| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
* Add 'dumpAction' hook to DynFlags.
It allows GHC API users to catch dumped intermediate codes and
information. The format of the dump (Core, Stg, raw text, etc.) is now
reported allowing easier automatic handling.
* Add 'traceAction' hook to DynFlags.
Some dumps go through the trace mechanism (for instance unfoldings that
have been considered for inlining). This is problematic because:
1) dumps aren't written into files even with -ddump-to-file on
2) dumps are written on stdout even with GHC API
3) in this specific case, dumping depends on unsafe globally stored
DynFlags which is bad for GHC API users
We introduce 'traceAction' hook which allows GHC API to catch those
traces and to avoid using globally stored DynFlags.
* Avoid dumping empty logs via dumpAction/traceAction (but still write
empty files to keep the existing behavior)
|
| | |
|
| | |
|
| | |
|
| | |
|
| | |
|
| |
|
|
|
|
|
|
|
|
|
|
| |
19 times out of 20 we already have dynflags in scope.
We could just always use `return dflags`. But this is in fact not free.
When looking at some STG code I noticed that we always allocate a
closure for this expression in the heap. Clearly a waste in these cases.
For the other cases we can either just modify the callsite to
get dynflags or use the _D variants of withTiming I added which
will use getDynFlags under the hood.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This is part two of fixing #17334.
There are two parts to this commit:
- A bugfix for computing loop levels
- A bugfix of basic block invariants in the NCG.
-----------------------------------------------------------
In the first bug we ended up with a CFG of the sort: [A -> B -> C]
This was represented via maps as fromList [(A,B),(B,C)] and later
transformed into a adjacency array. However the transformation did
not include block C in the array (since we only looked at the keys of
the map).
This was still fine until we tried to look up successors for C and tried
to read outside of the array bounds when accessing C.
In order to prevent this in the future I refactored to code to include
all nodes as keys in the map representation. And make this a invariant
which is checked in a few places.
Overall I expect this to make the code more robust as now any failed
lookup will represent an error, versus failed lookups sometimes being
expected and sometimes not.
In terms of performance this makes some things cheaper (getting a list
of all nodes) and others more expensive (adding a new edge). Overall
this adds up to no noteable performance difference.
-----------------------------------------------------------
Part 2: When the NCG generated a new basic block, it did
not always insert a NEWBLOCK meta instruction in the stream which
caused a quite subtle bug.
During instruction selection a statement `s`
in a block B with control of the sort: B -> C
will sometimes result in control
flow of the sort:
┌ < ┐
v ^
B -> B1 ┴ -> C
as is the case for some atomic operations.
Now to keep the CFG in sync when introducing B1 we clearly
want to insert it between B and C. However there is
a catch when we have to deal with self loops.
We might start with code and a CFG of these forms:
loop:
stmt1 ┌ < ┐
.... v ^
stmtX loop ┘
stmtY
....
goto loop:
Now we introduce B1:
┌ ─ ─ ─ ─ ─┐
loop: │ ┌ < ┐ │
instrs v │ │ ^
.... loop ┴ B1 ┴ ┘
instrsFromX
stmtY
goto loop:
This is simple, all outgoing edges from loop now simply
start from B1 instead and the code generator knows which
new edges it introduced for the self loop of B1.
Disaster strikes if the statement Y follows the same pattern.
If we apply the same rule that all outgoing edges change then
we end up with:
loop ─> B1 ─> B2 ┬─┐
│ │ └─<┤ │
│ └───<───┘ │
└───────<────────┘
This is problematic. The edge B1->B1 is modified as expected.
However the modification is wrong!
The assembly in this case looked like this:
_loop:
<instrs>
_B1:
...
cmpxchgq ...
jne _B1
<instrs>
<end _B1>
_B2:
...
cmpxchgq ...
jne _B2
<instrs>
jmp loop
There is no edge _B2 -> _B1 here. It's still a self loop onto _B1.
The problem here is that really B1 should be two basic blocks.
Otherwise we have control flow in the *middle* of a basic block.
A contradiction!
So to account for this we add yet another basic block marker:
_B:
<instrs>
_B1:
...
cmpxchgq ...
jne _B1
jmp _B1'
_B1':
<instrs>
<end _B1>
_B2:
...
Now when inserting B2 we will only look at the outgoing edges of B1' and
everything will work out nicely.
You might also wonder why we don't insert jumps at the end of _B1'. There is
no way another block ends up jumping to the labels _B1 or _B2 since they are
essentially invisible to other blocks. View them as control flow labels local
to the basic block if you'd like.
Not doing this ultimately caused (part 2 of) #17334.
|
| |
|
|
|
|
| |
This patch adds support for the s390x architecture for the LLVM code
generator. The patch includes a register mapping of STG registers onto
s390x machine registers which enables a registerised build.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
For backends maintaining the CFG during codegen
we can now find loops and their nesting level.
