Widening-Mul: Refine build_and_insert_cast when rhs is cast

The widening-mul will insert a cast for the widen-mul, the
function build_and_insert_cast is design to take care of it.

In some case the optimized gimple has some unnecessary cast,
for example as below code.

  #define SAT_U_MUL_FMT_5(NT, WT)                 \
  NT __attribute__((noinline))                    \
  sat_u_mul_##NT##_from_##WT##_fmt_5 (NT a, NT b) \
  {                                               \
    WT x = (WT)a * (WT)b;                         \
    NT hi = x >> (sizeof(NT) * 8);                \
    NT lo = (NT)x;                                \
    return lo | -!!hi;                            \
  }

  SAT_U_MUL_FMT_5(uint64_t, uint128_t)

There will be a additional cast to uint128_t after optimized,
this patch would like to refine this by checking the def of
the rhs cast, if it comes from a cast with less or equal
precision, the rhs of the def will be leveraged.

Before this patch:
  29   │   _1 = (__int128 unsigned) a_8(D);
  30   │   _2 = (__int128 unsigned) b_9(D);
  31   │   _35 = (unsigned long) _1;
  32   │   _34 = (unsigned long) _2;
  33   │   x_10 = _35 w* _34;

After this patch:
  27   │   _35 = (unsigned long) a_8(D);
  28   │   _34 = (unsigned long) b_9(D);
  29   │   x_10 = _35 w* _34;

gcc/ChangeLog:

	* tree-ssa-math-opts.cc (build_and_insert_cast): Refine
	the cast insert by check the rhs of val.

gcc/testsuite/ChangeLog:

	* gcc.target/riscv/sat/widen-mul-0.c: New test.
	* gcc.target/riscv/sat/widen-mul-1.c: New test.
	* gcc.target/riscv/sat/widen-mul-2.c: New test.
	* gcc.target/riscv/sat/widen-mul-3.c: New test.
	* gcc.target/riscv/sat/widen-mul.h: New test.

Signed-off-by: Pan Li <pan2.li@intel.com>

diff --git a/gcc/testsuite/gcc.target/riscv/sat/widen-mul-0.c b/gcc/testsuite/gcc.target/riscv/sat/widen-mul-0.c
new file mode 100644 (file)
index 0000000..1074fa9
--- /dev/null
+++ b/gcc/testsuite/gcc.target/riscv/sat/widen-mul-0.c
@@ -0,0 +1,8 @@
+/* { dg-do compile } */
+/* { dg-options "-march=rv64gc -mabi=lp64d -fdump-tree-optimized" } */
+
+#include "widen-mul.h"
+
+SAT_U_MUL_FMT_5(uint8_t, uint128_t)
+
+/* { dg-final { scan-tree-dump-not " = (__int128 unsigned) " "optimized" } } */

diff --git a/gcc/testsuite/gcc.target/riscv/sat/widen-mul-1.c b/gcc/testsuite/gcc.target/riscv/sat/widen-mul-1.c
new file mode 100644 (file)
index 0000000..5f8f0dd
--- /dev/null
+++ b/gcc/testsuite/gcc.target/riscv/sat/widen-mul-1.c
@@ -0,0 +1,8 @@
+/* { dg-do compile } */
+/* { dg-options "-march=rv64gc -mabi=lp64d -fdump-tree-optimized" } */
+
+#include "widen-mul.h"
+
+SAT_U_MUL_FMT_5(uint16_t, uint128_t)
+
+/* { dg-final { scan-tree-dump-not " = (__int128 unsigned) " "optimized" } } */

diff --git a/gcc/testsuite/gcc.target/riscv/sat/widen-mul-2.c b/gcc/testsuite/gcc.target/riscv/sat/widen-mul-2.c
new file mode 100644 (file)
index 0000000..4c54cc0
--- /dev/null
+++ b/gcc/testsuite/gcc.target/riscv/sat/widen-mul-2.c
@@ -0,0 +1,8 @@
+/* { dg-do compile } */
+/* { dg-options "-march=rv64gc -mabi=lp64d -fdump-tree-optimized" } */
+
+#include "widen-mul.h"
+
+SAT_U_MUL_FMT_5(uint32_t, uint128_t)
+
+/* { dg-final { scan-tree-dump-not " = (__int128 unsigned) " "optimized" } } */

diff --git a/gcc/testsuite/gcc.target/riscv/sat/widen-mul-3.c b/gcc/testsuite/gcc.target/riscv/sat/widen-mul-3.c
new file mode 100644 (file)
index 0000000..d3dd971
--- /dev/null
+++ b/gcc/testsuite/gcc.target/riscv/sat/widen-mul-3.c
@@ -0,0 +1,8 @@
+/* { dg-do compile } */
+/* { dg-options "-march=rv64gc -mabi=lp64d -fdump-tree-optimized" } */
+
+#include "widen-mul.h"
+
+SAT_U_MUL_FMT_5(uint64_t, uint128_t)
+
+/* { dg-final { scan-tree-dump-not " = (__int128 unsigned) " "optimized" } } */

diff --git a/gcc/testsuite/gcc.target/riscv/sat/widen-mul.h b/gcc/testsuite/gcc.target/riscv/sat/widen-mul.h
new file mode 100644 (file)
index 0000000..a0a4408
--- /dev/null
+++ b/gcc/testsuite/gcc.target/riscv/sat/widen-mul.h
@@ -0,0 +1,15 @@
+#include <stdint.h>
+
+#if __riscv_xlen == 64
+typedef unsigned __int128 uint128_t;
+#endif
+
+#define SAT_U_MUL_FMT_5(NT, WT)                 \
+NT __attribute__((noinline))                    \
+sat_u_mul_##NT##_from_##WT##_fmt_5 (NT a, NT b) \
+{                                               \
+  WT x = (WT)a * (WT)b;                         \
+  NT hi = x >> (sizeof(NT) * 8);                \
+  NT lo = (NT)x;                                \
+  return lo | -!!hi;                            \
+}

diff --git a/gcc/tree-ssa-math-opts.cc b/gcc/tree-ssa-math-opts.cc
index bfad4cf5c99dd06000a925b2457cf174d3abda9c..6fafa1b4096352ad81a6dd57703ec0272f90c6ef 100644 (file)
--- a/gcc/tree-ssa-math-opts.cc
+++ b/gcc/tree-ssa-math-opts.cc
@@ -1632,7 +1632,27 @@ build_and_insert_cast (gimple_stmt_iterator *gsi, location_t loc,
                       tree type, tree val)
 {
   tree result = make_ssa_name (type);
-  gassign *stmt = gimple_build_assign (result, NOP_EXPR, val);
+  tree rhs = val;
+
+  if (TREE_CODE (val) == SSA_NAME)
+    {
+      gimple *def = SSA_NAME_DEF_STMT (val);
+
+      if (is_gimple_assign (def)
+         && CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (def)))
+       {
+         tree cast_rhs = gimple_assign_rhs1 (def);
+         unsigned rhs_prec = TYPE_PRECISION (TREE_TYPE (cast_rhs));
+         unsigned type_prec = TYPE_PRECISION (type);
+         unsigned val_prec = TYPE_PRECISION (TREE_TYPE (val));
+
+         if (type_prec >= rhs_prec && val_prec >= rhs_prec)
+           rhs = cast_rhs;
+       }
+    }
+
+  gassign *stmt = gimple_build_assign (result, NOP_EXPR, rhs);
+
   gimple_set_location (stmt, loc);
   gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
   return result;