unroll: Run VN on unrolled-and-jammed loops

Message ID mptv955vv0s.fsf@arm.com
State New
Headers show
Series
  • unroll: Run VN on unrolled-and-jammed loops
Related show

Commit Message

Michael Meissner via Gcc-patches July 20, 2021, 3:56 p.m.
Unroll and jam can sometimes leave redundancies.  E.g. for:

  for (int j = 0; j < 100; ++j)
    for (int i = 0; i < 100; ++i)
      x[i] += y[i] * z[j][i];

the new loop will do the equivalent of:

  for (int j = 0; j < 100; j += 2)
    for (int i = 0; i < 100; ++i)
      {
        x[i] += y[i] * z[j][i];
        x[i] += y[i] * z[j + 1][i];
      }

with two reads of y[i] and with a round trip through memory for x[i].

At the moment these redundancies survive till vectorisation, so if
vectorisation succeeds, we're reliant on being able to remove the
redundancies from the vector form.  This can be hard to do if
a vector loop uses predication.  E.g. on SVE we end up with:

.L3:
        ld1w    z3.s, p0/z, [x3, x0, lsl 2]
        ld1w    z0.s, p0/z, [x5, x0, lsl 2]
        ld1w    z1.s, p0/z, [x2, x0, lsl 2]
        mad     z1.s, p1/m, z0.s, z3.s
        ld1w    z2.s, p0/z, [x4, x0, lsl 2]
        st1w    z1.s, p0, [x3, x0, lsl 2]    // store to x[i]
        ld1w    z1.s, p0/z, [x3, x0, lsl 2]  // load back from x[i]
        mad     z0.s, p1/m, z2.s, z1.s
        st1w    z0.s, p0, [x3, x0, lsl 2]
        add     x0, x0, x6
        whilelo p0.s, w0, w1
        b.any   .L3

This patch runs a value-numbering pass on loops after a successful
unroll-and-jam, which gets rid of the unnecessary load and gives
a more accurate idea of vector costs.  Unfortunately the redundant
store still persists without a pre-vect DSE, but that feels like
a separate issue.

Note that the pass requires the loop to have a single exit,
hence the simple calculation of exit_bbs.

Tested on aarch64-linux-gnu so far, will test on x86_64-linux-gnu too.
OK to install if testing passes?

Richard


gcc/
	* gimple-loop-jam.c: Include tree-ssa-sccvn.h.
	(tree_loop_unroll_and_jam): Run value-numbering on a loop that
	has been successfully unrolled.

gcc/testsuite/
	* gcc.dg/unroll-10.c: New test.
---
 gcc/gimple-loop-jam.c            | 14 +++++++++-----
 gcc/testsuite/gcc.dg/unroll-10.c | 13 +++++++++++++
 2 files changed, 22 insertions(+), 5 deletions(-)
 create mode 100644 gcc/testsuite/gcc.dg/unroll-10.c

Comments

Michael Meissner via Gcc-patches July 20, 2021, 5:59 p.m. | #1
On July 20, 2021 5:56:35 PM GMT+02:00, Richard Sandiford via Gcc-patches <gcc-patches@gcc.gnu.org> wrote:
>Unroll and jam can sometimes leave redundancies.  E.g. for:

>

>  for (int j = 0; j < 100; ++j)

>    for (int i = 0; i < 100; ++i)

>      x[i] += y[i] * z[j][i];

>

>the new loop will do the equivalent of:

>

>  for (int j = 0; j < 100; j += 2)

>    for (int i = 0; i < 100; ++i)

>      {

>        x[i] += y[i] * z[j][i];

>        x[i] += y[i] * z[j + 1][i];

>      }

>

>with two reads of y[i] and with a round trip through memory for x[i].

>

>At the moment these redundancies survive till vectorisation, so if

>vectorisation succeeds, we're reliant on being able to remove the

>redundancies from the vector form.  This can be hard to do if

>a vector loop uses predication.  E.g. on SVE we end up with:

>

>.L3:

>        ld1w    z3.s, p0/z, [x3, x0, lsl 2]

>        ld1w    z0.s, p0/z, [x5, x0, lsl 2]

>        ld1w    z1.s, p0/z, [x2, x0, lsl 2]

>        mad     z1.s, p1/m, z0.s, z3.s

>        ld1w    z2.s, p0/z, [x4, x0, lsl 2]

>        st1w    z1.s, p0, [x3, x0, lsl 2]    // store to x[i]

>        ld1w    z1.s, p0/z, [x3, x0, lsl 2]  // load back from x[i]

>        mad     z0.s, p1/m, z2.s, z1.s

>        st1w    z0.s, p0, [x3, x0, lsl 2]

>        add     x0, x0, x6

>        whilelo p0.s, w0, w1

>        b.any   .L3

>

>This patch runs a value-numbering pass on loops after a successful

>unroll-and-jam, which gets rid of the unnecessary load and gives

>a more accurate idea of vector costs.  Unfortunately the redundant

>store still persists without a pre-vect DSE, but that feels like

>a separate issue.

>

>Note that the pass requires the loop to have a single exit,

>hence the simple calculation of exit_bbs.

>

>Tested on aarch64-linux-gnu so far, will test on x86_64-linux-gnu too.

>OK to install if testing passes?


Ok. 

Thanks, 
Richard. 

>Richard

>

>

>gcc/

>	* gimple-loop-jam.c: Include tree-ssa-sccvn.h.

>	(tree_loop_unroll_and_jam): Run value-numbering on a loop that

>	has been successfully unrolled.

>

>gcc/testsuite/

>	* gcc.dg/unroll-10.c: New test.