This is based on the Cmm CFG and dominator analysis.
As a result we can estimate edge frequencies a lot better
for methods, resulting in far better code layout.
Speedup on nofib: ~1.5%
Increase in compile times: ~1.9%
To make this feasible this commit adds:
* Dominator analysis based on the Lengauer-Tarjan Algorithm.
* An algorithm estimating global edge frequences from branch
probabilities - In CFG.hs
A few static branch prediction heuristics:
* Expect to take the backedge in loops.
* Expect to take the branch NOT exiting a loop.
* Expect integer vs constant comparisons to be false.
We also treat heap/stack checks special for branch prediction
to avoid them being treated as loops.
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
Statements can change the basic block in which instructions
are placed during instruction selection.
We have to keep track of this switch of the current basic block
as we need this information in order to properly update the CFG.
This commit implements this change and fixes #17334.
We do so by having stmtToInstr return the new block id
if a statement changed the basic block.
|
| |
|
|
|
|
|
|
| |
- Remove unneeded ones
- Use <..> for inter-package.
Besides general clean up, helps distinguish between the RTS we link
against vs the RTS we compile for.
|
| | |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Introduces a new flag `-fmax-pmcheck-deltas` to achieve that. Deprecates
the old `-fmax-pmcheck-iter` mechanism in favor of this new flag.
From the user's guide:
Pattern match checking can be exponential in some cases. This limit makes sure
we scale polynomially in the number of patterns, by forgetting refined
information gained from a partially successful match. For example, when
matching `x` against `Just 4`, we split each incoming matching model into two
sub-models: One where `x` is not `Nothing` and one where `x` is `Just y` but
`y` is not `4`. When the number of incoming models exceeds the limit, we
continue checking the next clause with the original, unrefined model.
This also retires the incredibly hard to understand "maximum number of
refinements" mechanism, because the current mechanism is more general
and should catch the same exponential cases like PrelRules at the same
time.
-------------------------
Metric Decrease:
T11822
-------------------------
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
'withTiming' becomes a function that, when passed '-vN' (N >= 2) or
'-ddump-timings', will print timing (and possibly allocations) related
information. When additionally built with '-eventlog' and executed with
'+RTS -l', 'withTiming' will also emit both 'traceMarker' and 'traceEvent'
events to the eventlog.
'withTimingSilent' on the other hand will never print any timing information,
under any circumstance, and will only emit 'traceEvent' events to the eventlog.
As pointed out in !1672, 'traceMarker' is better suited for things that we
might want to visualize in tools like eventlog2html, while 'traceEvent'
is better suited for internal events that occur a lot more often and that we
don't necessarily want to visualize.
This addresses #17138 by using 'withTimingSilent' for all the codegen bits
that are expressed as a bunch of small computations over streams of codegen
ASTs.
|
| | |
|
| |
|
|
|
|
| |
Add StgToCmm module hierarchy. Platform modules that are used in several
other places (NCG, LLVM codegen, Cmm transformations) are put into
GHC.Platform.
|
| |
|
|
|
| |
The tightens up the kinds a bit. I use type synnonyms to avoid adding
promotion ticks everywhere.
|
| |
|
|
|
|
|
| |
- Fixes crazy indentation in -ddump-debug output
- We no longer dump empty sections in -ddump-debug when a code block
does not have any generated debug info
- Minor refactoring in Debug.hs and AsmCodeGen.hs
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This generalizes code generators (outputAsm, outputLlvm, outputC, and
the call site codeOutput) so that they'll return the return values of
the passed Cmm streams.
This allows accumulating data during Cmm generation and returning it to
the call site in HscMain.
Previously the Cmm streams were assumed to return (), so the code
generators returned () as well.
This change is required by !1304 and !1530.
Skipping CI as this was tested before and I only updated the commit
message.
[skip ci]
|
| | |
|
| | |
|
| |
|
|
|
|
|
|
|
|
|
| |
These kinds of imports are necessary in some cases such as
importing instances of typeclasses or intentionally creating
dependencies in the build system, but '-Wunused-imports' can't
detect when they are no longer needed. This commit removes the
unused ones currently in the code base (not including test files
or submodules), with the hope that doing so may increase
parallelism in the build system by removing unnecessary
dependencies.
|
| |
|
|
|
|
|
|
| |
We introduce a PlatformWordSize type and use it in platformWordSize
field.
This removes to panic/error calls called when platform word size is not
32 or 64. We now check for this when reading the platform config.
|
| | |
|
| |
|
|
|
|
|
| |
Unfortunately this will require more work; register allocation is
quite broken.
