The sh64-*-* port plays fast and loos with the distinction between 32 and 64 bit
values. In particular, it pretends a truncation from 64 to 32 bit is a no-op,
while a sign extension is needed, and that the sh5-32media and sh-compact
subtargets can expose the indexed addressing mode directly to the compiler,
which is wrong, too, since this addressing mode uses 64 bit arithmetic and
traps on all existing hardware if the result is not a valid 64 bit address.
On the other hand, sign extensions are usually no-ops.
Moreover, the target address of a branch should not be hardwired to DImode,
but always be Pmode.
The merge I am working on addresses all these issues, but it depends on all
LABEL_REFS to have Pmode; VOIDmode LABEL_REFs cause reload failures.
--
Summary: sh64-*-* port deos not handle 32 / 64 bit conversions
properly
Product: gcc
Version: 3.0.4
Status: UNCONFIRMED
Keywords: wrong-code, missed-optimization
Severity: normal
Priority: P2
Component: target
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: amylaar at gcc dot gnu dot org
CC: gcc-bugs at gcc dot gnu dot org
GCC target triplet: sh64-*-*
BugsThisDependsOn: 20413
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=20695
