[Intel] Failure to shift-in one bits via `shld`

[Validark]

Zig Godbolt

Attempt 1:

export fn shift_in_ones(a: u64, c: u8) u64 {
    return ~(~a << @truncate(c));
}

shift_in_ones:
        not     rdi
        shlx    rax, rdi, rsi
        not     rax
        ret

Attempt 2:

export fn shld(a: u64, b: u64, c: u8) u64 {
    return @truncate(@as(u128, @bitCast([2]u64{ a, b })) << @as(u6, @truncate(c)) >> 64);
}

export fn shift_in_ones_2(a: u64, c: u8) u64 {
    return shld(~@as(u64, 0), a, c);
}

shld:
        mov     ecx, edx
        mov     rax, rsi
        shld    rax, rdi, cl
        ret

shift_in_ones_2: ; for some reason inlining makes us not see that this a `shld`?
        shlx    rcx, rdi, rsi
        and     sil, 63
        not     sil
        movabs  rax, 9223372036854775807
        shrx    rax, rax, rsi
        or      rax, rcx
        ret

Should be:

shift_in_ones:
        mov     ecx, esi
        mov     rax, rdi
        mov     rdx, -1
        shld    rax, rdx, cl
        ret

[aengelke]

Dynamic count shld has a 2-5 cycle latency and 3-5 uops, shlx typically just 1 cycle and 1 uop. The code in the first output looks pretty optimal in terms of performance. Why should shld be preferred?

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: Niles Salter (Validark)

[Zig Godbolt](https://zig.godbo.lt/#g:!((g:!((g:!((h:codeEditor,i:(filename:'1',fontScale:14,fontUsePx:'0',j:1,lang:zig,selection:(endColumn:2,endLineNumber:11,positionColumn:1,positionLineNumber:5,selectionStartColumn:2,selectionStartLineNumber:11,startColumn:1,startLineNumber:5),source:'export+fn+shift_in_ones(a:+u64,+c:+u8)+u64+%7B%0A++++return+~(~a+%3C%3C+@truncate(c))%3B%0A%7D%0A%0Aexport+fn+shld(a:+u64,+b:+u64,+c:+u8)+u64+%7B%0A++++return+@truncate(@as(u128,+@bitCast(%5B2%5Du64%7B+a,+b+%7D))+%3C%3C+@as(u6,+@truncate(c))+%3E%3E+64)%3B%0A%7D%0A%0Aexport+fn+shift_in_ones_2(a:+u64,+c:+u8)+u64+%7B%0A++++return+shld(~@as(u64,+0),+a,+c)%3B%0A%7D%0A%0Aexport+fn+shift_in_ones_3(a:+u64,+c:+u8)+u64+%7B%0A++++return+shld_asm(~@as(u64,+0),+a,+c)%3B%0A%7D%0A%0Apub+fn+shld_asm(low:+u64,+high:+u64,+shift_amt:+u8)+u64+%7B%0A++++var+ret:+u64+%3D+high%3B%0A++++asm+(%22shldq+%25%5Bcnt%5D,+%25%5Bsrc%5D,+%25%5Bret%5D%22%0A++++++++:+%5Bret%5D+%22%2Br%22+(ret),%0A++++++++:+%5Bsrc%5D+%22r%22+(low),%0A++++++++++%5Bcnt%5D+%22%7Bcl%7D%22+(shift_amt),%0A++++++++:+.%7B+.cc+%3D+true+%7D)%3B%0A++++return+ret%3B%0A%7D'),l:'5',n:'0',o:'Zig+source+%231',t:'0')),k:55.575874538014034,l:'4',m:100,n:'0',o:'',s:0,t:'0'),(g:!((h:compiler,i:(compiler:ztrunk,filters:(b:'0',binary:'1',binaryObject:'1',commentOnly:'0',debugCalls:'1',demangle:'0',directives:'0',execute:'1',intel:'0',libraryCode:'0',trim:'1',verboseDemangling:'0'),flagsViewOpen:'1',fontScale:14,fontUsePx:'0',j:2,lang:zig,libs:!(),options:'-O+ReleaseFast+-target+x86_64-linux+-mcpu%3Dicelake_server+-fomit-frame-pointer',overrides:!(),selection:(endColumn:12,endLineNumber:27,positionColumn:1,positionLineNumber:21,selectionStartColumn:12,selectionStartLineNumber:27,startColumn:1,startLineNumber:21),source:1),l:'5',n:'0',o:'+zig+trunk+(Editor+%231)',t:'0')),header:(),k:44.42412546198598,l:'4',m:100.00000000000001,n:'0',o:'',s:0,t:'0')),l:'2',m:100,n:'0',o:'',t:'0')),version:4)

Attempt 1:

export fn shift_in_ones(a: u64, c: u8) u64 {
    return ~(~a &lt;&lt; @<!-- -->truncate(c));
}

shift_in_ones:
        not     rdi
        shlx    rax, rdi, rsi
        not     rax
        ret

Attempt 2:

export fn shld(a: u64, b: u64, c: u8) u64 {
    return @<!-- -->truncate(@<!-- -->as(u128, @<!-- -->bitCast([2]u64{ a, b })) &lt;&lt; @<!-- -->as(u6, @<!-- -->truncate(c)) &gt;&gt; 64);
}

export fn shift_in_ones_2(a: u64, c: u8) u64 {
    return shld(~@<!-- -->as(u64, 0), a, c);
}

shld:
        mov     ecx, edx
        mov     rax, rsi
        shld    rax, rdi, cl
        ret

shift_in_ones_2: ; for some reason inlining makes us not see that this a `shld`?
        shlx    rcx, rdi, rsi
        and     sil, 63
        not     sil
        movabs  rax, 9223372036854775807
        shrx    rax, rax, rsi
        or      rax, rcx
        ret

Should be:

shift_in_ones:
        mov     ecx, esi
        mov     rax, rdi
        mov     rdx, -1
        shld    rax, rdx, cl
        ret