Missed opportunity to use cmp with exx instead od rxx (code size optimization) (Bugzilla Bug 5)

[asl]

This issue was created automatically with bugzilla2github

Bugzilla Bug 5

Date: 2019-10-27T07:45:09+07:00
From: David Bolvansky <david.bolvansky@gmail.com>
To: Unassigned LLVM Bugs <unassignedbugs@nondot.org>
CC: @asl, llvm-bugs@lists.llvm.org, llvm-dev@redking.me.uk, @rotateright

Last updated: 2019-10-27T07:45:09+07:00

[asl]

Comment 190262

Date: 2019-10-27 07:45:09 -0700
From: David Bolvansky <david.bolvansky@gmail.com>

bool test1(U64 val)
{
    return ((U32)(val & 0x120)) == 0x20;
}

define dso_local zeroext i1 @test1(i64 %0) local_unnamed_addr #0 {
  %2 = and i64 %0, 288
  %3 = icmp eq i64 %2, 32
  ret i1 %3
}

test1:  
        and     edi, 288
        cmp     rdi, 32
        sete    al
        ret

(-O0's trunc LLVM IR instruction is eliminated)

but codegen could use "cmp edi" instead of "cmp rdi":

test1:
        and     edi, 288
        cmp     edi, 32
        sete    al
        ret

GCC does this code size optimization (eliminate REX prefix), Clang does not.

Another opportunity for "cmp edi":
bool test2(U64 val)
{
    return (val & 0x120) == 0x20;
}

test2:                     
        and     edi, 288
        cmp     rdi, 32
        sete    al
        ret

Both GCC and Clang fails to use rdi for "test2" case.

https://godbolt.org/z/A0hjTg


Another case:

bool test3(U64 val)
{
    return ((short)(val & 0x120)) == 0x20;
}

Clang
test3:                                  # @test3
        and     edi, 288
        cmp     rdi, 32
        sete    al
        ret

GCC
test3:
        and     di, 288
        cmp     di, 32
        sete    al
        ret