Incorrect fold unsigned icmp in glibc/wcsmbs/mbrtowc.c

[wzssyqa]

Since commit

commit ae73bc8e94b3174e2b495eee544cf843b6e51899                                                                                                             
Author: Nikita Popov <npopov@redhat.com>                                                                                                                    
Date:   Tue Dec 3 14:28:56 2024 +0100                                         
                                                                                                                                                            
    Reapply [InstCombine] Support gep nuw in icmp folds (#118472)             
                                                                                                                                                            
    The profile runtime test failure this caused has been addressed in:       
    https://github.com/llvm/llvm-project/pull/118782                                                                                                        
                                                                              
    -----                                                                     
                                                                              
    Unsigned icmp of gep nuw folds to unsigned icmp of offsets. Unsigned       
    icmp of gep nusw nuw folds to unsigned samesign icmp of offsets.
                                                                              
    Proofs: https://alive2.llvm.org/ce/z/VEwQY8                               
                                                                              
 .../Transforms/InstCombine/InstCombineCompares.cpp | 36 ++++++++---          
 llvm/test/Transforms/InstCombine/icmp-gep.ll       | 75 ++++++++++++++++++++++                                                                             
 2 files changed, 102 insertions(+), 9 deletions(-)

I meet a regression problem when build glibc with clang
In glibc/wcsmbs/mbrtowc.c, the follow code

  inbuf = (const unsigned char *) s;
  endbuf = inbuf + n;
  if (__glibc_unlikely (endbuf < inbuf))
    {
      endbuf = (const unsigned char *) ~(uintptr_t) 0;
      if (endbuf == inbuf)
        goto ilseq;
    }

is folded.

Howto reproduce:

1. git clone glibc and checkout branch azanella/clang
2. Build glibc with

 CC=/build/llvm-project/build/bin/clang CXX=/build/llvm-project/build/bin/clang++ CFLAGS="-g -O2 -Wno-unused-command-line-argument" CXXFLAGS="-g -O2 -Wno-unused-command-line-argument" ../configure --host=riscv64-linux-gnu --prefix=/usr
make

[EugeneZelenko]

@wzssyqa: Could you please elaborate?

[wzssyqa]

The problem is that is

char *a, *b;
size_t n;
b = a + n;
if (b < a)

different with

char *a, *b;
size_t n;
b = a + n;
if ((a + n) < a)

[wzssyqa]

I think that it is a bug, as we need a method to determine whether the add overflowed.
I have a try with this code even:

unsigned char* f(unsigned char *a, unsigned long int n) {
        unsigned char *t = a + n;
        volatile unsigned char *k = a + n;
        if (k<a)
                return (unsigned char *)-1;
        return t;
}