[IR] redefine 'reassoc' fast-math-flag and add 'trans' fast-math-flag

[spatel]

Summary

As discussed on llvm-dev:
http://lists.llvm.org/pipermail/llvm-dev/2016-November/107104.html
and again more recently:
http://lists.llvm.org/pipermail/llvm-dev/2017-October/118118.html

...this is a step in cleaning up our fast-math-flags implementation in IR.

As proposed in the above threads, I've replaced the 'UnsafeAlgebra' bit (which had the 'umbrella' meaning that all flags are set) with a new bit that only applies to algebraic reassociation - 'AllowReassoc'.

I've also added a bit to allow relaxed precision for transcendental functions called 'AllowTrans' (this was initially proposed as 'libm' or similar).

...and we're out of bits. 7 bits ought to be enough for anyone, right? :) FWIW, I did look at getting this out of SubclassOptionalData via SubclassData (spacious 16-bits), but that's apparently already used for other purposes. Also, I don't think we can just add a field to FPMathOperator because Operator is not intended to be instantiated. We'll defer movement of FMF to another day.

I'm not sure, but I may be diverging from the last proposal by keeping the 'fast' keyword. I thought about removing that, but seeing IR like this:
%f.fast = fadd reassoc nnan ninf nsz arcp contract trans float %op1, %op2
...made me think we want to keep the shortcut synonym.

I've also gone ahead and renamed the getter/setters, and mechanically added 'TODO' comments where we need to review how the old setUnsafeAlgebra() was used. In some cases, it's obvious that should be translated to setAllowReassoc(), but others may need to be discussed.

Finally, this change is binary incompatible with existing IR as seen in the compatibility tests. I'm hoping this:
"Newer releases can ignore features from older releases, but they cannot miscompile them. For example, if nsw is ever replaced with something else, dropping it would be a valid way to upgrade the IR." ( http://llvm.org/docs/DeveloperPolicy.html#ir-backwards-compatibility )
...provides the flexibility we want to make this change without requiring a new IR version. Ie, I don't think we're ever loosening the FP strictness of existing IR. At worst, we will fail to optimize some previously 'fast' code because it's no longer recognized as 'fast'. This should get fixed as we squash all of the TODO comments.

[spatel]

Patch updated:
I forgot to add assembler tests for the new keywords which revealed that I had forgotten to add 'reassoc' to the parser. Fixed the code and added tests.

[efriedma]

docs/LangRef.rst

2305

Could you mark up the LLVM intrinsics affected by the "trans" flag with the meaning with/without the flag?

Please clarify the language here to indicate this affects the semantics of both LLVM intrinsics and known library functions. And include a more complete description of what counts as a known library function. And explain what "relaxed" means, given that libm generally doesn't provide correctly rounded versions of transcendental functions.

Also, do we want to optimize sqrt() based on this flag? It technically isn't transcendental, but we currently generate an approximation in some cases based on fast-math flags.

[spatel]

docs/LangRef.rst

2305

I think we'd include sqrt() in the 'trans' bucket (so maybe 'libm' was the better name). But looking back through the dev thread, I don't see an actual definition of that term or what this flag would map to as a clang command-line param. @hfinkel / @wristow - suggestions?

[spatel]

D39319 (Set contract flag when setting unsafe algebra flag) proposes to fix a bug introduced with D31164 (Add AllowContract to FastMathFlags). That bug is also fixed here - but I should add that test.

This might suggest that we should split 'trans' into its own small follow-up patch. That would also allow the unsafe TODO fixes to proceed independently of 'trans' in case that takes longer to sort out. Let me know if splitting this patch up seems better.

[hfinkel]

I don't see a lot of value in having:

// TODO: This should use hasAllowReassoc()?

before every call to isFast. I can grep for isFast more easily than I can grep for the comment. Moreover, if something really needs 'isFast' because the transformation really needs "every possible liberty", then there should definitely be a comment explaining that.

Also, unless we make hasAllowReassoc() also imply hasNoSignedZeros() (as GCC does, FWIW), we'll almost certainly need that check as well in many places. Many need NoNans and NoInfs too.

docs/LangRef.rst

2305

You're correct, sqrt is not a transcendental function (it is an algebraic function because it is the root of a polynomial equation). The problem is that algebraic functions also include things that we don't want to include here (e.g., addition, division). I don't like 'libm' because it refers to a very specific set of functions, but maybe that's the best we can do. The best description I have of this set is: "All of the mathematical operations that generally produce irrational numbers and are not in the set of functions specified by the IEEE specification (e.g., +,-,*,/,%,sqrt)."

