| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| |
|
|
|
|
|
|
|
|
|
| |
With FP arithmetic, the negation of "x < y" is not "x >= y". For this reason, the back-end intermediate languages of CompCert used to have both "Ccompf c" and "Cnotcompf c" comparison operators, where "c" is of type Integers.comparison and "Cnotcompf c" denotes the negation of FP comparison c.
There are some problems with this approach:
- Beyond Cnotcompf we also need Cnotcompfs (for single precision FP) and, in case of ARM, special forms for not-comparison against 0.0. This duplication of comparison constructors inevitably causes some code and proof duplication.
- Cnotcompf Ceq is really Ccompf Cne, and likewise Cnotcompf Cne is really Ccompf Ceq, hence the representation of FP comparisons is not canonical, adding to the code and proof duplication mentioned above.
- Cnotcompf is introduced in CminorSel, but in Cminor we don't have it, making it impossible to express some transformations over comparisons at the machine-independent Cminor level.
This commit develops an alternate approach, whereas FP comparisons have their own type, defined as Floats.fp_comparison, and which includes constructors for "not <", "not <=", "not >" and "not >=". Hence this type is closed under boolean negation, so to speak, and there is no longer a need for "Cnotcompf", given that "Ccompf" takes a fp_comparison and can therefore express all FP comparisons of interest.
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
In 32-bit mode, a symbolic reference "symbol + ofs" (address of "symbol" plus "ofs" bytes) can always be resolved by the linker into a 32-bit quantity that fits the 32-bit displacement field of x86 addressing modes.
Not so in 64-bit mode: first, the displacement field is still 32 bits but the full address is 64 bits; second, the displacement is relative to the RIP instruction pointer. In the "small code model" that CompCert uses for x86-64, excessively large offsets lead to link-time overflows of this 32-bit displacement field.
This commit addresses the issue by limiting the "ofs" part of "symbol + ofs" global addressing models to the range [-2^24, 2^24 - 1]. As explained in the AMD64 ELF ABI document, this is a safe range in the small code model, under the assumption that no global symbol is bigger than 2^24 bytes. GCC seems to be using a wider range [-2^31, 2^24 - 1] but I'd rather be safe.
The limitation of the "ofs" offset is achieved by extending the mechanisms already present to ensure that "ofs" in "reg + ofs" indexed addressing modes fits in 32-bit signed:
- Op.addressing_valid checks that the "ofs" part of "symbol + ofs" addressing modes is in the correct interval;
- SelectOp.addressing turns invalid addressings into lea's + indexed addressings;
- Asmgen.normalize_addrmode_64 turns lea's with invalid addressings into simpler lea's + addq of the large offset.
|
| |
|
|
|
| |
Open Local becomes Local Open. This silences Coq 8.6's warning.
Also: remove one useless Require-inside-a-module that caused another warning.
|
|
|
-> x86/x86_32/x86_64
Having Archi.ptr64 as an opaque Parameter that is determined at run-time depending on compcert.ini is problematic for applications such as VST where functions such as Ctypes.sizeof must compute within Coq.
This commit introduces two versions of the Archi.v file, one for x86 32 bits (with ptr64 := false), one for x86 64 bits (with ptr64 := true). Unlike previous approaches, no other file is duplicated between these two variants of x86.
While we are at it, I renamed "ia32" into "x86" everywhere. "ia32" is Intel speak for the 32-bit architecture. It is not a good name to describe both the 32 and 64 bit architectures.
Finally, .depend is no longer under version control and is regenerated when the target architecture changes. That's because the location of Archi.v differs between the ports that have 32/64 bit variants (x86 so far) and the ports that have only one bitsize (ARM and PowerPC so far).
|
