[gcc r13-6418] vect: Don't apply masks to operations on invariants [PR108979]

[gcc r13-6418] vect: Don't apply masks to operations on invariants [PR108979]

[Richard Sandiford]

https://gcc.gnu.org/g:076d309e36c682176e9f85dc8593e6f2c9e6e75f

commit r13-6418-g076d309e36c682176e9f85dc8593e6f2c9e6e75f
Author: Richard Sandiford <richard.sandiford@arm.com>
Date:   Thu Mar 2 16:39:50 2023 +0000

    vect: Don't apply masks to operations on invariants [PR108979]
    
    The loop body in the testcase contains an operation on invariants.
    SLP detects this and can hoist/schedule the operation outside of
    the loop.  However, after the fix for PR96373, we would try to
    apply a loop mask to this operation, even though the mask is
    defined in the loop.
    
    The patch does what Richi suggested in the PR: suppress the
    masking for externs and constants.
    
    gcc/
            PR tree-optimization/108979
            * tree-vect-stmts.cc (vectorizable_operation): Don't mask
            operations on invariants.
    
    gcc/testsuite/
            PR tree-optimization/108979
            * gfortran.dg/vect/pr108979.f90: New test.

Diff:
---
 gcc/testsuite/gfortran.dg/vect/pr108979.f90 | 21 +++++++++++++++++++++
 gcc/tree-vect-stmts.cc                      | 25 +++++++++++++++++++------
 2 files changed, 40 insertions(+), 6 deletions(-)

diff --git a/gcc/testsuite/gfortran.dg/vect/pr108979.f90 b/gcc/testsuite/gfortran.dg/vect/pr108979.f90
new file mode 100644
index 00000000000..623eb67826f
--- /dev/null
+++ b/gcc/testsuite/gfortran.dg/vect/pr108979.f90
@@ -0,0 +1,21 @@
+! { dg-do compile }
+! { dg-additional-options "-fnon-call-exceptions" }
+! { dg-additional-options "-march=armv8.2-a+sve" { target aarch64*-*-* } }
+
+MODULE hfx_contract_block
+  INTEGER, PARAMETER :: dp=8
+CONTAINS
+  SUBROUTINE block_2_1_2_1(kbd,kbc,kad,kac,pbd,pbc,pad,pac,prim,scale)
+    REAL(KIND=dp) :: kbd(1*1), kbc(1*2), kad(2*1), kac(2*2), pbd(1*1), &
+      pbc(1*2), pad(2*1), pac(2*2), prim(2*1*2*1), scale
+      DO md = 1,1
+        DO mc = 1,2
+          DO mb = 1,1
+            DO ma = 1,2
+              kac((mc-1)*2+ma) = kac((mc-1)*2+ma)-tmp*p_bd
+            END DO
+          END DO
+        END DO
+      END DO
+  END SUBROUTINE block_2_1_2_1
+END MODULE hfx_contract_block
diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc
index 77ad8b78506..b56457617c0 100644
--- a/gcc/tree-vect-stmts.cc
+++ b/gcc/tree-vect-stmts.cc
@@ -6254,6 +6254,8 @@ vectorizable_operation (vec_info *vinfo,
                          "use not simple.\n");
       return false;
     }
+  bool is_invariant = (dt[0] == vect_external_def
+		       || dt[0] == vect_constant_def);
   /* If op0 is an external or constant def, infer the vector type
      from the scalar type.  */
   if (!vectype)
@@ -6307,6 +6309,8 @@ vectorizable_operation (vec_info *vinfo,
                              "use not simple.\n");
 	  return false;
 	}
+      is_invariant &= (dt[1] == vect_external_def
+		       || dt[1] == vect_constant_def);
       if (vectype2
 	  && maybe_ne (nunits_out, TYPE_VECTOR_SUBPARTS (vectype2)))
 	return false;
@@ -6321,6 +6325,8 @@ vectorizable_operation (vec_info *vinfo,
                              "use not simple.\n");
 	  return false;
 	}
+      is_invariant &= (dt[2] == vect_external_def
+		       || dt[2] == vect_constant_def);
       if (vectype3
 	  && maybe_ne (nunits_out, TYPE_VECTOR_SUBPARTS (vectype3)))
 	return false;
@@ -6426,16 +6432,23 @@ vectorizable_operation (vec_info *vinfo,
   int reduc_idx = STMT_VINFO_REDUC_IDX (stmt_info);
   vec_loop_masks *masks = (loop_vinfo ? &LOOP_VINFO_MASKS (loop_vinfo) : NULL);
   internal_fn cond_fn = get_conditional_internal_fn (code);
-  bool could_trap = gimple_could_trap_p (stmt);
+
+  /* If operating on inactive elements could generate spurious traps,
+     we need to restrict the operation to active lanes.  Note that this
+     specifically doesn't apply to unhoisted invariants, since they
+     operate on the same value for every lane.
+
+     Similarly, if this operation is part of a reduction, a fully-masked
+     loop should only change the active lanes of the reduction chain,
+     keeping the inactive lanes as-is.  */
+  bool mask_out_inactive = ((!is_invariant && gimple_could_trap_p (stmt))
+			    || reduc_idx >= 0);
 
   if (!vec_stmt) /* transformation not required.  */
     {
-      /* If this operation is part of a reduction, a fully-masked loop
-	 should only change the active lanes of the reduction chain,
-	 keeping the inactive lanes as-is.  */
       if (loop_vinfo
 	  && LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo)
-	  && (could_trap || reduc_idx >= 0))
+	  && mask_out_inactive)
 	{
 	  if (cond_fn == IFN_LAST
 	      || !direct_internal_fn_supported_p (cond_fn, vectype,
@@ -6578,7 +6591,7 @@ vectorizable_operation (vec_info *vinfo,
       vop1 = ((op_type == binary_op || op_type == ternary_op)
 	      ? vec_oprnds1[i] : NULL_TREE);
       vop2 = ((op_type == ternary_op) ? vec_oprnds2[i] : NULL_TREE);
-      if (masked_loop_p && (reduc_idx >= 0 || could_trap))
+      if (masked_loop_p && mask_out_inactive)
 	{
 	  tree mask = vect_get_loop_mask (gsi, masks, vec_num * ncopies,
 					  vectype, i);