target/8087: sparc-sun-solaris2.7 C testsuite failures in execute/20020720-1.c w/-m64 or on sparcv9/sparc64

[Roger Sayle]

>    The fixes that I'm still investigating address the sparc64 backend's
>    convoluted constant pool loading code.
>
> There is nothing convoluted about it.  That's a perfectly valid
> and fine way to output a load from the constant pool and expose
> all of the instructions to the compiler for scheduling purposes.

I beg to differ.

The sparc64 backend represents a load from the constant pool as:

(mem/u/f:DF (lo_sum:DI (reg/f:DI 110)
                       (symbol_ref/u:DI ("*.LLC0"))) [2 S8 A64]

The problem is that the code in "avoid_constant_pool_reference"
in simplify-rtx.c (line 149), assumes that constant pool references
are of the form "(mem (symbol_ref ...))".  Indeed the macro
CONSTANT_POOL_ADDRESS_P assumes that it is always passed a naked
symbol_ref.

A possible fix may be to extend this test is also allow the constant
pool to be indexed via LO_SUM.  Something like:

	if (GET_CODE (x) != MEM)
	  return x;
	addr = XEXP (x, 0)

+	if (GET_CODE (addr) == LO_SUM
+	    && GET_CODE (XEXP (addr, 1)) == SYMBOL_REF)
+	  addr = XEXP (addr, 1);

	if (GET_CODE (addr) != SYMBOL_REF
	    || ! CONSTANT_POOL_ADDRESS_P (addr))
	  return x;


Could someone that understands LO_SUM targets confirm that this is
a reasonable thing to do?  I'm assuming that the ".LLC0" symbol_ref
will be correctly marked with CONSTANT_POOL_ADDRESS_P.  If it is,
this should not only solve 20020720-1.c, but also enable many more
optimizations on sparc64.  RTH?

Roger
--