Get more indexed addressing modes from char pointers.

(1) From Chris Torek -- get the type right after deleting a shift and
upgrading an indexed address from byte to short or long offsets. (2) Clean
out trash in regs[r] for r<NREG when updating uses[r]

(1) From Chris Torek -- get the type right after deleting a shift and
upgrading an indexed address from byte to short or long offsets. (2) Clean
out trash in regs[r] for r<NREG when updating uses[r] and regs[r] in bflow(),
avoiding amazing substitution of operands. (3) flops() fix fails if there are
other uses between the store and the last use; we now loop on uses in flops().

show more ...

Chris Torek's fix for lost labels.

Fix botch with generating OREGs INTEMP.

Two fixes from Chris Torek -- get the type right after deleting a shift
and upgrading an indexed address from byte to short or long offsets; don't
lose a label number after deleting a test-and-branch.

Fix for awful walkf() botch from Chris Torek.

(1) Make the NOT operator work with floating constants. (2) Fix shifts so
that the type depends solely on the left operand, and so that the right
operand has type INT. (3) Eliminate warnings for po

(1) Make the NOT operator work with floating constants. (2) Fix shifts so
that the type depends solely on the left operand, and so that the right
operand has type INT. (3) Eliminate warnings for pointer conversions to
and from void *.

show more ...

Rearrange shift code in offstar() to produce more indexed operands.

Installed much simpler unsigned-to-floating conversion code suggested by
Robert Firth and the tahoe compiler.

(1) We must adjust inoff even if errors have occurred and we aren't
generating code for initializations, since error reporting upstream
depends on it. (2) Use char size offsets for field references

(1) We must adjust inoff even if errors have occurred and we aren't
generating code for initializations, since error reporting upstream
depends on it. (2) Use char size offsets for field references in
indexed structs since we can't generate code for long indexed OREG
field trees. I think the Tahoe compiler did this in the early days.

show more ...

Fix problem with OREGs placed INTEMP.

fix for ANSI C

fix for ANSI C

fix for ANSI C

Fix a goof originally discovered in the vax compiler -- in an assignment,
the temporary register may share with the rhs but not the lhs.

Fixed bug with unsigned short/unsigned char assigned to float/double --
share with rhs, not lhs.

For some reason the cc tests in sconv() were inverted; I believe this was
copied from the original code, but I should have checked this at the time...

Added a couple OREG INTEMP templates that I forgot to copy over from the VAX
table.c when I changed shtemp().

Deleted some puzzling code in flops(), a routine which locates redundant
std/ldd pairs and eliminates the ldd, turning the std into a write to a dead
register which is in turn removed by bflow(). Th

Deleted some puzzling code in flops(), a routine which locates redundant
std/ldd pairs and eliminates the ldd, turning the std into a write to a dead
register which is in turn removed by bflow(). The original coder apparently
wanted to re-run bflow() on all the instructions between the std and the ldd;
this is a gross error, since the uses[] information has been destroyed and
useful instructions can be inadvertently deleted.

show more ...

Better estimate of ediv register requirements in sucomp() -- a function call
in the right operand failed to force a store.

The situation with unsigned char/short comparisons and constants was worse
than I'd realized. The new code works and should be as compact as possible
given the architecture grot.

ZU doesn't handle floating operands, hence ZD was broken.

The tahoe ediv requires a register remainder operand (argh).

Finish installing vax fix for float/double increments.

Finish installing vax fix for float/double increments.