File _service:tar_scm:0129-LoongArch-Adjust-cost-of-... of Package gcc

File _service:tar_scm:0129-LoongArch-Adjust-cost-of-vector_stmt-that-match-mult.patch of Package gcc

From 825847768a29ec9d50e01015167002998150cb27 Mon Sep 17 00:00:00 2001
From: Li Wei <liwei@loongson.cn>
Date: Fri, 26 Jan 2024 16:41:11 +0800
Subject: [PATCH 129/188] LoongArch: Adjust cost of vector_stmt that match
 multiply-add pattern.

We found that when only 128-bit vectorization was enabled, 549.fotonik3d_r
failed to vectorize effectively. For this reason, we adjust the cost of
128-bit vector_stmt that match the multiply-add pattern to facilitate 128-bit
vectorization.
The experimental results show that after the modification, 549.fotonik3d_r
performance can be improved by 9.77% under the 128-bit vectorization option.

gcc/ChangeLog:

* config/loongarch/loongarch.cc (loongarch_multiply_add_p): New.
	(loongarch_vector_costs::add_stmt_cost): Adjust.

gcc/testsuite/ChangeLog:

* gfortran.dg/vect/vect-10.f90: New test.
---
 gcc/config/loongarch/loongarch.cc          | 48 +++++++++++++++
 gcc/testsuite/gfortran.dg/vect/vect-10.f90 | 71 ++++++++++++++++++++++
 2 files changed, 119 insertions(+)
 create mode 100644 gcc/testsuite/gfortran.dg/vect/vect-10.f90

diff --git a/gcc/config/loongarch/loongarch.cc b/gcc/config/loongarch/loongarch.cc
index b8f0291ab..526ea0bcb 100644
--- a/gcc/config/loongarch/loongarch.cc
+++ b/gcc/config/loongarch/loongarch.cc
@@ -4153,6 +4153,37 @@ loongarch_vector_costs::determine_suggested_unroll_factor (loop_vec_info loop_vi
   return 1 << ceil_log2 (uf);
 }
 
+/* Check if assign stmt rhs op comes from a multiply-add operation.  */
+static bool
+loongarch_multiply_add_p (vec_info *vinfo, stmt_vec_info stmt_info)
+{
+  gassign *assign = dyn_cast<gassign *> (stmt_info->stmt);
+  if (!assign)
+    return false;
+  tree_code code = gimple_assign_rhs_code (assign);
+  if (code != PLUS_EXPR && code != MINUS_EXPR)
+    return false;
+
+  auto is_mul_result = [&](int i)
+    {
+      tree rhs = gimple_op (assign, i);
+      if (TREE_CODE (rhs) != SSA_NAME)
+	return false;
+
+      stmt_vec_info def_stmt_info = vinfo->lookup_def (rhs);
+      if (!def_stmt_info
+	  || STMT_VINFO_DEF_TYPE (def_stmt_info) != vect_internal_def)
+	return false;
+      gassign *rhs_assign = dyn_cast<gassign *> (def_stmt_info->stmt);
+      if (!rhs_assign || gimple_assign_rhs_code (rhs_assign) != MULT_EXPR)
+	return false;
+
+      return true;
+    };
+
+  return is_mul_result (1) || is_mul_result (2);
+}
+
 unsigned
 loongarch_vector_costs::add_stmt_cost (int count, vect_cost_for_stmt kind,
 				       stmt_vec_info stmt_info, slp_tree,
@@ -4165,6 +4196,23 @@ loongarch_vector_costs::add_stmt_cost (int count, vect_cost_for_stmt kind,
     {
       int stmt_cost = loongarch_builtin_vectorization_cost (kind, vectype,
 							    misalign);
+      if (vectype && stmt_info)
+	{
+	  gassign *assign = dyn_cast<gassign *> (STMT_VINFO_STMT (stmt_info));
+	  machine_mode mode = TYPE_MODE (vectype);
+
+	  /* We found through testing that this strategy (the stmt that
+	     matches the multiply-add pattern) has positive returns only
+	     when applied to the 128-bit vector stmt, so this restriction
+	     is currently made.  */
+	  if (kind == vector_stmt && GET_MODE_SIZE (mode) == 16 && assign)
+	    {
+	      if (!vect_is_reduction (stmt_info)
+		  && loongarch_multiply_add_p (m_vinfo, stmt_info))
+		stmt_cost = 0;
+	    }
+	}
+
       retval = adjust_cost_for_freq (stmt_info, where, count * stmt_cost);
       m_costs[where] += retval;
 