[wristow]

docs/LangRef.rst

2305

For context, the original suggestion of the flag-name 'libm' was changed to 'trans' based on an observation that OpenCL has an option '-cl-fast-relaxed-math' that includes the semantics "This option also relaxes the precision of commonly used math functions" (see http://man.opencl.org/clCompileProgram.html), and so 'libm' may incorrectly imply it only applies to a more limited set of functions (math operations) only implemented in "libm.a". That led to the suggestion of 'trans', for transcendental functions. I originally liked the 'trans' suggestion, moving away from the libm.a implication. But I do feel that sqrt() should also be optimized on this flag, so as people said, 'trans' isn't perfect either.

I don't have a precise formal definition of what things would be controlled by this, but loosely, I'd say it's "All mathematical operations that have runtime library support on many platforms." In practice in LLVM, I think this means many (most?) of the math operations handled by "SimplifyLibCalls.cpp". So for me, in addition to transcendental functions, it would include sqrt(), and even simpler things like fmin() and fmax().

So after giving it more thought, I'm now back to preferring 'libm' over 'trans'. If someone has a better suggestion, that would be great. But I haven't thought of one.

[wristow]

In D39304#907880, @spatel wrote:

...
This might suggest that we should split 'trans' into its own small follow-up patch. That would also allow the unsafe TODO fixes to proceed independently of 'trans' in case that takes longer to sort out. Let me know if splitting this patch up seems better.

I'd like to see the new enumerations of the FastMathFlags bits all defined in one patch. So removing 'UnsafeAlgebra', and adding 'AllowReassoc' and 'AllowTrans' (along with all the changes required since 'UnsafeAlgebra' can no longer be referenced) in the first patch, is my preference. (As an aside, probably 'AllowTrans' will be changed to something like 'AllowMathLib', due to other discussions here.)

[hfinkel]

In D39304#908340, @wristow wrote:

In D39304#907880, @spatel wrote:

...
This might suggest that we should split 'trans' into its own small follow-up patch. That would also allow the unsafe TODO fixes to proceed independently of 'trans' in case that takes longer to sort out. Let me know if splitting this patch up seems better.

I'd like to see the new enumerations of the FastMathFlags bits all defined in one patch. So removing 'UnsafeAlgebra', and adding 'AllowReassoc' and 'AllowTrans' (along with all the changes required since 'UnsafeAlgebra' can no longer be referenced) in the first patch, is my preference. (As an aside, probably 'AllowTrans' will be changed to something like 'AllowMathLib', due to other discussions here.)

Makes sense to me as well.

[wristow]

docs/LangRef.rst

2305

So after giving it more thought, I'm now back to preferring 'libm' over 'trans'. If someone has a better suggestion, that would be great. But I haven't thought of one.

In the spirit of the flag 'arcp' for 'AllowReciprocal', and the possibility of 'AllowMathLib' for the internal enumeration name, how about 'amlib'?

With that, there's no direct implication of "this is only for libm.a operations", or "this is only for transcendental functions". That said, the 'am' part does give it an awkward "morning library" feeling. Maybe 'amathlib' instead?

[andrew.w.kaylor]

docs/LangRef.rst

2305

It seems that we're allowing something kind of open-ended here. That is, we don't seem to have an exact set of functions that will be covered. If that's the case then we should probably document it as such -- something like "allow substitution of approximate calculations for functions whose meaning are recognized by the optimizer." And maybe the flag could be "approx".

2308

Since reciprocal is an algebraically-equivalent transformation, this documentation isn't quite correct. Does this enable anything other than reassociation?

[hfinkel]

docs/LangRef.rst

2305

I like this suggestion.

[wristow]

docs/LangRef.rst

2305

Yes, I think it's open-ended. And I very much like the description "allow substitution of approximate calculations for functions whose meaning are recognized by the optimizer.".

