[llvm-dev] [cfe-dev] FE_INEXACT being set for an exact conversion from float to unsigned long long

Kaylor, Andrew via llvm-dev llvm-dev at lists.llvm.org
Thu Apr 20 17:30:06 PDT 2017

I think it’s generally true that whenever branches can reliably be predicted branching is faster than a cmov that involves speculative execution, and I would guess that your assessment regarding looping on input values is probably correct.

I believe the code that actually creates most of the transformation you’re interested in here is in SelectionDAGLegalize::ExpandNode() in LegalizeDAG.cpp.  The X86 backend sets a table entry indicating that FP_TO_UINT should be expanded for these value types, but the actual expansion is in target-independent code.  This is what it looks like in the version I last fetched:

  case ISD::FP_TO_UINT: {
    SDValue True, False;
    EVT VT =  Node->getOperand(0).getValueType();
    EVT NVT = Node->getValueType(0);
    APFloat apf(DAG.EVTToAPFloatSemantics(VT),
    APInt x = APInt::getSignBit(NVT.getSizeInBits());
    (void)apf.convertFromAPInt(x, false, APFloat::rmNearestTiesToEven);
    Tmp1 = DAG.getConstantFP(apf, dl, VT);
    Tmp2 = DAG.getSetCC(dl, getSetCCResultType(VT),
                        Tmp1, ISD::SETLT);
    True = DAG.getNode(ISD::FP_TO_SINT, dl, NVT, Node->getOperand(0));
    // TODO: Should any fast-math-flags be set for the FSUB?
    False = DAG.getNode(ISD::FP_TO_SINT, dl, NVT,
                        DAG.getNode(ISD::FSUB, dl, VT,
                                    Node->getOperand(0), Tmp1));
    False = DAG.getNode(ISD::XOR, dl, NVT, False,
                        DAG.getConstant(x, dl, NVT));
    Tmp1 = DAG.getSelect(dl, NVT, Tmp2, True, False);

The tricky bit here is that this code is asking for a Select and then something else will decide whether that select should be implemented as a branch or a cmov.

Circling back to what I said earlier, while I stand by the idea that we shouldn’t sacrifice performance just to maintain FP status bits, you’ve got me warming up to the idea that maybe an FP status safe implementation of this might be faster in most cases.  It’s at least worth taking some measurements to see if it is faster.


From: Michael Clark [mailto:michaeljclark at mac.com]
Sent: Thursday, April 20, 2017 4:02 PM
To: Kaylor, Andrew <andrew.kaylor at intel.com>
Cc: Flamedoge <code.kchoi at gmail.com>; llvm-dev <llvm-dev at lists.llvm.org>
Subject: Re: [llvm-dev] [cfe-dev] FE_INEXACT being set for an exact conversion from float to unsigned long long

On 21 Apr 2017, at 8:50 AM, Kaylor, Andrew <andrew.kaylor at intel.com<mailto:andrew.kaylor at intel.com>> wrote:

> This seems like it was done for perf reason (mispredict). Conditional-to-cmov transformation should keep
> from introducing additional observable side-effects, and it's clear that whatever did this did not account
> for floating point exception.

That’s a very reasonable statement, but I’m not sure it corresponds to the way we have typically approached this sort of problem.

Just out of interest, I ran the two code sequences (branch vs predicate logic) through Intel Architecture Code Analyzer:

- https://gist.github.com/michaeljclark/84f9e411ab959073996e10bff6ab8ec0

I am not sure how accurate the cycle estimate is for the GCC code sequence as it says “possibly". It would appear that if the branch is predicted in a loop of floating point values that are < LLONG_MAX then GCC would win. Some values > LLONG_MAX and some <= LLONG_MAX would fool the branch predictor.

In particular, there has been a de facto standard practice (and it has even been openly stated and agreed upon once or twice) of assuming default rounding mode and ignoring FP exceptions in the default (and currently only) optimization path for FP-related instructions.  That is, clang/LLVM didn’t just not support FENV_ACCESS by indifference but rather we have made conscious decisions to allow transformations that violate the needs of FENV_ACCESS when doing so can improve the performance of generated code.  Basically, we more or less pretend that floating point status bits don’t exist (at least before you get to the target-specific backend).

