On 05/01/17 12:01, Richard Biener wrote:
On Wed, Jan 4, 2017 at 4:07 PM, Kyrill Tkachov
<kyrylo.tkac...@foss.arm.com> wrote:
On 04/01/17 14:19, Richard Biener wrote:
On Wed, Dec 21, 2016 at 10:40 AM, Kyrill Tkachov
<kyrylo.tkac...@foss.arm.com> wrote:
On 20/12/16 17:30, Richard Biener wrote:
On December 20, 2016 5:01:19 PM GMT+01:00, Kyrill Tkachov
<kyrylo.tkac...@foss.arm.com> wrote:
Hi all,
The testcase in this patch generates bogus assembly for arm with -O1
-mfloat-abi=soft:
strd r4, [#0, r3]
This is due to non-canonical RTL being generated during expansion:
(set (mem:DI (plus:SI (const_int 0 [0])
(reg/f:SI 153)) [0 MEM[symbol: a, index: _26, offset: 0B]+0 S8
A64])
(reg:DI 154))
Note the (plus (const_int 0) (reg)). This is being generated in
gen_addr_rtx in tree-ssa-address.c
where it creates an explicit PLUS rtx through gen_rtx_PLUS, which
doesn't try to canonicalise its arguments
or simplify. The correct thing to do is to use simplify_gen_binary that
will handle all this properly.
But it has to match up the validity check which passes down exactly the
same RTL(?) Or does this stem from propagation simplifying a MEM after
IVOPTs?
You mean TARGET_LEGITIMATE_ADDRESS_P? Yes, it gets passed on to that, but
the arm implementation of that
doesn't try to handle non-canonical RTL (plus (const0_rtx) (reg) is not
canonical).
Or do you mean some other check?
Ok, now looking at the actual patch and the code in question. For your
testcase
it happens that symbol == const0_rtx? In this case please amend the
if (symbol) check like we do for the base, thus
if (symbol && symbol != const0_rtx)
No, symbol is not const0_rtx (it's just a symbol_ref).
index is const0_rtx and so gets assigned to *addr at the beginning of the
function.
base and step are NULL_RTX.
So at the time of the check:
if (*addr)
*addr = gen_rtx_PLUS (address_mode, *addr, act_elem);
else
*addr = act_elem;
*addr is const0_rtx. Do you want to amend that check to:
if (*addr && *addr != const0_rtx) ?
Hmm, I guess given the if (step) in if (index) *addr can end up being
a not simplified mult. So instead do
if (index && index != const0_rtx)
That works, I'll test a patch for this.
I haven't looked where the const0_rtx index comes from, so I don't know if
it
could have other CONST_INT values that may cause trouble.
Usually this happens when constant folding / propagation happens after
IVOPTs generates the TARGET_MEM_REF. We do have some canonicalization
foldings for TMR, see maybe_fold_tmr, but that should have made index NULL
if it was constant... So maybe we fail to fold a stmt at some point.
Btw, I fail to see the bogus asm with my arm-non-eabi cross with -O
-mfloat-abi=soft
so I can't tell how the TMR arrives at RTL expansion.
You'll also want to specify -marm (this doesn't trigger on Thumb) and perhaps
-march=armv7-a.
Thanks,
Kyrill
Richard.
Kyrill
Richard.
Thanks,
Kyrill
I didn't change the other gen_rtx_PLUS calls in this function as their
results is later used in XEXP operations
that seem to rely on a PLUS expression being explicitly produced, but
this particular call doesn't, so it's okay
to change it. With this patch the sane assembly is generated:
strd r4, [r3]
Bootstrapped and tested on arm-none-linux-gnueabihf, x86_64,
aarch64-none-linux-gnu.
Ok for trunk?
Thanks,
Kyrill
2016-12-20 Kyrylo Tkachov <kyrylo.tkac...@arm.com>
* tree-ssa-address.c (gen_addr_rtx): Use simplify_gen_binary to
add
*addr to act_elem.
2016-12-20 Kyrylo Tkachov <kyrylo.tkac...@arm.com>
* gcc.dg/20161219.c: New test.
