https://bugs.llvm.org/show_bug.cgi?id=44126

            Bug ID: 44126
           Summary: Improve IR for code which finds position of highest
                    bit
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Scalar Optimizations
          Assignee: [email protected]
          Reporter: [email protected]
                CC: [email protected]

https://github.com/facebook/zstd/blob/47034cd6c31125fdba3155abe9a618f580b4f3eb/programs/fileio.c#L1789

unsigned long long FIO_highbit64(unsigned long long v)
{
    unsigned  count = 0;
    v >>= 1;
    while (v) { v >>= 1; count++; }
    return count;
}


should be same as:

unsigned long long FIO_highbit64a(unsigned long long v)
{

    return 63 - __builtin_clzll(v);
}

But first version has worse IR and codegen:
define dso_local i64 @_Z13FIO_highbit64y(i64 %0) local_unnamed_addr #0 {
  %2 = lshr i64 %0, 1
  %3 = call i64 @llvm.ctlz.i64(i64 %2, i1 false), !range !2
  %4 = sub nuw nsw i64 64, %3
  ret i64 %4
}

=>

define dso_local i64 @_Z14FIO_highbit64ay(i64 %0) local_unnamed_addr #1 {
  %2 = tail call i64 @llvm.ctlz.i64(i64 %0, i1 true), !range !3
  %3 = xor i64 %2, 63
  ret i64 %3
}


It would be good to not forget on trunc variant:
unsigned FIO_highbit64(unsigned long long v)
{
    unsigned count = 0;
    v >>= 1;
    while (v) { v >>= 1; count++; }
    return count;
}

define dso_local i32 @_Z13FIO_highbit64y(i64 %0) local_unnamed_addr #0 {
  %2 = lshr i64 %0, 1
  %3 = call i64 @llvm.ctlz.i64(i64 %2, i1 false), !range !2
  %4 = trunc i64 %3 to i32
  %5 = sub nsw i32 64, %4
  ret i32 %5
}

=>

define dso_local i32 @_Z13FIO_highbit64y(i64 %0) local_unnamed_addr #0 {
  %2 = tail call i64 @llvm.cttz.i64(i64 %0, i1 true), !range !2
  %3 = trunc i64 %2 to i32
  %4 = xor i32 %3, 63
  ret i32 %4
}


FIO_highbit64(unsigned long long):
        shr     rdi
        je      .LBB0_1
        bsr     rcx, rdi
        xor     rcx, 63
        mov     eax, 64
        sub     eax, ecx
        ret
.LBB0_1:
        mov     ecx, 64
        mov     eax, 64
        sub     eax, ecx
        ret

vs:

FIO_highbit64(unsigned long long):
        bsf     rax, rdi
        xor     eax, 63
        ret

-- 
You are receiving this mail because:
You are on the CC list for the bug.
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs

Reply via email to