memory models, and we don't want to
duplicate any definitions or theorems. This adds a new memory model
theory which should be understandable in its own right. A heap is a
mapping from addresses to bytes, alongside a set of valid addresses, and
intervals that have been allocated already. Primitives are defined for
allocating and de-allocating as well as reading and writing chuncks of
bytes.
There is also a generic type of structured values, and functions for
converting them to/from byte arrays.
Reviewed By: jberdine
Differential Revision: D17074470
fbshipit-source-id: bdab6089f
Summary:
Each variable now contains its type, alongside its name. This is more
uniform than in LLVM, where the name is usually paired with a type, but
not always, for example, the register type of the result of an
extractvalue is left implicit.
Reviewed By: jberdine
Differential Revision: D16984630
fbshipit-source-id: 1c3bc4985
Summary:
HOL now lets us omit quotations on Datatypes and make them look more
like the other new-style HOL definitions.
Reviewed By: jberdine
Differential Revision: D16983934
fbshipit-source-id: f8ef3abb5
Summary:
This sketches out how translation can be approached. It is partially
based on the Sledge code.
For basic blocks, isn't based on the Sledge code, but just my own
thoughts as a starting point. Essentially, we are trying to build up
larger expressions, and so not assigning to temporary registers that
don't live past the end of the block. This does remove sharing, so a
fancier approach could check for multiple uses of end-of-block dead
registers, or look at the sizes of expressions. The approach should be
flexible enough to accommodate such changes.
Fix icmp syntax
Using finite maps is elegant in the semantics, but awkward for writing
the translation function. Refactor the mappings from labels to functions
and from labels to blocks to use association lists instead.
To remove phi nodes, the translation takes every edge in the control
flow graph and makes a new basic block that contains a single parallel
move instruction that corresponds to the action of the phi node of the
target block.
Reviewed By: jberdine
Differential Revision: D16831051
fbshipit-source-id: 005663e26
Summary:
The AST is not complete on expressions, but it should have most of the
important features.
The representation is in some ways very different from the OCaml
implementation, because the OCaml code uses mutation to build the CFG as
an actual pointer graph in memory, and also because the expression
representation is optimised for the backend. For the former, it should
be easy to see that the AST here is isomorphic, representing the CFG
with finite maps from block labels. The correspondence is less clear in
the latter case, but the point here is not to model or verify
implementation optimisations, but to give a semantics to llair as a
language.
Reviewed By: jberdine
Differential Revision: D16807132
fbshipit-source-id: b0f64b3ec
Summary:
Change loc_var (for local variable) to reg (for register) because
loc_var looks too much like a location tagged variable.
Reviewed By: jberdine
Differential Revision: D16827920
fbshipit-source-id: 5b11f1065
Summary:
There could very well still be bugs in the semantics, since the
invariant here doesn't say all that much, and it completely ignores
local registers. But most trivial things and typos are probably fixed.
Reviewed By: jberdine
Differential Revision: D16803281
fbshipit-source-id: 48ba2523b
Summary:
Made progress on the sanity checking lemma (that the step relation
preserves some simple invariants on the state). Proved the Ret
instruction case of the state invariant lemma. To do this, I fixed a few
bugs in the definition, and strengthened the invariants.
Reviewed By: jberdine
Differential Revision: D16786900
fbshipit-source-id: 6fa8cb170
Summary:
Global variables need allocating and initialising before the machine can
start. The definition here shouldn't constrain how and where they are
allocated. For example, they don't all need to have separate
allocations. We also tag allocated blocks so that the allocation for a
global can never be deallocated.
Start working on a sanity checking invariant on states.
Reviewed By: jberdine
Differential Revision: D16735068
fbshipit-source-id: 0d5e60e7a
Summary:
Start working on a simple model of LLVM with the ultimate goal of
handling relevant and/or tricky aspects of LLVM and LLAIR and then
formalising the translation from LLVM to LLAIR.
This is a complete initial model of everything that we are interested in
except for exceptions, which should be tricky. Also no thought has gone
into the treatment of poison and the undefined value, so the treatment
is naive, which is at least partially justified because we are
interested in the semantics of LLVM IR after the optimisation passes
have run.
Include some sanity checking theorems.
Reviewed By: jberdine
Differential Revision: D16731885
fbshipit-source-id: fd53949fe
Scott Owens