On 10/12/15 09:26, Christian Bruel wrote:
Hi Kyrill,
On 12/09/2015 06:32 PM, Kyrill Tkachov wrote:
Hi Christian,
On 08/12/15 12:53, Christian Bruel wrote:
Hi,
The order of the NEON builtins construction has led to complications since the
attribute target support. This was not a problem when driven from the command
line, but was causing various issues when the builtins was mixed between fpu
configurations or when used with LTO.
Firstly the builtin functions was not initialized before the parsing of
functions, leading to wrong type initializations.
Then error catching code when a builtin was used without the proper fpu flags
was incomprehensible for the user, for instance
#include "arm_neon.h"
int8x8_t a, b;
int16x8_t e;
void
main()
{
e = (int16x8_t)__builtin_neon_vaddlsv8qi (a, b);
}
compiled with default options (without -mfpu=neon -mfloat-abi=hard) gave pages
of
/arm-none-eabi/6.0.0/include/arm_neon.h:39:9: error: unknown type name
'__simd64_int8_t'
typedef __simd64_int8_t int8x8_t;
...
...
arm_neon.h:4724:3: error: can't convert a vector of type 'poly64x2_t {aka
__vector(4) int}' to type 'int' which has different size
return (poly64x2_t)__builtin_neon_vsli_nv2di ((int64x2_t) __a, (int64x2_t)
__b, __c);
^~~~~~
...
... and one for each arm_neon.h lines..
by postponing the check into arm_expand_builtin, we now emit something more
useful:
testo.c: In function 'main':
testo.c:9:7: error: '__builtin_neon_vaddlsv8qi' neon builtin is not supported
in this configuration.
e = (int16x8_t)__builtin_neon_vaddlsv8qi (a, b);
^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
One small side effect to note: The total memory allocated is 370k bigger when
neon is not used, so this support will have a follow-up to make their
initialization lazy. But I'd like first to stabilize the stuff for stage3 (or
get it
pre-approved if the memory is an issue)
tested without new failures with {,-mfpu=vfp,-mfpu=neon}{,-march=armv7-a\}
(a few tests that was fail are now unsupported)
I agree, the vector types (re)initialisation is a tricky part.
I've seen similar issues in the aarch64 work for target attributes
bool
arm_vector_mode_supported_p (machine_mode mode)
{
- /* Neon also supports V2SImode, etc. listed in the clause below. */
- if (TARGET_NEON && (mode == V2SFmode || mode == V4SImode || mode == V8HImode
+ if (mode == V2SFmode || mode == V4SImode || mode == V8HImode
|| mode == V4HFmode || mode == V16QImode || mode == V4SFmode
- || mode == V2DImode || mode == V8HFmode))
- return true;
-
- if ((TARGET_NEON || TARGET_IWMMXT)
- && ((mode == V2SImode)
- || (mode == V4HImode)
- || (mode == V8QImode)))
+ || mode == V2DImode || mode == V8HFmode
+ || mode == V2SImode || mode == V4HImode || mode == V8QImode)
return true;
So this allows vector modes unconditionally for all targets/fpu configurations?
I was tempted to do that in aarch64 when I was encountering similar issues.
In the end what worked for me was re-laying out the vector types in
SET_CURRENT_FUNCTION
if necessary (https://gcc.gnu.org/ml/gcc-patches/2015-08/msg01084.html)
yes my assumption was that arm_init_neon_builtins () is now called for all targets, since the check is done at expand time and that the builtins need to be known by lto, with the vector type initialization, before they are expanded.
However at that time, lto streaming-in have not yet processed the attributes and TARGET_NEON is not set for the function.
I had a look at your re-layout, but I'm not sure. it feels like a hack. I think this should be solved first place during the builtin construction. Also set_current_function is too late, builtin_expand that will explode because of the
unknown modes.
But raise the point. In fact I was not really happy with this
arm_vector_mode_supported_p neither as I was not sure about other contexts it
can be called from and I cannot clearly claim that this change is always
correct.
So the main usage of targetm.vector_mode_supported_p is in stor-layout.c and
vector_type_mode in particular seems
to have a relevant comment:
/* Vector types need to re-check the target flags each time we report
the machine mode. We need to do this because attribute target can
change the result of vector_mode_supported_p and have_regs_of_mode
on a per-function basis. Thus the TYPE_MODE of a VECTOR_TYPE can
change on a per-function basis. */
I think that implies that it expects targetm.vector_mode_supported_p to reject
vector modes in
contexts that don't support NEON...
I'd like to think about other way to set the vector modes from
arm_init_neon_builtins before the target flags are known. I'm thinking about
the lazy initialization at expand time, or using a contextual boolean flags.
how does that sound ?
Laying out the vector types during arm_init_neon_builtins sounds more promising
to me.
Changing layout of types during expand is risky, from what I remember.
In principle, the types and builtins created in arm_init_neon_builtins are only
ever supposed to be used in
a NEON context, so I thought that just turning on NEON upon entry into
arm_init_neon_builtins and resetting
it back upon exit would work. However, this won't work because we construct our
builtin types by copying existing
type nodes (e.g. intQI_type_node) that have been laid out earlier by the midend
(frontend?) assuming no NEON.
I wonder if we can explicitly layout these global types in the
arm_init_neon_builtins context...
Thanks,
Kyrill
many thanks,
Christian
Kyrill