I'm less enthusiastic about the flag-name 'approx', although I'm not horribly opposed to it (especially since I cannot come up with anything I really like). On it's own, a flag named 'approx' sounds too wide of scope. To me, it sounds like it might be describing all of what is enabled by -ffast-math. In short, it doesn't explicitly convey the concept of it being approximate calculations for functions whose meanings are recognized. To describe my concern "from the other direction", the reciprocal transformation is also an approximation (as is virtually everything else enabled by fast-math), and we don't intend to control the reciprocal transformation via this flag.

2308

Good point. I don't think 'reassoc' intended to enable anything else. Maybe:

Allow algebraic reassociation transformations that may dramatically change results in floating point.

For that matter, I think 'algebraic' could also be removed, so just "Allow reassociation transformations that..." would work.

[hfinkel]

docs/LangRef.rst

2305

We could call it funcapprox or approxfunc or something like that. We could use something shorter too (apfn perhaps).

[wristow]

docs/LangRef.rst

2305

I like it. I'm happy with any of those three. Given the brevity of many of the others (eg, 'arcp', 'nnan'), I lean toward 'apfn'.

[spatel]

Patch updated:

1. Changed 'AllowTrans' to 'AllowMathLib' and 'trans' to 'aml'. The IR abbreviation is more alphabet soup, but less "awkward morning" than 'amlib' - still open to suggestions :)
2. Updated LangRef with suggested fixes. Trying to toe the weasel-word line here with enough ambiguity to accomplish transforms but not be completely meaningless.
3. Removed 'TODO' comments at uses of isFast() / setFast(). We can just grep those calls and fix them.
4. Added a variant of the 'fast' means all flags test from D39319.

[spatel]

Sadly, I didn't get email notifications for the last couple of name suggestions before I posted the updated patch. I agree that 'ApproxFunc' and 'afn' are better than 'AllowMathLib' and 'aml', so I'll change that unless there are objections or better suggestions.

[spatel]

Patch updated:
'AllowMathLib' --> 'ApproxFunc'
'aml' --> 'afn'

[hfinkel]

docs/LangRef.rst

2304

I think that we should essentially use Andrew's proposed definition here:

Allow substitution of approximate calculations for functions (e.g., sin, log, sqrt).

10498

This conformance language is only necessary for sqrt. For the other functions, there is no standard for their accuracy/precision. You might say that with 'afn' the result may not match what would have been returned by the system's libm implementation.

[spatel]

Patch updated - edited LangRef phrasing.

[hfinkel]

docs/LangRef.rst

10866

I don't expect that afn would affect fma. I recommend removing the statement here. It might be true that fma is computed differently under -ffast-math because of denormal handling, but that applies to everything. The same is true of the conversion functions below (floor, rint, etc.). maxnum/minnum too.

[spatel]

Patch updated:
We don't need the 'afn' warning label on every FP intrinsic. The note is now only applied to these 10:
sqrt, powi, sin, cos, pow, exp, exp2, log, log10, log2

[efriedma]

docs/LangRef.rst

10498

We probably want different language here than for the transcendental functions; libm sqrt() is precisely the IEEE754 squareRoot().

10538

Does afn actually do anything here? I think "unspecified sequence of rounding operations" implies it isn't exact anyway. (And __powisf2 isn't really part of libm.)

10574

This language might give the wrong idea. Even without 'afn', the result may not match the target's libm: we constant-fold using a different implementation. The part to call out is that we might substitute an implementation which is less accurate.

10866

We might want to transform fma(a,b,c) to a*b+c in fast-math mode on targets which don't have a native fma instruction? Not sure if that makes sense.

[hfinkel]

docs/LangRef.rst

10498

True. On the other hand, the system's libm sqrt should be IEEE compliant, so saying that this differs from the libm result covers that (and also covers other cases, such as PPC long double, which aren't IEEE).

10574

we constant-fold using a different implementation.

Which, indeed, might be less accurate -- just hopefully not by much ;)

I realize that having these notes on the intrinsics seems like it could be helpful, but I'm leaning toward recommending that we don't have them at all.

10866

Good point. This would be relatively easy to implement as well:

afn fma(a, b, c) -> [afn] fmuladd(a, b, c)

fmuladd lowers either to an fma, or not, depending on target preferences.

[wristow]

include/llvm/IR/Operator.h

220

One loose end that needs to be taken care of more or less simultaneously is a Clang change. Specifically, the constructor for CodeGenFunction (in "CodeGenFunction.cpp") invokes FastMathFlags::setUnsafeAlgebra(), so it will need to be changed to setFast().

