On 7/24/24 7:31 AM, Christoph Müllner wrote:
When enabling XTheadMemIdx, we enable the pre- and post-modify
addressing modes in the RISC-V backend.
Unfortunately, the auto_inc_dec pass will then attempt to utilize
this feature regardless of the mode class (i.e. scalar or vector).
The assumption seems to be, that an enabled addressing mode for
scalar instructions will also be available for vector instructions.
In case of XTheadMemIdx and RVV, this is ovbiously not the case.
Still, auto_inc_dec (-O3) performs optimizations like the following:
(insn 23 20 27 3 (set (mem:V4QI (reg:DI 136 [ ivtmp.13 ]) [0 MEM <vector(4)
char> [(char *)_39]+0 S4 A32])
(reg:V4QI 168)) "gcc/testsuite/gcc.target/riscv/pr116033.c":12:27 3183
{*movv4qi}
(nil))
(insn 40 39 41 3 (set (reg:DI 136 [ ivtmp.13 ])
(plus:DI (reg:DI 136 [ ivtmp.13 ])
(const_int 20 [0x14]))) 5 {adddi3}
(nil))
====>
(insn 23 20 27 3 (set (mem:V4QI (post_modify:DI (reg:DI 136 [ ivtmp.13 ])
(plus:DI (reg:DI 136 [ ivtmp.13 ])
(const_int 20 [0x14]))) [0 MEM <vector(4) char> [(char
*)_39]+0 S4 A32])
(reg:V4QI 168)) "gcc/testsuite/gcc.target/riscv/pr116033.c":12:27 3183
{*movv4qi}
(expr_list:REG_INC (reg:DI 136 [ ivtmp.13 ])
(nil)))
The resulting memory-store with post-modify addressing cannot be
lowered to an existing instruction (see PR116033).
At a lower optimization level (-O2) this is currently fine,
but we can't rely on that.
One solution would be to introduce a target hook to check if a certain
type can be used for pre-/post-modify optimizations.
However, it will be hard to justify such a hook, if only a single
RISC-V vendor extension requires that.
Therefore, this patch takes a more drastic approach and disables
pre-/post-modify addressing if TARGET_VECTOR is set.
This results in not emitting pre-/post-modify instructions from
XTheadMemIdx if RVV is enabled.
Note, that this is not an issue with XTheadVector, because
we currently don't have auto-vectorization for that extension.
To ensure this won't changed without being noticed, an additional
test is added.
PR target/116033
gcc/ChangeLog:
* config/riscv/riscv.h (HAVE_POST_MODIFY_DISP): Disable for RVV.
(HAVE_PRE_MODIFY_DISP): Likewise.
This sounds like it's just papering over the real bug, probably in the
address checking code of the backend.
I haven't tried to debug this, but I'd look closely at
th_classify_address which seems to ignore the mode, which seems wrong.
jeff
