[optimization] gcc generate better code than clang base on value range propagation

[vfdff]

after fix the case #54132, the clang still need be improved for some more complex case （see detail on https://godbolt.org/z/xb9Mo85WM ）, such as:

#include <stdlib.h>

int my_abs (int x, int y) {
  return x>y ? abs (x-y+1): 0;
}

• clang's code：

my_abs: // @my_abs
  subs w8, w0, w1
  b.le .LBB0_2
  mvn w9, w8
  cmn w8, #1
  csinc w0, w9, w8, lt
  ret
.LBB0_2:
  mov w0, wzr
  ret

• gcc's code:

my_abs:
  subs w0, w0, w1
  csinc w0, wzr, w0, le
  ret

[vfdff]

the issue still exist, https://alive2.llvm.org/ce/z/ZTU2ct

[vfdff]

• base on the llvm15. it transform the abs into a select, and the newest upstream don't try this

IC: Replacing   %call = call i32 @abs(i32 noundef %add) #2
    with   %1 = select i1 %0, i32 %neg, i32 %add

This is because the new upstream transform call i32 @abs into call i32 @llvm.abs.i32 at the begin of the optimazation, which casues entry into different branches for II = dyn_cast<IntrinsicInst>(CI) in function LibCallSimplifier::optimizeCall. This changes from llvm17 to llvm18, https://gcc.godbolt.org/z/8jx4d13Wo