[spatel]

include/llvm/IR/Operator.h

220

This is correct - I didn't post it, but I have that one line patch in place locally, so I was planning to submit it to the clang repo as close as possible after this patch and reference this commit (if there's a way to avoid the build breakage cleanly, please let me know).

[spatel]

Patch updated - just changes to the LangRef text (this now attempts to subsume the similar changes in D28335) :

1. Add an extra line to the 'sqrt' semantics because that has well-defined behavior...unless the type is not IEEE.
2. Remove the 'afn' blurb for 'powi' because that's loose by definition.
3. Add the 'afn' blurb back to 'fma' because that actually could be affected.
4. Try to soften the existing libm language for all 10 'afn'-affected intrinsics while retaining some shred of accuracy. :)

[wristow]

include/llvm/IR/Operator.h

187

We'll need to add flags to SDNodeFlags that are analogous to AllowReassoc and ApproxFunc. Adding them in a separate patch seems fine, but in case the lack of that change in this patch was an oversight, I wanted to raise the point here.

220

I don't know of a clean way. Definitely fine with me to submit it right after this patch is submitted.

lib/Transforms/Scalar/Reassociate.cpp

2018

Very minor point/question: Since the test is no longer hasUnsafeALgebra(), are we OK with the comment still saying unsafe algebra? Or do we want to change the comment above to something like:

// Don't optimize floating point instructions that don't have fast-math.

I'm fine leaving it as-is, but I've found these sorts of things in a handful of places, so if we want to change them, I'll look through the patch more thoroughly, and identify each one I find.

test/Assembler/fast-math-flags.ll

97

Is testing the 'reassoc' flag on the 'call' instruction really what you intended here? I would have thought add/sub/mul/div, but 'call' surprised me.

[hfinkel]

I'm happy; please go ahead when the other reviewers are too.

[spatel]

include/llvm/IR/Operator.h

187

I want to fix the backend too, but that's separate patches. For example, we don't propagate the IR flags properly yet (see D37686).

Also IIRC, the backend does not treat the global "-enable-unsafe-fp-math" as an umbrella for other flags. So that setting does not imply "-enable-no-nans-fp-math" or other FP relaxations.

lib/Transforms/Scalar/Reassociate.cpp

2018

I'll fix the comment to match the current code. Since this is the reassociation pass, I would guess that 'reassoc' is all we need to enable transforms here, but we'll have to verify that that is correct.

test/Assembler/fast-math-flags.ll

97

It is surprising but intentional. I want to highlight the fact that we allow any FMF on any FPMathOperator. So things like this or 'fadd arcp ...' are legal but I'm not sure how that would be used for optimization. We may want to refine that someday?

[wristow]

include/llvm/IR/Operator.h

187

I see. Thanks for clarifying.

lib/Transforms/Scalar/Reassociate.cpp

2018

To be clear, I'm not suggesting that in this patch we change the code here to check for 'reassoc' (i.e., I'm not suggesting we change the isFast() call to hasAllowReassoc() at this time). I view that as a separate piece of work, where we go through and carefully audit existing FMF-related checks, and decide how to use the more precise flags. Possibly it's just 'reassoc' that is needed for this case, or possibly it's 'reassoc' and some other conditions.

All I was suggesting by my comment/question, is that code-comments referring to a no longer-existing "unsafe algebra" umbrella flag, are a bit misleading. So I wondered whether we wanted to change those comments to better match the new implementation. It's a pretty minor point, in my view.

From my POV, the main purpose of this patch is to fix the underlying implementation to allow us to go through and do that audit, and fix issues like this "use isFast() or use some finer check?" example, here.

test/Assembler/fast-math-flags.ll

97

OK, thanks for explaining.

[spatel]

Patch updated:
Change code comments that reference 'unsafe' to the new 'fast' vocabulary. This is not condoning what may be a wrongful use of 'fast'; it's just trying to keep our code and comments in sync. I didn't go out of my way to look beyond a few lines in the diffs, so there may still be 'unsafe' refs and wrapper functions out there, but we'll squash those as we audit the usage of 'isFast'.

[wristow]

In D39304#915601, @spatel wrote:

Patch updated:
Change code comments that reference 'unsafe' to the new 'fast' vocabulary. ...

