Hi Scott,
thanks for having a look.
On Mon, Nov 28, 2022 at 03:14:59PM +0000, Scott Ashcroft wrote:
> I've done some work on why the mips64el build fails.
> The issue is that the test script tests/sat/grom.ys makes yosys core in
> libs/fst/fstapi.cc:fstGetUint32
>
> Looking at the code it is clear that, when FST_DO_MISALIGNED_OPS is not
> defined, an address on the stack is returned to calling function.
I'm not sure what you mean here, my reading of the code is that it's just
tring to read a big-endian uint32_t at the given address regardless of
alignment. No address is returned.
> Because the Debian build uses various hardening flags to protect
> accesses to the stack that makes a segfault.
While the code is indeed a bit iffy TBH I don't really see why it should
crash even with the stack protector etc. enabled.
Regardless if you still have a test environment could you try the attached
patch that turns the roundabout u8 to u32 conversion into a (hopefully)
more portable roundtrip through a union.
--Daniel
--- a/libs/fst/fstapi.cc
+++ b/libs/fst/fstapi.cc
@@ -348,17 +348,17 @@ static void *fstMmap2(size_t __len, int
#ifdef FST_DO_MISALIGNED_OPS
#define fstGetUint32(x) (*(uint32_t *)(x))
#else
-static uint32_t fstGetUint32(unsigned char *mem)
+static inline uint32_t fstGetUint32(unsigned char *mem)
{
- uint32_t u32;
- unsigned char *buf = (unsigned char *)(&u32);
+ union {
+ uint8_t u8[sizeof(uint32_t)];
+ uint32_t u32;
+ } u;
- buf[0] = mem[0];
- buf[1] = mem[1];
- buf[2] = mem[2];
- buf[3] = mem[3];
+ for (int i=0; i < sizeof(u.u8); i++)
+ u.u8[i] = mem[i];
- return (*(uint32_t *)buf);
+ return u.u32;
}
#endif
