Unboxed arguments of type `int32` that are passed on the stack
are passed in 32-bit words instead of 64-bit words as in the AAPCS64 ABI.
To support this, we introduce a new specific operation, `Imove32`,
that compiles down to 32-bit moves or 32-bit stack loads or 32-bit
stack stores.
In the Selection pass, method `insert_move_extcall_arg`,
we generate `Imove32` instructions when required, i.e. if the
argument is an unboxed `int32` and needs to be passed on stack.
We then update `Proc.loc_external_arguments` to use 32-bit stack words
for `int32` arguments.
Introduce the type Cmm.exttype to precisely describe arguments to
external C functions, especially unboxed numerical arguments.
Annotate Cmm.Cextcall with the types of the arguments (Cmm.exttype list).
An empty list means "all arguments have default type XInt".
Annotate Mach.Iextcall with the type of the result (Cmm.machtype)
and the types of the arguments (Cmm.exttype list).
Change (slightly) the API for describing calling conventions in Proc:
- loc_external_arguments now takes a Cmm.exttype list,
in order to know more precisely the types of the arguments.
- loc_arguments, loc_parameters, loc_results, loc_external_results
now take a Cmm.machype instead of an array of pseudoregisters.
(Only the types of the pseudoregisters mattered anyway.)
Update the implementations of module Proc accordingly, in every port.
Introduce a new overridable method in Selectgen, insert_move_extcall_arg,
to produce the code that moves an argument of an external C function
to the locations returned by Proc.loc_external_arguments.
Revise the selection of external calls accordingly
(method emit_extcall_args in Selectgen).
Software emulation of floating-point arithmetic, as in the ARM EABI port,
use pairs of pseudoregisters of type Int to represent values of type Float.
This is achieved by the `regs_for` method of class `selector_generic`,
which defaults to `Reg.createv` but is overriden for ARM EABI so
as to perform the transformation Float -> Int,Int on the fly.
The method `bind_let_mut` uses `Reg.createv` to associate
pseudoregisters to bound variables. This is incorrect in a soft FP
context, as a bound variable of type Float will get a Float register
nonetheless. `self#regs_for` must be used instead. This is what this
commit does.
MSVC .lib format doesn't support having no .obj files in the library, so
ocamlopt -o foo.cmxa -a generates foo.cmxa but not foo.lib (there's no
error from the Microsoft Linker). The resulting foo.cmxa is unlinkable,
since OCaml passes foo.lib on to the linker.
This patch relaxes the requirement for foo.lib if the .cmxa contains no
units.
Lambda and Clambda distinguish Const_int from Const_pointer only so
that they can pass the information to Cmm. But now that that
Const_pointer is gone from Cmm (#9578), there's no need for the
distinction in Lambda either.
This PR requires a bootstrap, because the .cmo format changes:
Lambda.structured_constant has one fewer constructor. The bootstrap
is in the following commit.
Since #9316 was merged, Cconst_pointer is compiled in exactly the same way as Cconst_int. This commit removes the now-redundant Cconst_pointer and Cconst_natpointer.
The original implementation of loc_external_arguments and
loc_external_results was following an older ABI,
where an FP argument passed in an FP register "burns" an integer register.
In the ELF psABI, integer registers and FP registers are used independently,
as in the OCaml calling convention. Plus, if all FP registers are used
but an integer register remains, the integer register is used to pass
the next FP argument.
Fixes: #9515
Using instruction fmv.x.d.
This is necessary to implement the ELF psABI calling conventions,
whereas some FP arguments may have to be passed in integer registers.
Fix tail-call optimisation with a mutable ref
(Clet_mutable was not recognized properly in tail position.)
Add a test for tail-call optimisation with a mutable ref
There's no need for allocation on amd64 to clobber the %rax register. It's only used in one case (-compact out-of-line allocation of >3 words), and only used there to do a single subtraction. That subtraction can be done by the caller at no code size penalty, freeing up %rax.
Inside amd64.S functions, %r11 can be used instead of %rax as temporary. %r11 is destroyed by PLT stub code, so on ELF platforms it costs nothing to use.
This PR was tested with also the -dsel, -dlinear output (also fixed to
not-print locations), but the output is architecture-dependent so this
part of the test was removed.
amd64: remove caml_call_gc{1,2,3} and simplify caml_alloc{1,2,3,N}
by tail-calling caml_call_gc.
i386: simplify caml_alloc{1,2,3,N} by tail-calling caml_call_gc.
these functions do not need to preserve ebx.
arm: simplify caml_alloc{1,2,3,N} by tail-calling caml_call_gc.
partial revert of #8619.
arm64: simplify caml_alloc{1,2,3,N} by tail-calling caml_call_gc.
partial revert of #8619.
power: partial revert of #8619.
avoid restarting allocation sequence after failure.
s390: partial revert of #8619.
avoid restarting allocation seqeunce after failure.
Sébastien Hinderer
Xavier Leroy
Fourchaux
EduardoRFS
iOS Porting Team
Xavier Leroy
David Allsopp
Stephen Dolan
Greta Yorsh
Nicolás Ojeda Bär
Gabriel Scherer
Vincent Laviron
Anukriti12
Anukriti12
Greta Yorsh
octachron