Hi,
GMP 6.1.1 (and probably earlier versions) uses a wrong constraint in the x86
inline assembler code for the MPN_IORD_U macro that blows up (fails to
assemble) on x86-64 when compiled using GCC 6 with link-time optimization
enabled. The problem occurs when a compile-time constant does not fit in a
32-bit immediate operand, so my guess is that it is specifically triggered by
constant propagation across source files.
The code currently uses a constraint that allows any compile-time constant as
an immediate operand, but x86-64 supports only 32-bit immediate values (even in
64-bit instructions), whereas the value MPN_IORD_U uses with this constraint
may be longer than that. GCC, at least since version 3.1.1, actually provides a
constraint just for such an occasion, which allows a compile-time constant as
an immediate operand if and only if its value fits in 32 bits. In fact, GCC
provides two such constraints: one for signed values (for sign-extending
instructions) and one for unsigned values (for zero-extending instructions).
MPN_IORD_U is used only with sign-extending instructions (ADD and SUB), so the
fix is easy: simply replace the currently used constraint with the signed
32-bit immediate constraint.
I have not investigated whether the same bug appears in other places. GMP
currently builds successfully for me with the below patch applied.
Here is a patch:
diff --git a/gmp-impl.h b/gmp-impl.h
index 6987581..dcc233b 100644
--- a/gmp-impl.h
+++ b/gmp-impl.h
@@ -2687,7 +2687,7 @@ __GMP_DECLSPEC mp_bitcnt_t mpn_remove (mp_ptr, mp_size_t
*, mp_srcptr, mp_size_t
ASM_L(done) ":\n" \
: "=r" (__ptr_dummy) \
: "0" (ptr), \
- "ri" ((mp_limb_t) (incr)), "n" (sizeof(mp_limb_t)) \
+ "re" ((mp_limb_t) (incr)), "n" (sizeof(mp_limb_t)) \
: "memory"); \
}
\
} while (0)
_______________________________________________
gmp-bugs mailing list
[email protected]
https://gmplib.org/mailman/listinfo/gmp-bugs