This reverts commit acd795583625401c5554f8e04ec7efca18814011.
|
| |
|
|
|
|
|
| |
This adds support for constructing vector types from Float#, Double# etc
and performing arithmetic operations on them
Cleaned-Up-By: Ben Gamari <ben@well-typed.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Here the following changes are introduced:
- A read barrier machine op is added to Cmm.
- The order in which a closure's fields are read and written is changed.
- Memory barriers are added to RTS code to ensure correctness on
out-or-order machines with weak memory ordering.
Cmm has a new CallishMachOp called MO_ReadBarrier. On weak memory machines, this
is lowered to an instruction that ensures memory reads that occur after said
instruction in program order are not performed before reads coming before said
instruction in program order. On machines with strong memory ordering properties
(e.g. X86, SPARC in TSO mode) no such instruction is necessary, so
MO_ReadBarrier is simply erased. However, such an instruction is necessary on
weakly ordered machines, e.g. ARM and PowerPC.
Weam memory ordering has consequences for how closures are observed and mutated.
For example, consider a closure that needs to be updated to an indirection. In
order for the indirection to be safe for concurrent observers to enter, said
observers must read the indirection's info table before they read the
indirectee. Furthermore, the entering observer makes assumptions about the
closure based on its info table contents, e.g. an INFO_TYPE of IND imples the
closure has an indirectee pointer that is safe to follow.
When a closure is updated with an indirection, both its info table and its
indirectee must be written. With weak memory ordering, these two writes can be
arbitrarily reordered, and perhaps even interleaved with other threads' reads
and writes (in the absence of memory barrier instructions). Consider this
example of a bad reordering:
- An updater writes to a closure's info table (INFO_TYPE is now IND).
- A concurrent observer branches upon reading the closure's INFO_TYPE as IND.
- A concurrent observer reads the closure's indirectee and enters it. (!!!)
- An updater writes the closure's indirectee.
Here the update to the indirectee comes too late and the concurrent observer has
jumped off into the abyss. Speculative execution can also cause us issues,
consider:
- An observer is about to case on a value in closure's info table.
- The observer speculatively reads one or more of closure's fields.
- An updater writes to closure's info table.
- The observer takes a branch based on the new info table value, but with the
old closure fields!
- The updater writes to the closure's other fields, but its too late.
Because of these effects, reads and writes to a closure's info table must be
ordered carefully with respect to reads and writes to the closure's other
fields, and memory barriers must be placed to ensure that reads and writes occur
in program order. Specifically, updates to a closure must follow the following
pattern:
- Update the closure's (non-info table) fields.
- Write barrier.
- Update the closure's info table.
Observing a closure's fields must follow the following pattern:
- Read the closure's info pointer.
- Read barrier.
- Read the closure's (non-info table) fields.
This patch updates RTS code to obey this pattern. This should fix long-standing
SMP bugs on ARM (specifically newer aarch64 microarchitectures supporting
out-of-order execution) and PowerPC. This fixes issue #15449.
Co-Authored-By: Ben Gamari <ben@well-typed.com>
|
| |
|
|
|
|
|
| |
ghc-pkg needs to be aware of platforms so it can figure out which
subdire within the user package db to use. This is admittedly
roundabout, but maybe Cabal could use the same notion of a platform as
GHC to good affect too.
|
| | |
|
| |
|
|
|
| |
Previously log and exp were primitives yet log1p and expm1 were FFI
calls. Fix this non-uniformity.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
When ghc was built for powerpc32 built failed as:
It's a fallout of commit 3f46cffcc2850e68405a1
("PPC NCG: Refactor stack allocation code") where
word size used to be
II32/II64
and changed to
II8/panic "no width for given number of bytes"
widthFromBytes ((platformWordSize platform) `quot` 8)
The change restores initial behaviour by removing extra division.
Signed-off-by: Sergei Trofimovich <slyfox@gentoo.org>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
On powerpc32 64-bit comparison code generated dangling
target labels. This caused ghc build failure as:
$ ./configure --target=powerpc-unknown-linux-gnu && make
...
SCCs aren't in reverse dependent order
bad blockId n3U
This happened because condIntCode' in PPC codegen generated
label name but did not place the label into `cmp_lo` code block.
The change adds the `cmp_lo` label into the case of negative
comparison.
Signed-off-by: Sergei Trofimovich <slyfox@gentoo.org>
|
| |
|
|
| |
See #13101 and #15454
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
1. If GHC is to be multi-target, these cannot be baked in at compile
time.
2. Compile-time flags have a higher maintenance than run-time flags.
3. The old way makes build system implementation (various bootstrapping
details) with the thing being built. E.g. GHC doesn't need to care
about which integer library *will* be used---this is purely a crutch
so the build system doesn't need to pass flags later when using that
library.