>---

> gcc/gimple-loop-jam.c            | 14 +++++++++-----

> gcc/testsuite/gcc.dg/unroll-10.c | 13 +++++++++++++

> 2 files changed, 22 insertions(+), 5 deletions(-)

> create mode 100644 gcc/testsuite/gcc.dg/unroll-10.c

>

>diff --git a/gcc/gimple-loop-jam.c b/gcc/gimple-loop-jam.c

>index 4842f0dff80..544ad779dd6 100644

>--- a/gcc/gimple-loop-jam.c

>+++ b/gcc/gimple-loop-jam.c

>@@ -38,6 +38,7 @@ along with GCC; see the file COPYING3.  If not see

> #include "tree-data-ref.h"

> #include "tree-ssa-loop-ivopts.h"

> #include "tree-vectorizer.h"

>+#include "tree-ssa-sccvn.h"

> 

> /* Unroll and Jam transformation

>    

>@@ -487,7 +488,7 @@ static unsigned int

> tree_loop_unroll_and_jam (void)

> {

>   class loop *loop;

>-  bool changed = false;

>+  unsigned int todo = 0;

> 

>   gcc_assert (scev_initialized_p ());

> 

>@@ -591,7 +592,11 @@ tree_loop_unroll_and_jam (void)

> 			    &desc);

> 	  free_original_copy_tables ();

> 	  fuse_loops (outer->inner);

>-	  changed = true;

>+	  todo |= TODO_cleanup_cfg;

>+

>+	  auto_bitmap exit_bbs;

>+	  bitmap_set_bit (exit_bbs, single_dom_exit (outer)->dest->index);

>+	  todo |= do_rpo_vn (cfun, loop_preheader_edge (outer), exit_bbs);

> 	}

> 

>       loop_nest.release ();

>@@ -599,13 +604,12 @@ tree_loop_unroll_and_jam (void)

>       free_data_refs (datarefs);

>     }

> 

>-  if (changed)

>+  if (todo)

>     {

>       scev_reset ();

>       free_dominance_info (CDI_DOMINATORS);

>-      return TODO_cleanup_cfg;

>     }

>-  return 0;

>+  return todo;

> }

> 

> /* Pass boilerplate */

>diff --git a/gcc/testsuite/gcc.dg/unroll-10.c

>b/gcc/testsuite/gcc.dg/unroll-10.c

>new file mode 100644

>index 00000000000..0559915f2fc

>--- /dev/null

>+++ b/gcc/testsuite/gcc.dg/unroll-10.c

>@@ -0,0 +1,13 @@

>+/* { dg-options "-O3 -fdump-tree-unrolljam" } */

>+

>+void

>+f (int *restrict x, int *restrict y, int z[restrict 100][100])

>+{

>+  for (int j = 0; j < 100; ++j)

>+    for (int i = 0; i < 100; ++i)

>+      x[i] += y[i] * z[j][i];

>+}

>+

>+/* The loop should be unrolled 2 times, leaving one load from x,

>+   one load from y and 2 loads from z.  */

>+/* { dg-final { scan-tree-dump-times { = \(*\*} 4 "unrolljam" } } */

Patch

diff --git a/gcc/gimple-loop-jam.c b/gcc/gimple-loop-jam.c
index 4842f0dff80..544ad779dd6 100644
--- a/gcc/gimple-loop-jam.c
+++ b/gcc/gimple-loop-jam.c
@@ -38,6 +38,7 @@  along with GCC; see the file COPYING3.  If not see
 #include "tree-data-ref.h"
 #include "tree-ssa-loop-ivopts.h"
 #include "tree-vectorizer.h"
+#include "tree-ssa-sccvn.h"
 
 /* Unroll and Jam transformation
    
@@ -487,7 +488,7 @@  static unsigned int
 tree_loop_unroll_and_jam (void)
 {
   class loop *loop;
-  bool changed = false;
+  unsigned int todo = 0;
 
   gcc_assert (scev_initialized_p ());
 
@@ -591,7 +592,11 @@  tree_loop_unroll_and_jam (void)
 			    &desc);
 	  free_original_copy_tables ();
 	  fuse_loops (outer->inner);
-	  changed = true;
+	  todo |= TODO_cleanup_cfg;
+
+	  auto_bitmap exit_bbs;
+	  bitmap_set_bit (exit_bbs, single_dom_exit (outer)->dest->index);
+	  todo |= do_rpo_vn (cfun, loop_preheader_edge (outer), exit_bbs);
 	}
 
       loop_nest.release ();
@@ -599,13 +604,12 @@  tree_loop_unroll_and_jam (void)
       free_data_refs (datarefs);
     }
 
-  if (changed)
+  if (todo)
     {
       scev_reset ();
       free_dominance_info (CDI_DOMINATORS);
-      return TODO_cleanup_cfg;
     }
-  return 0;
+  return todo;
 }
 
 /* Pass boilerplate */
diff --git a/gcc/testsuite/gcc.dg/unroll-10.c b/gcc/testsuite/gcc.dg/unroll-10.c
new file mode 100644
index 00000000000..0559915f2fc
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/unroll-10.c
@@ -0,0 +1,13 @@ 
+/* { dg-options "-O3 -fdump-tree-unrolljam" } */
+
+void
+f (int *restrict x, int *restrict y, int z[restrict 100][100])
+{
+  for (int j = 0; j < 100; ++j)
+    for (int i = 0; i < 100; ++i)
+      x[i] += y[i] * z[j][i];
+}
+
+/* The loop should be unrolled 2 times, leaving one load from x,
+   one load from y and 2 loads from z.  */
+/* { dg-final { scan-tree-dump-times { = \(*\*} 4 "unrolljam" } } */