openmp: Adjust outer bounds of non-rect loops

In loops like:
  #pragma omp parallel for collapse(2)
  for (i = -4; i < 8; i++)
    for (j = 3 * i; j > 2 * i; j--)
for some outer loop iterations there are no inner loop iterations at all,
the condition is false.  In order to use Summæ Potestate to count number
of iterations or to transform the logical iteration number to actual
iterator values using quadratic non-equation root discovery the outer
iterator range needs to be adjusted, such that the inner loop has at least
one iteration for each of the outer loop iterator value in the reduced
range.  Sometimes this adjustment is done at the start of the range,
at other times at the end.

This patch implements it during the compile time number of loop computation
(if all expressions are compile time constants).

2020-07-14  Jakub Jelinek  <jakub@redhat.com>

	* omp-general.h (struct omp_for_data): Add adjn1 member.
	* omp-general.c (omp_extract_for_data): For non-rect loop, punt on
	count computing if n1, n2 or step are not INTEGER_CST earlier.
	Narrow the outer iterator range if needed so that non-rect loop
	has at least one iteration for each outer range iteration.  Compute
	adjn1.
	* omp-expand.c (expand_omp_for_init_vars): Use adjn1 if non-NULL
	instead of the outer loop's n1.

	* testsuite/libgomp.c/loop-21.c: New test.

4 files changed, 360 insertions, 4 deletions

diff --git a/gcc/omp-expand.c b/gcc/omp-expand.c
index c3b8820..a721940 100644
--- a/gcc/omp-expand.c
+++ b/gcc/omp-expand.c
@@ -2262,6 +2262,7 @@ expand_omp_for_init_vars (struct omp_for_data *fd, gimple_stmt_iterator *gsi,
 	      && (optab_handler (sqrt_optab, TYPE_MODE (double_type_node))
 		  != CODE_FOR_nothing))
 	    {
+	      tree outer_n1 = fd->adjn1 ? fd->adjn1 : fd->loops[i - 1].n1;
 	      tree itype = TREE_TYPE (fd->loops[i].v);
 	      tree min_inner_iterations = fd->min_inner_iterations;
 	      tree factor = fd->factor;
@@ -2384,7 +2385,7 @@ expand_omp_for_init_vars (struct omp_for_data *fd, gimple_stmt_iterator *gsi,
 	      *gsi = gsi_after_labels (e->dest);
 	      t = fold_convert (itype, c);
 	      t = fold_build2 (MULT_EXPR, itype, t, fd->loops[i - 1].step);
-	      t = fold_build2 (PLUS_EXPR, itype, fd->loops[i - 1].n1, t);
+	      t = fold_build2 (PLUS_EXPR, itype, outer_n1, t);
 	      t = force_gimple_operand_gsi (gsi, t, true, NULL_TREE, false,
 					    GSI_CONTINUE_LINKING);
 	      expand_omp_build_assign (gsi, fd->loops[i - 1].v, t, true);
diff --git a/gcc/omp-general.c b/gcc/omp-general.c
index c6878cf..b2ce408 100644
--- a/gcc/omp-general.c
+++ b/gcc/omp-general.c
@@ -214,6 +214,7 @@ omp_extract_for_data (gomp_for *for_stmt, struct omp_for_data *fd,
   fd->simd_schedule = false;
   fd->min_inner_iterations = NULL_TREE;
   fd->factor = NULL_TREE;
+  fd->adjn1 = NULL_TREE;
   collapse_iter = NULL;
   collapse_count = NULL;
 
@@ -508,7 +509,10 @@ omp_extract_for_data (gomp_for *for_stmt, struct omp_for_data *fd,
 		continue;
 	      if (single_nonrect == -1
 		  || (loop->m1 && TREE_CODE (loop->m1) != INTEGER_CST)
-		  || (loop->m2 && TREE_CODE (loop->m2) != INTEGER_CST))
+		  || (loop->m2 && TREE_CODE (loop->m2) != INTEGER_CST)
+		  || TREE_CODE (loop->n1) != INTEGER_CST
+		  || TREE_CODE (loop->n2) != INTEGER_CST
+		  || TREE_CODE (loop->step) != INTEGER_CST)
 		{
 		  count = NULL_TREE;
 		  continue;
@@ -574,12 +578,129 @@ omp_extract_for_data (gomp_for *for_stmt, struct omp_for_data *fd,
 	      else if (t && t2 && integer_zerop (t) && integer_zerop (t2))
 		/* No iterations of the inner loop.  count will be set to
 		   zero cst below.  */;
-	      else
+	      else if (TYPE_UNSIGNED (itype)
+		       || t == NULL_TREE
+		       || t2 == NULL_TREE
+		       || TREE_CODE (t) != INTEGER_CST
+		       || TREE_CODE (t2) != INTEGER_CST)
 		{
 		  /* Punt (for now).  */
 		  count = NULL_TREE;
 		  continue;
 		}
+	      else
+		{
+		  /* Some iterations of the outer loop have zero iterations
+		     of the inner loop, while others have at least one.