diff --git a/gcc/testsuite/gfortran.dg/vect/vect-10.f90 b/gcc/testsuite/gfortran.dg/vect/vect-10.f90
new file mode 100644
index 000000000..b85bc2702
--- /dev/null
+++ b/gcc/testsuite/gfortran.dg/vect/vect-10.f90
@@ -0,0 +1,71 @@
+! { dg-do compile }
+! { dg-additional-options "-Ofast -mlsx -fvect-cost-model=dynamic" { target loongarch64*-*-* } }
+
+MODULE material_mod
+
+IMPLICIT NONE
+
+integer, parameter :: dfp = selected_real_kind (13, 99)
+integer, parameter :: rfp = dfp
+
+PUBLIC Mat_updateE, iepx, iepy, iepz
+
+PRIVATE
+
+integer, dimension (:, :, :), allocatable :: iepx, iepy, iepz
+real (kind = rfp), dimension (:), allocatable :: Dbdx, Dbdy, Dbdz
+integer :: imin, jmin, kmin
+integer, dimension (6) :: Exsize
+integer, dimension (6) :: Eysize
+integer, dimension (6) :: Ezsize
+integer, dimension (6) :: Hxsize
+integer, dimension (6) :: Hysize
+integer, dimension (6) :: Hzsize
+
+CONTAINS
+
+SUBROUTINE mat_updateE (nx, ny, nz, Hx, Hy, Hz, Ex, Ey, Ez)
+
+integer, intent (in) :: nx, ny, nz
+
+real (kind = rfp), intent (inout),                                         &
+  dimension (Exsize (1) : Exsize (2), Exsize (3) : Exsize (4), Exsize (5) : Exsize (6)) :: Ex
+real (kind = rfp), intent (inout),                                         &
+  dimension (Eysize (1) : Eysize (2), Eysize (3) : Eysize (4), Eysize (5) : Eysize (6)) :: Ey
+real (kind = rfp), intent (inout),                                         &
+  dimension (Ezsize (1) : Ezsize (2), Ezsize (3) : Ezsize (4), Ezsize (5) : Ezsize (6)) :: Ez
+real (kind = rfp), intent (in),                                            &
+  dimension (Hxsize (1) : Hxsize (2), Hxsize (3) : Hxsize (4), Hxsize (5) : Hxsize (6)) :: Hx
+real (kind = rfp), intent (in),                                            &
+  dimension (Hysize (1) : Hysize (2), Hysize (3) : Hysize (4), Hysize (5) : Hysize (6)) :: Hy
+real (kind = rfp), intent (in),                                            &
+  dimension (Hzsize (1) : Hzsize (2), Hzsize (3) : Hzsize (4), Hzsize (5) : Hzsize (6)) :: Hz
+
+integer :: i, j, k, mp
+
+do k = kmin, nz
+  do j = jmin, ny
+    do i = imin, nx
+      mp = iepx (i, j, k)
+      Ex (i, j, k) = Ex (i, j, k) +                                        &
+                  Dbdy (mp) * (Hz (i, j, k  ) - Hz (i, j-1, k)) +          &
+                  Dbdz (mp) * (Hy (i, j, k-1) - Hy (i, j  , k))
+
+      mp = iepy (i, j, k)
+      Ey (i, j, k) = Ey (i, j, k) +                                        &
+                  Dbdz (mp) * (Hx (i  , j, k) - Hx (i, j, k-1)) +          &
+                  Dbdx (mp) * (Hz (i-1, j, k) - Hz (i, j, k  ))
+
+      mp = iepz (i, j, k)
+      Ez (i, j, k) = Ez (i, j, k) +                                        &
+                  Dbdx (mp) * (Hy (i, j  , k) - Hy (i-1, j, k)) +          &
+                  Dbdy (mp) * (Hx (i, j-1, k) - Hx (i  , j, k))
+    end do
+  end do
+end do
+
+END SUBROUTINE mat_updateE
+
+END MODULE material_mod
+
+! { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" { target loongarch64*-*-* } } }
-- 
2.43.0