It is a shame as Clang/LLVM does surprisingly well at honouring floating point accrued exceptions on x86_64 as is (see below).

As far as I can tell, besides the eager constant folding optimisation, the only issues I am facing is with the float to u64 and double to u64 intrinsics.

I have expressed the RISC-V floating point ISA in a C notation, which I use to build a RISC-V ISA simulator, and the floating point accrued exceptions are tested for the complete set of RISC-V FPU operations. I’ve been using the RISC-V ISA test suite essentially to test gcc and clang floating point accrued exception state and the test suite has relatively complete accrued exception state tests after any floating point operations that update the accrued exception state. Clang does very well, with the exception of float to u64 and double to u64. They are currently the only two operations that fail the RISC-V QEMU tests, which is essentially testing all of the compiler floating point intrinsics.


These are the tests that I am using to exercise clang and gcc FENV_ACCESS along with the C code in the link above. riscv-qemu-tests is a user-mode port of the RISC-V ISA conformance test suite.

- https://github.com/arsv/riscv-qemu-tests/tree/master/rv64f
- https://github.com/arsv/riscv-qemu-tests/tree/master/rv64d

The tests are not exhaustive. There are some special cases with rounding modes that are not tested and the code is isolated from certain optimisations and is essentially testing just the intrinsics. That said, it makes sense for the intrinsics to correctly set the floating point accrued exception state to allow any more complete handling of optimisation around floating point operations.

In any case I need these conversions to work on compilers today so I came up with this work-around:

- https://godbolt.org/g/Ez8RZA

With the workaround above I am able to pass all of the RISC-V tests, however I know of some corner cases that are not tested. Interestingly float and double to unsigned need to round or clamp negative values to zero. I guess this is undefined behaviour in C.

You’ll find that the X86 backend doesn’t even model MXCSR at the moment.  I tried to add it recently and it kind of blew up before I had even modeled it for anything other than LDMXCSR and STMCXSR.  We may want to address that at some point, but right now it just isn’t there.

When we discussed how FENV_ACCESS support should be implemented, Chandler proposed that when restricted modes (whether FENV_ACCESS or any other front end-specific analogous behavior) were not being used the optimizer should be able to behave as described above and that nothing done to support restricted FP behavior should complicate or restrict the default optimizer behavior.  This was met with general agreement at the time.

I mention all this as prologue to saying that while we should do something to get FPToUI lowered without incorrectly setting FP exception status bits, it isn’t necessarily what we want as the default behavior.

I would also like to add that I personally am very pleased that you discovered this issue and have gotten as far as you have in the analysis of the problem.  I’m in the process of adding constrained versions of various FP intrinsics (I have a patch ready to be sent out today) and what I’ve done up to now has been to simply translate the constrained operations into their traditional representations somewhere in the ISel process.  I was aware that something would need to be done in the codegen space to continue protecting these operations, but I was kind of hoping that the actual instruction selection would be reasonably safe.  FWIW, FPToUI is not one of the parts my pending patch addresses.


From: llvm-dev [mailto:llvm-dev-bounces at lists.llvm.org] On Behalf Of Flamedoge via llvm-dev
Sent: Wednesday, April 19, 2017 10:14 AM
To: Michael Clark <michaeljclark at mac.com<mailto:michaeljclark at mac.com>>
Cc: llvm-dev <llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>>
Subject: Re: [llvm-dev] [cfe-dev] FE_INEXACT being set for an exact conversion from float to unsigned long long

> Are we better off using branches instead of cmove to implement FP to
unsigned i64?

This seems like it was done for perf reason (mispredict). Conditional-to-cmov transformation should keep from introducing additional observable side-effects, and it's clear that whatever did this did not account for floating point exception.

On Wed, Apr 19, 2017 at 10:01 AM, Michael Clark via llvm-dev <llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>> wrote:
Changing the list from cfe-dev to llvm-dev

On 20 Apr 2017, at 4:52 AM, Michael Clark <michaeljclark at mac.com<mailto:michaeljclark at mac.com>> wrote:

I’m getting close. I think it may be an issue with an individual intrinsic. I’m looking for the X86 lowering of Instruction::FPToUI.

