JIT: Optimize movzx after setcc

[EgorBo]

Clear target reg via xor reg, reg instead of relying on zero-extend after SETCC which is heavier and slower

bool Test(int x) => x == 42;

; Method Test(int):bool:this
G_M55728_IG01:              ;; offset=0000H
G_M55728_IG02:              ;; offset=0000H
+      33C0                 xor      eax, eax
       83FA2A               cmp      edx, 42
       0F94C0               sete     al
-      0FB6C0               movzx    rax, al
G_M55728_IG03:              ;; offset=0009H
       C3                   ret      
-; Total bytes of code: 10
+; Total bytes of code: 9

Nice diffs:

benchmarks.run.Linux.x64.checked.mch:
Total bytes of delta: -5230 (-0.03 % of base)

coreclr_tests.pmi.Linux.x64.checked.mch:
Total bytes of delta: -210860 (-0.16 % of base)

libraries.crossgen2.Linux.x64.checked.mch:
Total bytes of delta: -4420 (-0.06 % of base)

libraries.pmi.Linux.x64.checked.mch:
Total bytes of delta: -25312 (-0.05 % of base)

libraries_tests.pmi.Linux.x64.checked.mch:
Total bytes of delta: -49314 (-0.04 % of base)

[ghost]

Tagging subscribers to this area: @JulieLeeMSFT
See info in area-owners.md if you want to be subscribed.

Issue Details

---

Clear target reg via xor reg, reg instead of relying on zero-extend after SETCC which is heavier and slower

bool Test(int x) => x == 42;

; Method Test(int):bool:this
G_M55728_IG01:              ;; offset=0000H
G_M55728_IG02:              ;; offset=0000H
+      33C0                 xor      eax, eax
       83FA2A               cmp      edx, 42
       0F94C0               sete     al
-      0FB6C0               movzx    rax, al
G_M55728_IG03:              ;; offset=0009H
       C3                   ret      
; Total bytes of code: 10

Local diffs were promising...

Author:	EgorBo
Assignees:	EgorBo
Labels:	area-CodeGen-coreclr
Milestone:	-

[EgorBo]

it's not safe apparently, xor clears flags

[jakobbotsch]

it's not safe apparently, xor clears flags

In which cases did it matter? Doesn't all of these cases emit a cmp after?

[EgorBo]

it's not safe apparently, xor clears flags

In which cases did it matter? Doesn't all of these cases emit a cmp after?

I wasn't able to detect - tests were keeping failing, I suspect some complex logic in emitInsBinary?

[EgorBo]

Ah, I think I've found out why

[EgorBo]

My understanding that it used to crash on things like

xor      eax, eax
cmp      dword ptr [rax+8], 0

the fix is quite conservative

[EgorBo]

cc @dotnet/jit-contrib @jakobbotsch

[jakobbotsch]

LGTM. What do the diffs look like if you do something less conservative? e.g.

regMaskTP GenTree::gtGetContainedRegMask()
{
    regMaskTP mask = gtGetRegMask();
    for (GenTree* operand : Operands())
    {
        if (operand->isContained())
        {
            mask |= operand->gtGetContainedRegMask();
        }
    }

    return mask;
}

and do the optimization unless (op1->gtGetContainedRegMask() | op2->gtGetContainedRegMask()) & genRegMask(targetReg)? (not sure if that's completely right)

[kasperk81]

Nice diffs:

nice. and regressions are just cse diffs or something more interesting?

[EgorBo]

Nice diffs:

nice. and regressions are just cse diffs or something more interesting?

No, it's too late for CSE, from my understanding it happens because we used to end up with:

cmp      edx, 42
sete     al
movzx    rax, al
movzx    rax, al

where the 2nd mov was optimized out via RedundantMov

 -- suppressing mov because last instruction already moved from src to dst register.

with my PR we end up with:

xor      eax, eax
cmp      edx, 42
sete     al
movzx    rax, al

and we don't benefit from RedundantMov here.

it can be fixed I guess, but not as part of this PR

Repro:

public enum ResultCode
{
    Success = 0,
    SaslBindInProgress = 14,
    NoSuchAttribute = 16,
    InvalidAttributeSyntax = 21,
    NoSuchObject = 32,
    InvalidDNSyntax = 34,
    AliasDereferencingProblem = 36,
    InappropriateAuthentication = 48,
}

internal static bool IsResultCode(ResultCode code)
{
    if (code >= ResultCode.Success && code <= ResultCode.SaslBindInProgress)
    {
        return true;
    }

    if (code >= ResultCode.NoSuchAttribute && code <= ResultCode.InvalidAttributeSyntax)
    {
        return true;
    }

    if (code >= ResultCode.NoSuchObject && code <= ResultCode.InvalidDNSyntax)
    {
        return true;
    }

    return (code == ResultCode.AliasDereferencingProblem || 
            code == ResultCode.InappropriateAuthentication);
}

[jakobbotsch]

This still seems more conservative than necessary, doesn't it only need to or in the operand regs if it itself is contained?

[jakobbotsch]

I.e. I guess this is a more correct version than my original:

regMaskTP GenTree::gtGetContainedRegMask()
{
    if (!isContained())
        return gtGetRegMask();

    regMaskTP mask = 0;
    for (GenTree* operand : Operands())
    {
        mask |= operand->gtGetContainedRegMask();
    }
    return mask;
}

[kasperk81]

thanks. overall it is nice win (regression in bignumber format method is a red-herring).
c++ compiler generated interesting asm with that repro (almost half the no. of lines)

IsResultCode(ResultCode):
        cmp     edi, 34
        ja      .L2
        movabs  rdx, -30068932608
        mov     eax, 1
        bt      rdx, rdi
        jc      .L2
        ret
.L2:
        cmp     edi, 36
        sete    al
        cmp     edi, 48
        sete    dl
        or      eax, edx
        ret

[EgorBo]

c++ compiler generated interesting asm with that repr

Yeah, the bitmask hack is a nice thing to have, e.g. here it's used in C# https://github.com/dotnet/runtime/blob/main/src%2Flibraries%2FSystem.Private.CoreLib%2Fsrc%2FSystem%2FRuntime%2FCompilerServices%2FRuntimeHelpers.cs#L87-L89 (because jit is not yet able to optimize it)

[EgorBo]

@jakobbotsch thank you, diffs are even bigger now