Re: [PATCH] Lift restrictions on SLP permutation for loop vect

[Richard Biener <rguenther@suse.de>]

On Thu, 11 Jun 2015, James Greenhalgh wrote:

> On Thu, Jun 11, 2015 at 03:08:59PM +0100, Richard Biener wrote:
> > On Thu, 11 Jun 2015, Uros Bizjak wrote:
> > 
> > > > So this turned up other issues thus the following is what I have
> > > > committed after bootstrapping and testing on x86_64-unknown-linux-gnu.
> > > >
> > > > Richard.
> > > >
> > > > 2015-06-08  Richard Biener  <rguenther@suse.de>
> > > >
> > > > * tree-vect-stmts.c (vectorizable_load): Compute the pointer
> > > > adjustment for gaps at the end of a SLP load group properly.
> > > > * tree-vect-slp.c (vect_supported_load_permutation_p): Allow
> > > > all permutations we can generate.
> > > > (vect_transform_slp_perm_load): Use the correct group-size.
> > > >
> > > > * gcc.dg/vect/slp-perm-10.c: New testcase.
> > > > * gcc.dg/vect/slp-23.c: Adjust.
> > > > * gcc.dg/torture/pr53366-2.c: Also verify cross-iteration vector pointer update.
> > > 
> > > This patch caused:
> > > 
> > > FAIL: gcc.target/i386/pr61403.c scan-assembler blend
> > 
> > Yeah, I noticed.  We now want to vectorize this differently but
> > fail due to the cost model.  I'm working on enhancing the vectorizer
> > here.
> 
> It also caused an ICE in the ARM port (arm-none-eabi,
> arm-none-linux-gnueabihf):
> 
>     FAIL: gcc.target/arm/pr53636.c (internal compiler error)
> 
> Full ICE text below, and reduced testcase attached, compile with:
> 
>     arm-none-eabi-gcc -O -ftree-vectorize -mfpu=neon -mcpu=cortex-a9 bug.c
> 
> I tried to take a look to see what was happening, but I couldn't see
> the root of the problem. The access to dr_chain in
> vect_create_mask_and_perm:
> 
>     second_vec = dr_chain[second_vec_indx];
> 
> Fails as dr_chain has length 1, and second_vec_indx is 2.
> 
> I think that the mask that the code is trying to produce is { 1, 2, 3, 4 }.
> 
>     bug.c:4:3: note: add new stmt: vect__8.6_108 = VEC_PERM_EXPR <vect__8.4_104, vect__8.5_106, { 1, 2, 3, 4 }>;
> 
> But that's about as far as I got.

I will have a look next week - mind opening a bugreport for this so
I dont' forget?

Thanks,
Richard.

> Thanks,
> James
> 
> ---
> bug.c: In function 'test':
> bug.c:1:6: internal compiler error: in operator[], at vec.h:738
>  void test(unsigned char *dst) {
>       ^
> 0xd759fe vec<tree_node*, va_heap, vl_embed>::operator[](unsigned int)
> 	.../src/gcc/gcc/vec.h:738
> 0xd759fe vec<tree_node*, va_heap, vl_ptr>::operator[](unsigned int)
> 	.../src/gcc/gcc/vec.h:1204
> 0xd759fe vect_create_mask_and_perm
> 	.../src/gcc/gcc/tree-vect-slp.c:3072
> 0xd759fe vect_transform_slp_perm_load(_slp_tree*, vec<tree_node*, va_heap, vl_ptr>, gimple_stmt_iterator*, int, _slp_instance*, bool)
> 	.../src/gcc/gcc/tree-vect-slp.c:3350
> 0xd51613 vectorizable_load
> 	.../src/gcc/gcc/tree-vect-stmts.c:6847
> 0xd57ad2 vect_transform_stmt(gimple_statement_base*, gimple_stmt_iterator*, bool*, _slp_tree*, _slp_instance*)
> 	.../src/gcc/gcc/tree-vect-stmts.c:7490
> 0xd7aac1 vect_schedule_slp_instance
> 	.../src/gcc/gcc/tree-vect-slp.c:3500
> 0xd7a117 vect_schedule_slp_instance
> 	.../src/gcc/gcc/tree-vect-slp.c:3381
> 0xd7a117 vect_schedule_slp_instance
> 	.../src/gcc/gcc/tree-vect-slp.c:3381
> 0xd7a117 vect_schedule_slp_instance
> 	.../src/gcc/gcc/tree-vect-slp.c:3381
> 0xd7a117 vect_schedule_slp_instance
> 	.../src/gcc/gcc/tree-vect-slp.c:3381
> 0xd7a117 vect_schedule_slp_instance
> 	.../src/gcc/gcc/tree-vect-slp.c:3381
> 0xd7a117 vect_schedule_slp_instance
> 	.../src/gcc/gcc/tree-vect-slp.c:3381
> 0xd7a117 vect_schedule_slp_instance
> 	.../src/gcc/gcc/tree-vect-slp.c:3381
> 0xd7abce vect_schedule_slp(_loop_vec_info*, _bb_vec_info*)
> 	.../src/gcc/gcc/tree-vect-slp.c:3570
> 0xd5e564 vect_transform_loop(_loop_vec_info*)
> 	.../src/gcc/gcc/tree-vect-loop.c:6223
> 0xd7eca8 vectorize_loops()
> 	.../src/gcc/gcc/tree-vectorizer.c:499
> 0xc88c54 execute
> 	.../src/gcc/gcc/tree-ssa-loop.c:292
> Please submit a full bug report,
> with preprocessed source if appropriate.
> Please include the complete backtrace with any bug report.
> See <http://gcc.gnu.org/bugs.html> for instructions.
> 
> 

-- 
Richard Biener <rguenther@suse.de>
SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard, Dilip Upmanyu, Graham Norton, HRB 21284 (AG Nuernberg)