I found a comment around the rationale for using a conditional move versus a branch. I believe the predicate logic using a conditional move is causing INEXACT to be set from the other side of the predicate as the lowered x86_64 code executes both conversions whereas GCC uses a branch. That seems to be the difference.

I can’t find FPToUI in llvm/lib/Target/X86 so I’m trying to figure out what the cast gets renamed to in the target layer so I can find where the sequence is emitted.

$ more llvm/lib/Target/X86//README-X86-64.txt
Are we better off using branches instead of cmove to implement FP to
unsigned i64?

        ucomiss LC0(%rip), %xmm0
        cvttss2siq      %xmm0, %rdx
        jb      L3
        subss   LC0(%rip), %xmm0
        movabsq $-9223372036854775808, %rax
        cvttss2siq      %xmm0, %rdx
        xorq    %rax, %rdx
        movq    %rdx, %rax

instead of

        movss LCPI1_0(%rip), %xmm1
        cvttss2siq %xmm0, %rcx
        movaps %xmm0, %xmm2
        subss %xmm1, %xmm2
        cvttss2siq %xmm2, %rax
        movabsq $-9223372036854775808, %rdx
        xorq %rdx, %rax
        ucomiss %xmm1, %xmm0
        cmovb %rcx, %rax

On 19 Apr 2017, at 2:10 PM, Michael Clark <michaeljclark at mac.com<mailto:michaeljclark at mac.com>> wrote:

On 19 Apr 2017, at 1:14 PM, Tim Northover <t.p.northover at gmail.com<mailto:t.p.northover at gmail.com>> wrote:

On 18 April 2017 at 15:54, Michael Clark via cfe-dev
<cfe-dev at lists.llvm.org<mailto:cfe-dev at lists.llvm.org>> wrote:

The only way towards completing a milestone is via fixing a number of small issues along
the way…

I believe there's more to it than that. None of LLVM's optimizations
are aware of this extra side-channel of information (with possible
exceptions like avoiding speculating fdiv because of unavoidable

From what I remember, the real proposal is to replace all
floating-point IR with intrinsics when FENV_ACCESS is on, which the
optimizers by default won't have a clue about and will treat
conservatively (essentially like they're modifying external memory).

So be careful with drawing conclusions from small snippets; you're
probably not seeing the full range of LLVM's behaviour.

Yes. I’m sure.

It reproduces with just the cast on its own: https://godbolt.org/g/myUoL2

It appears to be in the LLVM lowering of the fptoui intrinsic so it must MC layer optimisations.

; Function Attrs: noinline nounwind uwtable
define i64 @_Z7fcvt_luf(float %f) #0 {
  %1 = alloca float, align 4
  store float %f, float* %1, align 4
  %2 = load float, float* %1, align 4
  %3 = fptoui float %2 to i64
  ret i64 %3

GCC performs a comparison with ucomiss and branches whereas Clang computes both forms and predicates the result using a conditional move. One of the conversions obviously is setting the INEXACT MXCSR flag.

Clang lowering (inexact set when result is exact):

        movss   xmm1, dword ptr [rip + .LCPI1_0] # xmm1 = mem[0],zero,zero,zero
        movaps  xmm2, xmm0
        subss   xmm2, xmm1
        cvttss2si       rax, xmm2
        movabs  rcx, -9223372036854775808
        xor     rcx, rax
        cvttss2si       rax, xmm0
        ucomiss xmm0, xmm1
        cmovae  rax, rcx

GCC lowering (sets flags correctly):

        ucomiss xmm0, DWORD PTR .LC0[rip]
        jnb     .L4
        cvttss2si       rax, xmm0
        subss   xmm0, DWORD PTR .LC0[rip]
        movabs  rdx, -9223372036854775808
        cvttss2si       rax, xmm0
        xor     rax, rdx

LLVM Developers mailing list
llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20170421/97a716e6/attachment-0001.html>

More information about the llvm-dev mailing list