+		     In this case, we need to adjust one of those outer
+		     loop bounds.  If ADJ_FIRST, we need to adjust outer n1
+		     (first), otherwise outer n2 (last).  */
+		  bool adj_first = integer_zerop (t);
+		  tree n1 = fold_convert (itype, loop->n1);
+		  tree n2 = fold_convert (itype, loop->n2);
+		  tree m1 = loop->m1 ? fold_convert (itype, loop->m1)
+				     : build_zero_cst (itype);
+		  tree m2 = loop->m2 ? fold_convert (itype, loop->m2)
+				     : build_zero_cst (itype);
+		  t = fold_binary (MINUS_EXPR, itype, n1, n2);
+		  t2 = fold_binary (MINUS_EXPR, itype, m2, m1);
+		  t = fold_binary (TRUNC_DIV_EXPR, itype, t, t2);
+		  t2 = fold_binary (MINUS_EXPR, itype, t, first);
+		  t2 = fold_binary (TRUNC_MOD_EXPR, itype, t2, ostep);
+		  t = fold_binary (MINUS_EXPR, itype, t, t2);
+		  tree n1cur
+		    = fold_binary (PLUS_EXPR, itype, n1,
+				   fold_binary (MULT_EXPR, itype, m1, t));
+		  tree n2cur
+		    = fold_binary (PLUS_EXPR, itype, n2,
+				   fold_binary (MULT_EXPR, itype, m2, t));
+		  t2 = fold_binary (loop->cond_code, boolean_type_node,
+				    n1cur, n2cur);
+		  tree t3 = fold_binary (MULT_EXPR, itype, m1, ostep);
+		  tree t4 = fold_binary (MULT_EXPR, itype, m2, ostep);
+		  tree diff;
+		  if (adj_first)
+		    {
+		      tree new_first;
+		      if (integer_nonzerop (t2))
+			{
+			  new_first = t;
+			  n1first = n1cur;
+			  n2first = n2cur;
+			  if (flag_checking)
+			    {
+			      t3 = fold_binary (MINUS_EXPR, itype, n1cur, t3);
+			      t4 = fold_binary (MINUS_EXPR, itype, n2cur, t4);
+			      t3 = fold_binary (loop->cond_code,
+						boolean_type_node, t3, t4);
+			      gcc_assert (integer_zerop (t3));
+			    }
+			}
+		      else
+			{
+			  t3 = fold_binary (PLUS_EXPR, itype, n1cur, t3);
+			  t4 = fold_binary (PLUS_EXPR, itype, n2cur, t4);
+			  new_first = fold_binary (PLUS_EXPR, itype, t, ostep);
+			  n1first = t3;
+			  n2first = t4;
+			  if (flag_checking)
+			    {
+			      t3 = fold_binary (loop->cond_code,
+						boolean_type_node, t3, t4);
+			      gcc_assert (integer_nonzerop (t3));
+			    }
+			}
+		      diff = fold_binary (MINUS_EXPR, itype, new_first, first);
+		      first = new_first;
+		      fd->adjn1 = first;
+		    }
+		  else
+		    {
+		      tree new_last;
+		      if (integer_zerop (t2))
+			{
+			  t3 = fold_binary (MINUS_EXPR, itype, n1cur, t3);
+			  t4 = fold_binary (MINUS_EXPR, itype, n2cur, t4);
+			  new_last = fold_binary (MINUS_EXPR, itype, t, ostep);
+			  n1last = t3;
+			  n2last = t4;
+			  if (flag_checking)
+			    {
