…types usi… (#144676)"
This reverts commit 68471d29eed2c49f9b439e505b3f24d387d54f97.
IntegralAP contains a union:
union {
uint64_t *Memory = nullptr;
uint64_t Val;
};
On 64bit systems, both Memory and Val have the same size. However, on 32
bit system, Val is 64bit and Memory only 32bit. Which means the default
initializer for Memory will only zero half of Val. We fixed this by
zero-initializing Val explicitly in the IntegralAP(unsigned BitWidth)
constructor.
See also the discussion in
https://github.com/llvm/llvm-project/pull/144246
Both `APInt` and `APFloat` will heap-allocate memory themselves using
the system allocator when the size of their data exceeds 64 bits.
This is why clang has `APNumericStorage`, which allocates its memory
using an allocator (via `ASTContext`) instead. Calling `getValue()` on
an ast node like that will then create a new `APInt`/`APFloat` , which
will copy the data (in the `APFloat` case, we even copy it twice).
That's sad but whatever.
In the bytecode interpreter, we have a similar problem. Large integers
and floating-point values are placement-new allocated into the
`InterpStack` (or into the bytecode, which is a `vector<std::byte>`).
When we then later interrupt interpretation, we don't run the destructor
for all items on the stack, which means we leak the memory the
`APInt`/`APFloat` (which backs the `IntegralAP`/`Floating` the
interpreter uses).
Fix this by using an approach similar to the one used in the AST. Add an
allocator to `InterpState`, which is used for temporaries and local
values. Those values will be freed at the end of interpretation. For
global variables, we need to promote the values to global lifetime,
which we do via `InitGlobal` and `FinishInitGlobal` ops.
Interestingly, this results in a slight _improvement_ in compile times:
https://llvm-compile-time-tracker.com/compare.php?from=6bfcdda9b1ddf0900f82f7e30cb5e3253a791d50&to=88d1d899127b408f0fb0f385c2c58e6283195049&stat=instructions:u
(but don't ask me why).
Fixes https://github.com/llvm/llvm-project/issues/139012
In certain cases (i.e. long double on x86), the bit with we get from the
floating point semantics is different than the type size we compute for
the BitCast instruction. Pass this along to DoBitCast, so in there we
can check only the relevant bits for being initialized.
This also fixes a weirdness we still had in DoBitCast.
