Try to clarify what subreg_get_info does and doesn't check.
gcc/
* rtl.h (subreg_info): Expand commentary
* rtlanal.c (subreg_get_info): Likewise.
Index: gcc/rtl.h
===================================================================
--- gcc/rtl.h 2014-09-15 10:00:14.693366097 +0100
+++ gcc/rtl.h 2014-09-15 10:00:14.689366147 +0100
@@ -2866,10 +2866,13 @@ struct subreg_info
{
/* Offset of first hard register involved in the subreg. */
int offset;
- /* Number of hard registers involved in the subreg. */
+ /* Number of hard registers involved in the subreg. In the case of
+ a paradoxical subreg, this is the number of registers that would
+ be modified by writing to the subreg; some of them may be don't-care
+ when reading from the subreg. */
int nregs;
/* Whether this subreg can be represented as a hard reg with the new
- mode. */
+ mode (by adding OFFSET to the original hard register). */
bool representable_p;
};
Index: gcc/rtlanal.c
===================================================================
--- gcc/rtlanal.c 2014-09-15 10:00:14.693366097 +0100
+++ gcc/rtlanal.c 2014-09-15 10:00:14.689366147 +0100
@@ -3411,7 +3411,20 @@ subreg_lsb (const_rtx x)
xmode - The mode of xregno.
offset - The byte offset.
ymode - The mode of a top level SUBREG (or what may become one).
- info - Pointer to structure to fill in. */
+ info - Pointer to structure to fill in.
+
+ Rather than considering one particular inner register (and thus one
+ particular "outer" register) in isolation, this function really uses
+ XREGNO as a model for a sequence of isomorphic hard registers. Thus the
+ function does not check whether adding INFO->offset to XREGNO gives
+ a valid hard register; even if INFO->offset + XREGNO is out of range,
+ there might be another register of the same type that is in range.
+ Likewise it doesn't check whether HARD_REGNO_MODE_OK accepts the new
+ register, since that can depend on things like whether the final
+ register number is even or odd. Callers that want to check whether
+ this particular subreg can be replaced by a simple (reg ...) should
+ use simplify_subreg_regno. */
+
void
subreg_get_info (unsigned int xregno, enum machine_mode xmode,
unsigned int offset, enum machine_mode ymode,