+			      t3 = fold_binary (loop->cond_code,
+						boolean_type_node, t3, t4);
+			      gcc_assert (integer_nonzerop (t3));
+			    }
+			}
+		      else
+			{
+			  new_last = t;
+			  n1last = n1cur;
+			  n2last = n2cur;
+			  if (flag_checking)
+			    {
+			      t3 = fold_binary (PLUS_EXPR, itype, n1cur, t3);
+			      t4 = fold_binary (PLUS_EXPR, itype, n2cur, t4);
+			      t3 = fold_binary (loop->cond_code,
+						boolean_type_node, t3, t4);
+			      gcc_assert (integer_zerop (t3));
+			    }
+			}
+		      diff = fold_binary (MINUS_EXPR, itype, last, new_last);
+		    }
+		  if (TYPE_UNSIGNED (itype)
+		      && single_nonrect_cond_code == GT_EXPR)
+		    diff = fold_binary (TRUNC_DIV_EXPR, itype,
+					fold_unary (NEGATE_EXPR, itype, diff),
+					fold_unary (NEGATE_EXPR, itype,
+						    ostep));
+		  else
+		    diff = fold_binary (TRUNC_DIV_EXPR, itype, diff, ostep);
+		  diff = fold_convert (long_long_unsigned_type_node, diff);
+		  single_nonrect_count
+		    = fold_binary (MINUS_EXPR, long_long_unsigned_type_node,
+				   single_nonrect_count, diff);
+		  t = NULL_TREE;
+		}
 	    }
 	  else
 	    t = fold_binary (loop->cond_code, boolean_type_node,
@@ -715,10 +836,11 @@ omp_extract_for_data (gomp_for *for_stmt, struct omp_for_data *fd,
 	  *collapse_count = fold_convert_loc (loc, iter_type, count);
 	  if (fd->min_inner_iterations && fd->factor)
 	    {
-	      t = make_tree_vec (3);
+	      t = make_tree_vec (4);
 	      TREE_VEC_ELT (t, 0) = *collapse_count;
 	      TREE_VEC_ELT (t, 1) = fd->min_inner_iterations;
 	      TREE_VEC_ELT (t, 2) = fd->factor;
+	      TREE_VEC_ELT (t, 3) = fd->adjn1;
 	      *collapse_count = t;
 	    }
 	}
@@ -736,6 +858,7 @@ omp_extract_for_data (gomp_for *for_stmt, struct omp_for_data *fd,
 	  gcc_assert (fd->non_rect);
 	  fd->min_inner_iterations = TREE_VEC_ELT (fd->loop.n2, 1);
 	  fd->factor = TREE_VEC_ELT (fd->loop.n2, 2);
+	  fd->adjn1 = TREE_VEC_ELT (fd->loop.n2, 3);
 	  fd->loop.n2 = TREE_VEC_ELT (fd->loop.n2, 0);
 	}
       fd->loop.step = build_int_cst (TREE_TYPE (fd->loop.v), 1);
diff --git a/gcc/omp-general.h b/gcc/omp-general.h
index ec0f2a4..2da4d14 100644
--- a/gcc/omp-general.h
+++ b/gcc/omp-general.h
@@ -85,6 +85,8 @@ struct omp_for_data
 				outer iterator, depending on which
 				results in fewer iterations.  */
   tree factor; /* (m2 - m1) * outer_step / inner_step.  */
+  /* Adjusted n1 of the outer loop in such loop nests (if needed).  */
+  tree adjn1;
 };
 
 #define OACC_FN_ATTRIB "oacc function"
