Thank you for the review Richard!

I re-worked the patch based on your suggestions. I combined the two patterns. Neither one requires a signedness check as long as the type of the 'add' has overflow wrap semantics.

I had to modify the regular expression in no-strict-overflow-4.c test. In that test the following function is compiled with -fno-strict-overflow :

int
foo (int i)
{
  return i + 1 > i;
}

We now optimize this function so that the tree-optimized dump has

;; Function foo (foo, funcdef_no=0, decl_uid=1931, cgraph_uid=1, symbol_order=0)

foo (int i)
{
  _Bool _1;
  int _3;

  <bb 2> [local count: 1073741824]:
  _1 = i_2(D) != 2147483647;
  _3 = (int) _1;
  return _3;
}

This is a correct optimization since -fno-strict-overflow implies -fwrapv.

Eugene

-----Original Message-----
From: Richard Biener <richard.guenther@gmail.com> 
Sent: Tuesday, October 27, 2020 2:23 AM
To: Eugene Rozenfeld <Eugene.Rozenfeld@microsoft.com>
Cc: gcc-patches@gcc.gnu.org
Subject: [EXTERNAL] Re: [PATCH] [tree-optimization] Fix for PR97223

On Sat, Oct 24, 2020 at 2:20 AM Eugene Rozenfeld via Gcc-patches <gcc-patches@gcc.gnu.org> wrote:
>
> This patch adds a pattern for folding
>                 x < (short) ((unsigned short)x + const) to
>          x <= SHORT_MAX - const
> (and similarly for other integral types) if const is not 0.
> as described in PR97223.
>
> For example, without this patch the x86_64-pc-linux code generated for 
> this function
>
> bool f(char x)
> {
>     return x < (char)(x + 12);
> }
>
> is
>
> lea    eax,[rdi+0xc]
> cmp    al,dil
> setg   al
> ret
>
> With the patch the code is
>
> cmp    dil,0x73
> setle  al
> ret
>
> Tested on x86_64-pc-linux.

+/* Similar to the previous pattern but with additional casts. */ (for 
+cmp (lt le ge gt)
+     out (gt gt le le)
+ (simplify
+  (cmp:c (convert@3 (plus@2 (convert@4 @0) INTEGER_CST@1)) @0)
+  (if (!TYPE_UNSIGNED (TREE_TYPE (@0))
+       && types_match (TREE_TYPE (@0), TREE_TYPE (@3))
+       && types_match (TREE_TYPE (@4), unsigned_type_for (TREE_TYPE (@0)))
+       && TYPE_OVERFLOW_WRAPS (TREE_TYPE (@4))
+       && wi::to_wide (@1) != 0
+       && single_use (@2))
+   (with { unsigned int prec = TYPE_PRECISION (TREE_TYPE (@0)); }
+    (out @0 { wide_int_to_tree (TREE_TYPE (@0),
+                               wi::max_value (prec, SIGNED)
+                               - wi::to_wide (@1)); })))))

I think it's reasonable but the comment can be made more precise.
In particular I wonder why we require a signed comparison here while the previous pattern requires an unsigned comparison.  It might be an artifact and the restriction instead only applies to the plus?

Note that

+       && types_match (TREE_TYPE (@4), unsigned_type_for (TREE_TYPE 
+ (@0)))

unsigned_type_for should be avoided since it's quite expensive.  May I suggest

          && TYPE_UNSIGNED (TREE_TYPE (@4))
          && tree_nop_conversion_p (TREE_TYPE (@4), TREE_TYPE (@0))

instead?

I originally wondered if "but with additional casts" could be done in a single pattern via (convert? ...) uses but then I noticed the strange difference in the comparison signedness requirement ...

Richard.

> Eugene
>