Re: [aarch64] PR107920 - Fix incorrect handling of virtual operands in svld1rq_impl::fold

[Richard Sandiford <richard.sandiford@arm.com>]

Prathamesh Kulkarni <prathamesh.kulkarni@linaro.org> writes:
> Hi,
> The following test:
>
> #include "arm_sve.h"
>
> svint8_t
> test_s8(int8_t *x)
> {
>   return svld1rq_s8 (svptrue_b8 (), &x[0]);
> }
>
> ICE's with -march=armv8.2-a+sve -O1 -fno-tree-ccp -fno-tree-forwprop:
> during GIMPLE pass: fre
> pr107920.c: In function ‘test_s8’:
> pr107920.c:7:1: internal compiler error: in execute_todo, at passes.cc:2140
>     7 | }
>       | ^
> 0x7b03d0 execute_todo
>         ../../gcc/gcc/passes.cc:2140
>
> because of incorrect handling of virtual operands in svld1rq_impl::fold:
>  # VUSE <.MEM>
>   _5 = MEM <vector(16) signed char> [(signed char * {ref-all})x_3(D)];
>   _4 = VEC_PERM_EXPR <_5, _5, { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11,
> 12, 13, 14, 15, ... }>;
>   # VUSE <.MEM_2(D)>
>   return _4;
>
> The attached patch tries to fix the issue by building the replacement
> statements in gimple_seq, and passing it to gsi_replace_with_seq_vops,
> which resolves the ICE, and results in:
>   <bb 2> :
>   # VUSE <.MEM_2(D)>
>   _5 = MEM <vector(16) signed char> [(signed char * {ref-all})x_3(D)];
>   _4 = VEC_PERM_EXPR <_5, _5, { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11,
> 12, 13, 14, 15, ... }>;
>   # VUSE <.MEM_2(D)>
>   return _4;
>
> Bootstrapped+tested on aarch64-linux-gnu.
> OK to commit ?

Looks good, but we also need to deal with the -fnon-call-exceptions
point that Andrew made in the PR trail.  An easy way of testing
would be to make sure that:

#include "arm_sve.h"

svint8_t
test_s8(int8_t *x)
{
  try
    {
      return svld1rq_s8 (svptrue_b8 (), &x[0]);
    }
  catch (...)
    {
      return svdup_s8 (1);
    }
}

compiled with -fnon-call-exceptions still has a call to __cxa_begin_catch.

I don't think it's worth optimising this case.  Let's just add
!flag_non_call_exceptions to the test.

The patch is missing a changelog btw.

Thanks,
Richard

> Thanks,
> Prathamesh
>
> diff --git a/gcc/config/aarch64/aarch64-sve-builtins-base.cc b/gcc/config/aarch64/aarch64-sve-builtins-base.cc
> index 6347407555f..f5546a65d22 100644
> --- a/gcc/config/aarch64/aarch64-sve-builtins-base.cc
> +++ b/gcc/config/aarch64/aarch64-sve-builtins-base.cc
> @@ -45,6 +45,7 @@
>  #include "aarch64-sve-builtins-base.h"
>  #include "aarch64-sve-builtins-functions.h"
>  #include "ssa.h"
> +#include "gimple-fold.h"
>  
>  using namespace aarch64_sve;
>  
> @@ -1232,7 +1233,9 @@ public:
>  	tree mem_ref_op = fold_build2 (MEM_REF, access_type, arg1, zero);
>  	gimple *mem_ref_stmt
>  	  = gimple_build_assign (mem_ref_lhs, mem_ref_op);
> -	gsi_insert_before (f.gsi, mem_ref_stmt, GSI_SAME_STMT);
> +
> +	gimple_seq stmts = NULL;
> +	gimple_seq_add_stmt_without_update (&stmts, mem_ref_stmt);
>  
>  	int source_nelts = TYPE_VECTOR_SUBPARTS (access_type).to_constant ();
>  	vec_perm_builder sel (lhs_len, source_nelts, 1);
> @@ -1245,8 +1248,11 @@ public:
>  						   indices));
>  	tree mask_type = build_vector_type (ssizetype, lhs_len);
>  	tree mask = vec_perm_indices_to_tree (mask_type, indices);
> -	return gimple_build_assign (lhs, VEC_PERM_EXPR,
> -				    mem_ref_lhs, mem_ref_lhs, mask);
> +	gimple *g2 = gimple_build_assign (lhs, VEC_PERM_EXPR,
> +					  mem_ref_lhs, mem_ref_lhs, mask);
> +	gimple_seq_add_stmt_without_update (&stmts, g2);
> +	gsi_replace_with_seq_vops (f.gsi, stmts);
> +	return g2;
>        }
>  
>      return NULL;
> diff --git a/gcc/gimple-fold.cc b/gcc/gimple-fold.cc
> index c2d9c806aee..03cdb2f9f49 100644
> --- a/gcc/gimple-fold.cc
> +++ b/gcc/gimple-fold.cc
> @@ -591,7 +591,7 @@ fold_gimple_assign (gimple_stmt_iterator *si)
>     If the statement has a lhs the last stmt in the sequence is expected
>     to assign to that lhs.  */
>  
> -static void
> +void
>  gsi_replace_with_seq_vops (gimple_stmt_iterator *si_p, gimple_seq stmts)
>  {
>    gimple *stmt = gsi_stmt (*si_p);
> diff --git a/gcc/gimple-fold.h b/gcc/gimple-fold.h
> index 7d29ee9a9a4..87ed4e56d25 100644
> --- a/gcc/gimple-fold.h
> +++ b/gcc/gimple-fold.h
> @@ -63,6 +63,7 @@ extern bool arith_code_with_undefined_signed_overflow (tree_code);
>  extern gimple_seq rewrite_to_defined_overflow (gimple *, bool = false);
>  extern void replace_call_with_value (gimple_stmt_iterator *, tree);
>  extern tree tree_vec_extract (gimple_stmt_iterator *, tree, tree, tree, tree);
> +extern void gsi_replace_with_seq_vops (gimple_stmt_iterator *, gimple_seq);
>  
>  /* gimple_build, functionally matching fold_buildN, outputs stmts
>     int the provided sequence, matching and simplifying them on-the-fly.
> diff --git a/gcc/testsuite/gcc.target/aarch64/sve/acle/general/pr107920.c b/gcc/testsuite/gcc.target/aarch64/sve/acle/general/pr107920.c
> new file mode 100644
> index 00000000000..11448ed5e68
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/aarch64/sve/acle/general/pr107920.c
> @@ -0,0 +1,10 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O1 -fno-tree-ccp -fno-tree-forwprop" } */
> +
> +#include "arm_sve.h"
> +
> +svint8_t
> +test_s8(int8_t *x)
> +{
> +  return svld1rq_s8 (svptrue_b8 (), &x[0]);
> +}