I was experimenting in Rust with implementing the ASCII versions of certain unicode predicates in terms of an implementation on ASCII bytes, something like this https://rust.godbolt.org/z/c495YKTsM
pub fn demo_is_ascii_digit(c: char) -> bool {
if let Some(c) = AsciiChar::new(c) {
c.is_digit()
} else {
false
}
}Unfortunately, that optimizes much worse than checking against the 32-bit type directly:
define noundef zeroext i1 @_ZN10playground19demo_is_ascii_digit17h37e7c57418c2459fE(i32 noundef %c) unnamed_addr #0 {
start:
%_4.i = icmp ult i32 %c, 128
%_7.i = trunc i32 %c to i8
%0 = add i8 %_7.i, -48
%1 = icmp ult i8 %0, 10
%.0 = and i1 %_4.i, %1
ret i1 %.0
}
define noundef zeroext i1 @_ZN10playground18std_is_ascii_digit17he953dd4cafd53a1eE(i32 noundef %0) unnamed_addr #0 {
start:
%1 = add i32 %0, -48
%2 = icmp ult i32 %1, 10
ret i1 %2
}Alive2 confirms that converting the former there into the latter is legal: https://alive2.llvm.org/ce/z/_vhQJE
But opt seemingly can't do it today: https://llvm.godbolt.org/z/zdsEsvrsP
Perhaps the way to get there would be to do something like this:
start:
%_4.i = icmp ult i32 %c, 128
- %_7.i = trunc i32 %c to i8
- %0 = add i8 %_7.i, -48
+ %0 = add i32 %c, -48
- %1 = icmp ult i8 %0, 10
+ %1 = icmp ult i32 %0, 10
%.0 = and i1 %_4.i, %1
ret i1 %.0
Since alive2 says that's legal https://alive2.llvm.org/ce/z/bETvXP and after that other existing optimizations can get the rest of the way: https://llvm.godbolt.org/z/aEhThW96M
I tried to make a C++ repro as well, which came out with a slightly different, but still suboptimal, output: https://cpp.godbolt.org/z/r5xGxa8bc
bool is_ascii_char(uint32_t c) {
if (c < 128) {
auto b = (uint8_t)c;
return (b >= (uint8_t)'0') && (b <= (uint8_t)'9');
} else {
return false;
}
}define dso_local noundef zeroext i1 @_Z13is_ascii_charj(i32 noundef %0) local_unnamed_addr #0 {
%2 = icmp ult i32 %0, 128
%3 = and i32 %0, 254 ; <-- the comparisons got widened, but the `and` didn't disappear
%4 = add nsw i32 %3, -48
%5 = icmp ult i32 %4, 10
%6 = select i1 %2, i1 %5, i1 false
ret i1 %6
}
I was experimenting in Rust with implementing the ASCII versions of certain unicode predicates in terms of an implementation on ASCII bytes, something like this https://rust.godbolt.org/z/c495YKTsM
Unfortunately, that optimizes much worse than checking against the 32-bit type directly:
Alive2 confirms that converting the former there into the latter is legal: https://alive2.llvm.org/ce/z/_vhQJE
But opt seemingly can't do it today: https://llvm.godbolt.org/z/zdsEsvrsP
Perhaps the way to get there would be to do something like this:
Since alive2 says that's legal https://alive2.llvm.org/ce/z/bETvXP and after that other existing optimizations can get the rest of the way: https://llvm.godbolt.org/z/aEhThW96M
I tried to make a C++ repro as well, which came out with a slightly different, but still suboptimal, output: https://cpp.godbolt.org/z/r5xGxa8bc