diff --git a/libgomp/testsuite/libgomp.c/loop-21.c b/libgomp/testsuite/libgomp.c/loop-21.c
new file mode 100644
index 0000000..1baf13d
--- /dev/null
+++ b/libgomp/testsuite/libgomp.c/loop-21.c
@@ -0,0 +1,230 @@
+/* { dg-do run } */
+
+extern void abort (void);
+
+int x, i, j;
+volatile int a, b, c, d, e, f, g, h;
+int k[13][27];
+
+int
+main ()
+{
+  int niters;
+  for (i = -4; i < 8; i++)
+    for (j = 3 * i; j > 2 * i; j--)
+      k[i + 5][j + 5] = 1;
+  a = -4; b = 8; c = 1; d = 3; e = 0; f = 2; g = 0; h = -1;
+  niters = 0; i = -100; j = -100; x = -100;
+  #pragma omp parallel for collapse(2) lastprivate (i, j, x) reduction(+:niters)
+  for (i = -4; i < 8; i++)
+    for (j = 3 * i; j > 2 * i; j--)
+      {
+	if (i < -4 || i >= 8 || j > 3 * i || j <= i * 2 || k[i + 5][j + 5] != 1)
+	  abort ();
+	k[i + 5][j + 5]++;
+	x = i * 1024 + (j & 1023);
+	niters++;
+      }
+  if (i != 8 || j != 14 || x != 7183 || niters != 28)
+    abort ();
+  niters = 0; i = -100; j = -100; x = -100;
+  #pragma omp parallel for collapse(2) lastprivate (i, j, x) reduction(+:niters)
+  for (i = a; i < b; i += c)
+    for (j = d * i + e; j > g + i * f; j += h)
+      {
+	if (i < -4 || i >= 8 || j > 3 * i || j <= i * 2 || k[i + 5][j + 5] != 2)
+	  abort ();
+	k[i + 5][j + 5]++;
+	x = i * 1024 + (j & 1023);
+	niters++;
+      }
+  if (i != 8 || j != 14 || x != 7183 || niters != 28)
+    abort ();
+  for (int i = -4; i < 8; i++)
+    for (int j = 3 * i; j > 2 * i; j--)
+      if (k[i + 5][j + 5] == 3)
+	k[i + 5][j + 5] = 0;
+      else
+	abort ();
+  for (i = -2; i < 4; i++)
+    for (j = -2 * i + 3; j > -3; j -= 2)
+      k[i + 5][j + 5] = 1;
+  a = -2; b = 4; c = 1; d = -2; e = 3; f = 0; g = -3; h = -2;
+  niters = 0; i = -100; j = -100; x = -100;
+  #pragma omp parallel for collapse(2) lastprivate (i, j, x) reduction(+:niters)
+  for (i = -2; i < 4; i++)
+    for (j = -2 * i + 3; j > -3; j -= 2)
+      {
+	if (i < -2 || i >= 4 || j <= -3 || j > -2 * i + 3 || k[i + 5][j + 5] != 1)
+	  abort ();
+	k[i + 5][j + 5]++;
+	x = i * 1024 + (j & 1023);
+	niters++;
+      }
+  if (/* i != 4 || j != -3 || */x != 3071 || niters != 15)
+    abort ();
+  niters = 0; i = -100; j = -100; x = -100;
+  #pragma omp parallel for collapse(2) lastprivate (i, j, x) reduction(+:niters)
+  for (i = a; i < b; i += c)
+    for (j = d * i + e; j > g + i * f; j += h)
+      {
+	if (i < -2 || i >= 4 || j <= -3 || j > -2 * i + 3 || k[i + 5][j + 5] != 2)
+	  abort ();
+	k[i + 5][j + 5]++;
+	x = i * 1024 + (j & 1023);
+	niters++;
+      }
+  if (/*i != 4 || j != -3 || */x != 3071 || niters != 15)
