From 6684509e81e4341675c73a7dc853180229a8abcb Mon Sep 17 00:00:00 2001
From: Pronin Alexander 00812787 <pronin.alexander@huawei.com>
Date: Tue, 24 Jan 2023 16:43:40 +0300
Subject: [PATCH 04/18] Add option to allow matching uaddsub overflow for widen
 ops too.

---
 gcc/common.opt                 |   5 ++
 gcc/testsuite/gcc.dg/uaddsub.c | 143 +++++++++++++++++++++++++++++++++
 gcc/tree-ssa-math-opts.cc      |  43 ++++++++--
 3 files changed, 184 insertions(+), 7 deletions(-)
 create mode 100644 gcc/testsuite/gcc.dg/uaddsub.c

diff --git a/gcc/common.opt b/gcc/common.opt
index dac477c04..39c90604e 100644
--- a/gcc/common.opt
+++ b/gcc/common.opt
@@ -3106,6 +3106,11 @@ freciprocal-math
 Common Var(flag_reciprocal_math) SetByCombined Optimization
 Same as -fassociative-math for expressions which include division.
 
+fuaddsub-overflow-match-all
+Common Var(flag_uaddsub_overflow_match_all)
+Match unsigned add/sub overflow even if the target does not support
+the corresponding instruction.
+
 ; Nonzero means that unsafe floating-point math optimizations are allowed
 ; for the sake of speed.  IEEE compliance is not guaranteed, and operations
 ; are allowed to assume that their arguments and results are "normal"
diff --git a/gcc/testsuite/gcc.dg/uaddsub.c b/gcc/testsuite/gcc.dg/uaddsub.c
new file mode 100644
index 000000000..96c26d308
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/uaddsub.c
@@ -0,0 +1,143 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -fuaddsub-overflow-match-all -fdump-tree-optimized" } */
+#include <stdint.h>
+
+typedef unsigned __int128 uint128_t;
+typedef struct uint256_t
+{
+  uint128_t lo;
+  uint128_t hi;
+} uint256_t;
+
+uint16_t add16 (uint8_t a, uint8_t b)
+{
+  uint8_t tmp = a + b;
+  uint8_t overflow = 0;
+  if (tmp < a)
+    overflow = 1;
+
+  uint16_t res = overflow;
+  res <<= 8;
+  res += tmp;
+  return res;
+}
+
+uint32_t add32 (uint16_t a, uint16_t b)
+{
+  uint16_t tmp = a + b;
+  uint16_t overflow = 0;
+  if (tmp < a)
+    overflow = 1;
+
+  uint32_t res = overflow;
+  res <<= 16;
+  res += tmp;
+  return res;
+}
+
+uint64_t add64 (uint32_t a, uint32_t b)
+{
+  uint32_t tmp = a + b;
+  uint32_t overflow = 0;
+  if (tmp < a)
+    overflow = 1;
+
+  uint64_t res = overflow;
+  res <<= 32;
+  res += tmp;
+  return res;
+}
+
+uint128_t add128 (uint64_t a, uint64_t b)
+{
+  uint64_t tmp = a + b;
+  uint64_t overflow = 0;
+  if (tmp < a)
+    overflow = 1;
+
+  uint128_t res = overflow;
+  res <<= 64;
+  res += tmp;
+  return res;
+}
+
+uint256_t add256 (uint128_t a, uint128_t b)
+{
+  uint128_t tmp = a + b;
+  uint128_t overflow = 0;
+  if (tmp < a)
+    overflow = 1;
+
+  uint256_t res;
+  res.hi = overflow;
+  res.lo = tmp;
+  return res;
+}
+
+uint16_t sub16 (uint8_t a, uint8_t b)
+{
+  uint8_t tmp = a - b;
+  uint8_t overflow = 0;
+  if (tmp > a)
+    overflow = -1;
+
+  uint16_t res = overflow;
+  res <<= 8;
+  res += tmp;
+  return res;
+}
+
+uint32_t sub32 (uint16_t a, uint16_t b)
+{
+  uint16_t tmp = a - b;
+  uint16_t overflow = 0;
+  if (tmp > a)
+    overflow = -1;
+
+  uint32_t res = overflow;
+  res <<= 16;
+  res += tmp;
+  return res;
+}
+
+uint64_t sub64 (uint32_t a, uint32_t b)
+{
+  uint32_t tmp = a - b;
+  uint32_t overflow = 0;
+  if (tmp > a)
+    overflow = -1;
+
+  uint64_t res = overflow;
+  res <<= 32;
+  res += tmp;
+  return res;
+}
+
+uint128_t sub128 (uint64_t a, uint64_t b)
+{
+  uint64_t tmp = a - b;
+  uint64_t overflow = 0;
+  if (tmp > a)
+    overflow = -1;
+
+  uint128_t res = overflow;
+  res <<= 64;
+  res += tmp;
+  return res;
+}
+
+uint256_t sub256 (uint128_t a, uint128_t b)
+{
+  uint128_t tmp = a - b;
+  uint128_t overflow = 0;
+  if (tmp > a)
+    overflow = -1;
+
+  uint256_t res;
+  res.hi = overflow;
+  res.lo = tmp;
+  return res;
+}
+
+/* { dg-final { scan-tree-dump-times "= .ADD_OVERFLOW \\(a_\[0-9\]+\\(D\\), b_\[0-9\]+\\(D\\)\\)" 5 "optimized" } } */
+/* { dg-final { scan-tree-dump-times "= .SUB_OVERFLOW \\(a_\[0-9\]+\\(D\\), b_\[0-9\]+\\(D\\)\\)" 5 "optimized" } } */
diff --git a/gcc/tree-ssa-math-opts.cc b/gcc/tree-ssa-math-opts.cc
index 232e903b0..55d6ee8ae 100644
--- a/gcc/tree-ssa-math-opts.cc
+++ b/gcc/tree-ssa-math-opts.cc
@@ -3468,6 +3468,27 @@ convert_mult_to_fma (gimple *mul_stmt, tree op1, tree op2,
     }
 }
 
