| Age | Commit message (Collapse) | Author |
|---|
|
Previously, cproc effectively used used
typedef struct { /* 32 bytes, 8-byte aligned */ } __builtin_va_list[1];
However, this is not quite correct for x86_64 nor aarch64, though
it was close enough for both to work in most cases.
In actuality, for x86_64 we want
typedef struct { /* 24 bytes, 8-byte aligned */ } __builtin_va_list[1];
and for aarch64 we want
typedef struct { /* 32 bytes, 8-byte aligned */ } __builtin_va_list;
The difference only appears when the size of va_list matters, or
when va_list is passed as a parameter. However, the former is not
often the case, and the aarch64 ABI replaces aggregate arguments
with pointers to caller-allocated memory, which is quite similar
to arrays decaying to pointers in C except that the struct is not
copied.
Additionally, riscv64 simply uses
typedef void *__builtin_va_list;
which again has a different size and calling convention.
To fix this, make the __builtin_va_list type architecture-specific
and use architecture-specific tests for varargs-related functionality.
|
|
|
|
|
|
|
|
|
|
Now that we don't track QBE types within values, we don't have to
worry about generating incorrect SSA when passing an 'l' value to
a function taking a 'w'. So we can just return the source value
instead of emitting a dummy copy instruction.
|
|
This way we avoid leaking backend-specific details of type
representation outside qbe.c. It also facilitates some future
simplifications.
|
|
We always store the result to a w temporary, so there is no difference.
In fact, QBE provides loadw as an alias for loadsw precisely for
this reason.
|
|
Make the ID an unsigned int. This will make it small enough to
efficiently pass struct value by value. It also simplifies things
slightly.
|
|
|
|
|
|
It appears that some operating systems don't yet support the C11
DBL_DECIMAL_DIG. In order to ensure consistent output no matter the
precision of long double (which varies from arch to arch), just use
a fixed 17, which is sufficient for IEEE 754 binary64.
|
|
|
|
DECIMAL_DIG may vary from system to system depending on the width
of long double, causing one of the tests to fail.
|
|
Previously, this was needed so that an aggregate type value was
updated to be an 'l' type value. However, since 5ff1d2fa the aggregate
type name is stored in a separate parameter in IARG/ICALL instructions,
so we can just re-use the same pointer value.
|
|
The operands are already promoted, so never need to be extended.
These extend()s were added in 7d746860 in an attempt to fix a bug
related to the comparison of values with type smaller than int.
However, the real bug was that the operands should have been promoted
to int by usual arithmetic conversions, which was fixed later in
a8131372.
|
|
We no longer define __GNUC__ so we don't have to work around glibc's
assert definition anymore.
|
|
|
|
|
|
|
|
Labels are no longer used as instruction arguments.
|
|
Move jump and phi instructions to struct block, and function arguments
to their own instruction.
This will facilitate allocating instructions as an array.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Parameters for K&R function definitions that are affected by default
argument promotion are passed as their promoted type, so need to be
converted back before storing into memory allocated for the parameter.
|
|
|
|
|
|
|
|
This is C99, but is not supported by fscanf in OpenBSD, so floating
constants can't be read back in by QBE.
|
|
The result of funcexpr might be NULL if we are in a dead branch, but
this doesn't mean that we had an invalid object.
|
|
|
|
When we are evaluating an arithmetic constant expression, we don't want
to indroduce static data definitions for string or compound literals.
Fixes #59.
|
|
|
|
Fixes #51.
Thanks to Andrew Chambers for the bug report and test case.
|
|
|
|
|
|
|
|
|
|
|
|
It would be correct in any order, but this is a bit simpler and guarantees
the same output.
|
|
Fixes #47.
|
|
|
|
|
|
Fixes #46. Thanks to Andrew Chambers for the bug report.
|
|
|
|
|