+    abort ();
+  for (i = -2; i < 4; i++)
+    for (j = -2 * i + 3; j > -3; j -= 2)
+      if (k[i + 5][j + 5] == 3)
+	k[i + 5][j + 5] = 0;
+      else
+	abort ();
+  for (i = 3; i > -3; i--)
+    for (j = -2 * i + 7; j > 2 * i + 1; j--)
+      k[i + 5][j + 5] = 1;
+  a = 3; b = -3; c = -1; d = -2; e = 7; f = 2; g = 1; h = -1;
+  niters = 0; i = -100; j = -100; x = -100;
+  #pragma omp parallel for collapse(2) lastprivate (i, j, x) reduction(+:niters)
+  for (i = 3; i > -3; i--)
+    for (j = -2 * i + 7; j > 2 * i + 1; j--)
+      {
+	if (i <= -3 || i > 3 || j <= 2 * i + 1 || j > -2 * i + 7 || k[i + 5][j + 5] != 1)
+	  abort ();
+	k[i + 5][j + 5]++;
+	x = i * 1024 + (j & 1023);
+	niters++;
+      }
+  if (i != -3 || j != -3 || x != -1026 || niters != 32)
+    abort ();
+  niters = 0; i = -100; j = -100; x = -100;
+  #pragma omp parallel for collapse(2) lastprivate (i, j, x) reduction(+:niters)
+  for (i = a; i > b; i += c)
+    for (j = d * i + e; j > g + i * f; j += h)
+      {
+	if (i <= -3 || i > 3 || j <= 2 * i + 1 || j > -2 * i + 7 || k[i + 5][j + 5] != 2)
+	  abort ();
+	k[i + 5][j + 5]++;
+	x = i * 1024 + (j & 1023);
+	niters++;
+      }
+  if (i != -3 || j != -3 || x != -1026 || niters != 32)
+    abort ();
+  for (i = 3; i > -3; i--)
+    for (j = -2 * i + 7; j > 2 * i + 1; j--)
+      if (k[i + 5][j + 5] == 3)
+	k[i + 5][j + 5] = 0;
+      else
+	abort ();
+  for (i = 3; i > -3; i--)
+    for (j = 2 * i + 7; j > -2 * i + 1; j--)
+      k[i + 5][j + 5] = 1;
+  a = 3; b = -3; c = -1; d = 2; e = 7; f = -2; g = 1; h = -1;
+  niters = 0; i = -100; j = -100; x = -100;
+  #pragma omp parallel for collapse(2) lastprivate (i, j, x) reduction(+:niters)
+  for (i = 3; i > -3; i--)
+    for (j = 2 * i + 7; j > -2 * i + 1; j--)
+      {
+	if (i <= -3 || i > 3 || j <= -2 * i + 1 || j > 2 * i + 7 || k[i + 5][j + 5] != 1)
+	  abort ();
+	k[i + 5][j + 5]++;
+	x = i * 1024 + (j & 1023);
+	niters++;
+      }
+  if (/*i != -3 || j != 3 || */x != -1020 || niters != 50)
+    abort ();
+  niters = 0; i = -100; j = -100; x = -100;
+  #pragma omp parallel for collapse(2) lastprivate (i, j, x) reduction(+:niters)
+  for (i = a; i > b; i += c)
+    for (j = d * i + e; j > g + i * f; j += h)
+      {
+	if (i <= -3 || i > 3 || j <= -2 * i + 1 || j > 2 * i + 7 || k[i + 5][j + 5] != 2)
+	  abort ();
+	k[i + 5][j + 5]++;
+	x = i * 1024 + (j & 1023);
+	niters++;
+      }
+  if (/*i != -3 || j != 3 || */x != -1020 || niters != 50)
+    abort ();
+  for (i = 3; i > -3; i--)
+    for (j = 2 * i + 7; j > -2 * i + 1; j--)
