-
ac9fb269
by Simon Peyton Jones at 2025-05-20T09:19:04-04:00
Track rewriter sets more accurately in constraint solving
This MR addresses #26003, by refactoring the arcane
intricacies of Note [Equalities with incompatible kinds].
NB: now retitled to
Note [Equalities with heterogeneous kinds].
and the main Note for this MR.
In particular:
* Abandon invariant (COERCION-HOLE) in Note [Unification preconditions] in
GHC.Tc.Utils.Unify.
* Abandon invariant (TyEq:CH)) in Note [Canonical equalities] in
GHC.Tc.Types.Constraint.
* Instead: add invariant (REWRITERS) to Note [Unification preconditions]:
unify only if the constraint has an empty rewriter set.
Implementation:
* In canEqCanLHSFinish_try_unification, skip trying unification if there is a
non-empty rewriter set.
* To do this, make sure the rewriter set is zonked; do so in selectNextWorkItem,
which also deals with prioritisation.
* When a coercion hole is filled, kick out inert equalities that have that hole
as a rewriter. It might now be unlocked and available to unify.
* Remove the ad-hoc `ch_hetero_kind` field of `CoercionHole`.
* In `selectNextWorkItem`, priorities equalities withan empty rewriter set.
* Defaulting: see (DE6) in Note [Defaulting equalities]
and Note [Limited defaulting in the ambiguity check]
* Concreteness checks: there is some extra faff to try to get decent
error messages when the FRR (representation-polymorphism) checks
fail. In partiular, add a "When unifying..." explanation when the
representation-polymorphism check arose from another constraint.
-
86406f48
by Cheng Shao at 2025-05-20T09:19:47-04:00
rts: fix rts_clearMemory logic when sanity checks are enabled
This commit fixes an RTS assertion failure when invoking
rts_clearMemory with +RTS -DS. -DS implies -DZ which asserts that free
blocks contain 0xaa as the designated garbage value. Also adds the
sanity way to rts_clearMemory test to prevent future regression.
Closes #26011.
ChatGPT Codex automatically diagnosed the issue and proposed the
initial patch in a single shot, given a GHC checkout and the following
prompt:
---
Someone is reporting the following error when attempting to use `rts_clearMemory` with the RTS option `-DS`:
```
test.wasm: internal error: ASSERTION FAILED: file rts/sm/Storage.c, line 1216
(GHC version 9.12.2.20250327 for wasm32_unknown_wasi)
Please report this as a GHC bug: https://www.haskell.org/ghc/reportabug
```
What's the culprit? How do I look into this issue?
---
I manually reviewed & revised the patch, tested and submitted it.
-
2b795c2f
by Cheng Shao at 2025-05-20T09:51:24-04:00
compiler: do not allocate strings in bytecode assembler
This patch refactors the compiler to avoid allocating iserv buffers
for BCONPtrStr at assemble-time. Now BCONPtrStr ByteStrings are
recorded as a part of CompiledByteCode, and actual allocation only
happens at link-time. This refactoring is necessary for adding
bytecode serialization functionality, as explained by the revised
comments in this commit.
-
2914d829
by Cheng Shao at 2025-05-20T09:51:24-04:00
compiler: make bc_strs serializable
This commit makes the bc_strs field in CompiledByteCode serializable;
similar to previous commit, we preserve the ByteString directly and
defer the actual allocation to link-time, as mentioned in updated
comment.
-
a6ec1765
by Cheng Shao at 2025-05-20T09:51:24-04:00
compiler: make bc_itbls serializable
This commit makes bc_itbls in CompiledByteCode serializable. A
dedicated ConInfoTable datatype has been added in ghci which is the
recipe for dynamically making a datacon's info table, containing the
payload of the MkConInfoTable iserv message.
-
a238a259
by Cheng Shao at 2025-05-20T09:51:24-04:00
compiler: remove FFIInfo bookkeeping in BCO
This commit removes the bc_ffis field from CompiledByteCode
completely, as well as all the related bookkeeping logic in
GHC.StgToByteCode. bc_ffis is actually *unused* in the rest of GHC
codebase! It is merely a list of FFIInfo, which is just a remote
pointer of the libffi ffi_cif struct; once we allocate the ffi_cif
struct and put its pointer in a CCALL instruction, we'll never free it
anyway. So there is no point of bookkeeping.
-
1f5cc26b
by Cheng Shao at 2025-05-20T09:51:24-04:00
compiler: make FFIInfo serializable in BCO
This commit makes all the FFIInfo needed in CCALL instructions
serializable. Previously, when doing STG to BCO lowering, we would
allocate a libffi ffi_cif struct and keep its remote pointer as
FFIInfo; but actually we can just keep the type signature as FFIInfo
and defer the actual allocation to link-time.
-
20a60364
by Cheng Shao at 2025-05-20T09:51:24-04:00
ghci: remove redundant NewBreakModule message
This commit removes the redundant NewBreakModule message from ghci: it
just allocates two strings! This functionality can be implemented with
existing MallocStrings in one iserv call.
-
bbfd5a5c
by Cheng Shao at 2025-05-20T09:51:24-04:00
compiler: make breakpoint module name and unit id serializable
This commit makes breakpoint module name and unit id serializable, in
BRK_FUN instructions as well as ModBreaks. We can simply keep the
module name and unit ids, and defer the buffer allocation to link
time.
-
aaaf4576
by Cheng Shao at 2025-05-20T09:51:24-04:00
compiler: remove unused newModule
This commit removes the now unused newModule function from GHC.
-
ea0f5482
by Cheng Shao at 2025-05-20T09:51:24-04:00
compiler: add BCONPtrFS for interned top level string literals in BCO
This commit adds BCONPtrFS as a BCO non-pointer literal kind, which
has the same semantics of BCONPtrStr, except it contains a FastString
instead of a ByteString. By using BCONPtrFS to represent top level
string literals that are already FastString in the compilation
pipeline, we enjoy the FastString interning logic and avoid allocating
a bunch of redundant ByteStrings for the same FastStrings, especially
when we lower the BRK_FUN instruction.
-
55b0dfd8
by Peng Fan at 2025-05-20T09:51:36-04:00
hadrian: enable GHCi for loongarch64
-
6c1c9110
by kwxm at 2025-05-20T09:51:45-04:00
Fix bugs in `integerRecipMod` and `integerPowMod`
This fixes #26017.
* `integerRecipMod x 1` now returns `(# 1 | #)` for all x; previously it
incorrectly returned `(# | () #)`, indicating failure.
* `integerPowMod 0 e m` now returns `(# | () #)` for e<0 and m>1, indicating
failure; previously it incorrectly returned `(# 0 | #)`.
-
0155f550
by Andreas Klebinger at 2025-05-20T09:51:46-04:00
Specialise: Don't float out constraint components.
It was fairly complex to do so and it doesn't seem to improve anything.
Nofib allocations were unaffected as well.
See also Historical Note [Floating dictionaries out of cases]