[Bug target/78037] Incorrect code generated at optimization level -O2 for tzcnt and binary and

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78037

--- Comment #9 from Uroš Bizjak <ubizjak at gmail dot com> ---
Created attachment 39845
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=39845&action=edit
Patch that introduces new builtins with UNSPEC_LZCNT and UNSPEC_TZCNT RTX
patterns

This patch introduces new builtins with UNSPEC_LZCNT and UNSPEC_TZCNT RTX
patterns.

The patch fails following tests:

FAIL: gcc.target/i386/bmi-4.c (test for excess errors)
FAIL: gcc.target/i386/bmi-5.c (test for excess errors)
FAIL: gcc.target/i386/bmi-6.c (test for excess errors)

These tests check that:

/* Test that a constant operand 0 to tzcnt gets folded.  */
extern void link_error(void);
int main()
{
  if (__tzcnt_u32(0) != 32)
    link_error();
  return 0;
}

But unpatched gcc simply calls generic __builtin_ctz in the above case. This is
undefined behaviour, so it looks that the test works only by luck...

OTOH, we know that __tzcnt_* and __lzcnt_* builtin functions implement CTZ and
CLZ functionality with defined output for all input arguments. Instead of
simply removing mentioned tests, we can introduce certain constant-folding
functionality to ix86_fold_builtin that would handle constant arguments in a
__builtin_ctz/__builtin_clz way.

The patch is thoroughly tested and works OK (modulo testsuite failures
mentioned above), but I have no idea on how to implement tree-folding stuff. 

So, some kind soul that would take the patch from this point is most welcome.