+      if (k[i + 5][j + 5] == 3)
+	k[i + 5][j + 5] = 0;
+      else
+	abort ();
+  for (i = 6; i > -6; i--)
+    for (j = 2 * i + 7; j <= -2 * i + 1; j++)
+      k[i + 5][j + 5] = 1;
+  a = 6; b = -6; c = -1; d = 2; e = 7; f = -2; g = 2; h = 1;
+  niters = 0; i = -100; j = -100; x = -100;
+  #pragma omp parallel for collapse(2) lastprivate (i, j, x) reduction(+:niters)
+  for (i = 6; i > -6; i--)
+    for (j = 2 * i + 7; j <= -2 * i + 1; j++)
+      {
+	if (i <= -6 || i > 6 || j < 2 * i + 7 || j >= -2 * i + 2 || k[i + 5][j + 5] != 1)
+	  abort ();
+	k[i + 5][j + 5]++;
+	x = i * 1024 + (j & 1023);
+	niters++;
+      }
+  if (i != -6 || j != 12 || x != -5109 || niters != 36)
+    abort ();
+  niters = 0; i = -100; j = -100; x = -100;
+  #pragma omp parallel for collapse(2) lastprivate (i, j, x) reduction(+:niters)
+  for (i = a; i > b; i += c)
+    for (j = d * i + e; j < g + i * f; j += h)
+      {
+	if (i <= -6 || i > 6 || j < 2 * i + 7 || j >= -2 * i + 2 || k[i + 5][j + 5] != 2)
+	  abort ();
+	k[i + 5][j + 5]++;
+	x = i * 1024 + (j & 1023);
+	niters++;
+      }
+  if (i != -6 || j != 12 || x != -5109 || niters != 36)
+    abort ();
+  for (i = 6; i > -6; i--)
+    for (j = 2 * i + 7; j <= -2 * i + 1; j++)
+      if (k[i + 5][j + 5] == 3)
+	k[i + 5][j + 5] = 0;
+      else
+	abort ();
+  for (i = 6; i > -6; i -= 2)
+    for (j = -2 * i + 7; j <= 2 * i + 1; j++)
+      k[i + 5][j + 5] = 1;
+  a = 6; b = -6; c = -2; d = -2; e = 7; f = 2; g = 2; h = 1;
+  niters = 0; i = -100; j = -100; x = -100;
+  #pragma omp parallel for collapse(2) lastprivate (i, j, x) reduction(+:niters)
+  for (i = 6; i > -6; i -= 2)
+    for (j = -2 * i + 7; j <= 2 * i + 1; j++)
+      {
+	if (i <= -6 || i > 6 || j < -2 * i + 7 || j >= 2 * i + 2 || k[i + 5][j + 5] != 1)
+	  abort ();
+	k[i + 5][j + 5]++;
+	x = i * 1024 + (j & 1023);
+	niters++;
+      }
+  if (/*i != -6 || j != 15 || */x != 2053 || niters != 33)
+    abort ();
+  niters = 0; i = -100; j = -100; x = -100;
+  #pragma omp parallel for collapse(2) lastprivate (i, j, x) reduction(+:niters)
+  for (i = a; i > b; i += c)
+    for (j = d * i + e; j < g + i * f; j += h)
+      {
+	if (i <= -6 || i > 6 || j < -2 * i + 7 || j >= 2 * i + 2 || k[i + 5][j + 5] != 2)
+	  abort ();
+	k[i + 5][j + 5]++;
+	x = i * 1024 + (j & 1023);
+	niters++;
+      }
+  if (/*i != -6 || j != 15 || */x != 2053 || niters != 33)
+    abort ();
+  for (i = 6; i > -6; i -= 2)
+    for (j = -2 * i + 7; j <= 2 * i + 1; j++)
+      if (k[i + 5][j + 5] == 3)
+	k[i + 5][j + 5] = 0;
+      else
+	abort ();
+  return 0;
+}