4. Experience with cross compilation in Nixpkgs has shown things work
nicer when compiler's can *optionally* delegate the bootstrapping the
package manager. The package manager knows the entire end-goal build
plan, and thus can make top-down decisions on bootstrapping. GHC can
just worry about GHC, not even core library like base and ghc-prim!
|
| |
|
|
|
|
|
|
|
|
|
|
| |
When a new closure identifier is being established to a
local or exported closure already emitted into the same
module, refrain from adding an IND_STATIC closure, and
instead emit an assembly-language alias.
Inter-module IND_STATIC objects still remain, and need to be
addressed by other measures.
Binary-size savings on nofib are around 0.1%.
|
| | |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
* simplifies registers to have GPR, Float and Double, by removing the SSE2 and X87 Constructors
* makes -msse2 assumed/default for x86 platforms, fixing a long standing nondeterminism in rounding
behavior in 32bit haskell code
* removes the 80bit floating point representation from the supported float sizes
* theres still 1 tiny bit of x87 support needed,
for handling float and double return values in FFI calls wrt the C ABI on x86_32,
but this one piece does not leak into the rest of NCG.
* Lots of code thats not been touched in a long time got deleted as a
consequence of all of this
all in all, this change paves the way towards a lot of future further
improvements in how GHC handles floating point computations, along with
making the native code gen more accessible to a larger pool of contributors.
|
| | |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Fixes #16052
When the offset in `setByteArray#` is statically known, we can provide
better alignment guarantees then just 1 byte.
Also, memset can now do 64-bit wide sets.
The current memset intrinsic is not optimal however and can be
improved for the case when we know that we deal with
(baseAddress at known alignment) + offset
For instance, on 64-bit
`setByteArray# s 1# 23# 0#`
given that bytearray is 8 bytes aligned could be unrolled into
`movb, movw, movl, movq, movq`; but currently it is
`movb x23` since alignment of 1 is all we can embed into MO_Memset op.
|
| |
|
|
|
|
| |
This commit includes the necessary changes in code and
documentation to support a primop that reverses a word's
bits. It also includes a test.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This moves all URL references to Trac Wiki to their corresponding
GitLab counterparts.
This substitution is classified as follows:
1. Automated substitution using sed with Ben's mapping rule [1]
Old: ghc.haskell.org/trac/ghc/wiki/XxxYyy...
New: gitlab.haskell.org/ghc/ghc/wikis/xxx-yyy...
2. Manual substitution for URLs containing `#` index
Old: ghc.haskell.org/trac/ghc/wiki/XxxYyy...#Zzz
New: gitlab.haskell.org/ghc/ghc/wikis/xxx-yyy...#zzz
3. Manual substitution for strings starting with `Commentary`
Old: Commentary/XxxYyy...
New: commentary/xxx-yyy...
See also !539
[1]: https://gitlab.haskell.org/bgamari/gitlab-migration/blob/master/wiki-mapping.json
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
We make liveness information for global registers
available on `JMP` and `BCTR`, which were the last instructions
missing. With complete liveness information we do not need to
reserve global registers in `freeReg` anymore. Moreover we
assign R9 and R10 to callee saves registers.
Cleanup by removing `Reg_Su`, which was unused, from `freeReg`
and removing unused register definitions.
The calculation of the number of floating point registers is too
conservative. Just follow X86 and specify the constants directly.
Overall on PowerPC this results in 0.3 % smaller code size in nofib
while runtime is slightly better in some tests.
|
| |
|
|
|
| |
This moves all URL references to Trac tickets to their corresponding
GitLab counterparts.
|
| |
|
|
|
| |
GHC native code generator generates .incbin and .file directives. We
need to escape those strings correctly on Windows (see #16389).
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The splitter is an evil Perl script that processes assembler code.
Its job can be done better by the linker's --gc-sections flag. GHC
passes this flag to the linker whenever -split-sections is passed on
the command line.
This is based on @DemiMarie's D2768.
Fixes Trac #11315
Fixes Trac #9832
Fixes Trac #8964
Fixes Trac #8685
Fixes Trac #8629
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
It never really encoded a invariant.
* The linear register allocator just did partial pattern matches
* The graph allocator just set it to (Just mapEmpty) for Nothing
So I changed LiveInfo to directly contain the map.
Further natCmmTopToLive which filled in Nothing is no longer exported.
Instead we know call cmmTopLiveness which changes the type AND fills
in the map.
|
| |
|
|
|
|
|
|
|
|
| |
This patch adds an optimization into the NCG: for large strings
(threshold configurable via -fbinary-blob-threshold=NNN flag), instead
of printing `.asciz "..."` in the generated ASM source, we print
`.incbin "tmpXXX.dat"` and we dump the contents of the string into a
temporary "tmpXXX.dat" file.
See the note for more details.
|
| | |
|
| | |
|