+/* Check if the corresponding operation has wider equivalent on the target.  */
+
+static bool
+wider_optab_check_p (optab op, machine_mode mode, int unsignedp)
+{
+  machine_mode wider_mode;
+  FOR_EACH_WIDER_MODE (wider_mode, mode)
+    {
+      machine_mode next_mode;
+      if (optab_handler (op, wider_mode) != CODE_FOR_nothing
+	  || (op == smul_optab
+	      && GET_MODE_WIDER_MODE (wider_mode).exists (&next_mode)
+	      && (find_widening_optab_handler ((unsignedp
+						? umul_widen_optab
+						: smul_widen_optab),
+						next_mode, mode))))
+	return true;
+    }
+
+  return false;
+}
 
 /* Helper function of match_arith_overflow.  For MUL_OVERFLOW, if we have
    a check for non-zero like:
@@ -3903,15 +3924,22 @@ match_arith_overflow (gimple_stmt_iterator *gsi, gimple *stmt,
 		       || code == MINUS_EXPR
 		       || code == MULT_EXPR
 		       || code == BIT_NOT_EXPR);
+  int unsignedp = TYPE_UNSIGNED (type);
   if (!INTEGRAL_TYPE_P (type)
-      || !TYPE_UNSIGNED (type)
-      || has_zero_uses (lhs)
-      || (code != PLUS_EXPR
-	  && code != MULT_EXPR
-	  && optab_handler (code == MINUS_EXPR ? usubv4_optab : uaddv4_optab,
-			    TYPE_MODE (type)) == CODE_FOR_nothing))
+      || !unsignedp
+      || has_zero_uses (lhs))
     return false;
 
+  if (code == PLUS_EXPR || code == MINUS_EXPR)
+    {
+      machine_mode mode = TYPE_MODE (type);
+      optab op = code == PLUS_EXPR ? uaddv4_optab : usubv4_optab;
+      if (optab_handler (op, mode) == CODE_FOR_nothing
+	  && (!flag_uaddsub_overflow_match_all
+	      || !wider_optab_check_p (op, mode, unsignedp)))
+	return false;
+    }
+
   tree rhs1 = gimple_assign_rhs1 (stmt);
   tree rhs2 = gimple_assign_rhs2 (stmt);
   FOR_EACH_IMM_USE_FAST (use_p, iter, lhs)
@@ -3986,7 +4014,8 @@ match_arith_overflow (gimple_stmt_iterator *gsi, gimple *stmt,
       || (code != MULT_EXPR && (code == BIT_NOT_EXPR ? use_seen : !use_seen))
       || (code == PLUS_EXPR
 	  && optab_handler (uaddv4_optab,
-			    TYPE_MODE (type)) == CODE_FOR_nothing)
+			    TYPE_MODE (type)) == CODE_FOR_nothing
+	  && !flag_uaddsub_overflow_match_all)
       || (code == MULT_EXPR
 	  && optab_handler (cast_stmt ? mulv4_optab : umulv4_optab,
 			    TYPE_MODE (type)) == CODE_FOR_nothing))
-- 
2.33.0