This patch fixes a false-positive warning in df-scan, at bootstrap-O3
failed, and enables GCC to optimize out the code that leads to the
warning.
df_ref_create_structure was inlined into the else part of
df_ref_record. Due to the condition of the corresponding if, In the
else part, VRP deduced unsigned regno >= FIRST_PSEUDO_REGISTER.
In df_ref_create_structure, there's another regno variable,
initialized with the same expression and value as the caller's. GCC
can tell as much, but this regno variable is signed. It is used,
shifted right, to index a hard regset bit array within a path that
tests that this signed regno < FIRST_PSEUDO_REGISTER.
GCC warned about the possible out-of-range indexing into the hard
regset array. It shouldn't, after all, the same regno can't possibly
be both < FIRST_PSEUDO_REGISTER and >= FIRST_PSEUDO_REGISTER, can it?
Well, the optimizers correctly decide it could, if it was a negative
int that, when converted to unsigned, became larger than
FIRST_PSEUDO_REGISTER. But GCC doesn't know regno can't be negative,
so the test could not be optimize out. What's more, given the
constraints, VRP correctly concluded the hard regset array would
always be indexed by a value way outside the array index range.
This patch changes the inlined regno to unsigned, like the caller's,
so that we can now tell the conditions can't both hold, so we optimize
out the path containing the would-be out-of-range array indexing.
Regstrapped on x86_64-linux-gnu and i686-linux-gnu. OK to install?
for gcc/ChangeLog
* df-scan.c (df_ref_create_structure): Make regno unsigned,
to match the caller.