That covers my last questions.
LGTM.

[mcberg2017]

As an llvm vendor for Apple targets we have actually run into the incompatibility issues with old fast and current fast, we elected to use version markers in the IR for in house code to tell us that the old fast is relevant in the bit code reader during auto upgrade for IR into new Fast format.

[spatel]

In D39304#1001295, @mcberg2017 wrote:

As an llvm vendor for Apple targets we have actually run into the incompatibility issues with old fast and current fast, we elected to use version markers in the IR for in house code to tell us that the old fast is relevant in the bit code reader during auto upgrade for IR into new Fast format.

Thanks for letting me know. Did this manifest as a performance problem only or was there a visible functional difference too?

[qcolombet]

Hi Sanjay,

Did this manifest as a performance problem only or was there a visible functional difference too?

I guess it depends how you qualify visible functional difference :).
With the previous layout of the bits, reassociation was performed, but with the new layout, it is not.
The reason behind that reassociation checks isFast, which with the new layout is true only if all the 7 bits are set to true and that doesn't happen when the upgrade path is used (we get only 5 of them).

So this is a performance problem that is exposed by a functional difference in the compiler (an optimization was triggered and now it is not).

Given this new format was not released yet, do you see a way we could make the autoupgrade path to preserve the isFast semantic from the old format to the new format?

I haven't audited the code, but potential we are going to regress all the code that relied on isFast and given this is not publicly available yet (unless I am mistaken), I was wondering if there is a way to fix that.

Cheers,
-Quentin

[spatel]

In D39304#1002599, @qcolombet wrote:

Hi Sanjay,

Did this manifest as a performance problem only or was there a visible functional difference too?

I guess it depends how you qualify visible functional difference :).
With the previous layout of the bits, reassociation was performed, but with the new layout, it is not.
The reason behind that reassociation checks isFast, which with the new layout is true only if all the 7 bits are set to true and that doesn't happen when the upgrade path is used (we get only 5 of them).

So this is a performance problem that is exposed by a functional difference in the compiler (an optimization was triggered and now it is not).

Right - we expected that could happen with existing IR compiled with -ffast-math. I don't think we could go wrong in that scenario given that 'fast' gives us license to do all kinds of transforms, but doesn't require it - although people have different expectations once they get accustomed to the optimizations. :)

Given this new format was not released yet, do you see a way we could make the autoupgrade path to preserve the isFast semantic from the old format to the new format?
I haven't audited the code, but potential we are going to regress all the code that relied on isFast and given this is not publicly available yet (unless I am mistaken), I was wondering if there is a way to fix that.

Michael has this patch already, right? I think we have to create a new version of the IR since the bits changed meaning (we can't just flip 'on' new bits).

I have no objection to that, but that use case wasn't important to me, so that's why I didn't bother in this patch. AFAIK, this is new for v6.0, so yes, it's still possible to do this before that window closes.

[qcolombet]

I think we have to create a new version of the IR since the bits changed meaning (we can't just flip 'on' new bits).

Yeah, exactly.

I have no objection to that, but that use case wasn't important to me, so that's why I didn't bother in this patch. AFAIK, this is new for v6.0, so yes, it's still possible to do this before that window closes.

That would be ideal, but on the other hand, like you said, the new code won't be wrong.
I don't know what it takes to bump the bitcode version nor what are the implications, so maybe that's the right call, but let us have this conversation on LLVM dev.

Let me start a thread on that.

[qcolombet]

That's http://lists.llvm.org/pipermail/llvm-dev/2018-February/121114.html

[qcolombet]

BTW, while writing the RFC, I realized that we could potentially generated incorrect code if we were silently downgrade a post-r317488 bitcode with a pre-r317488 compiler. (I.e., running fast math optimizations whereas we only wanted reassoc)

docs/LangRef.rst

 seq\_cst total orderings of other operations that are not marked
 ``syncscope("singlethread")`` or ``syncscope("<target-scope>")``.
 
 .. _fastmath:
 
 Fast-Math Flags
 ---------------
 
-LLVM IR floating-point binary ops (:ref:`fadd <i_fadd>`,
+LLVM IR floating-point operations (:ref:`fadd <i_fadd>`,
 :ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
 :ref:`frem <i_frem>`, :ref:`fcmp <i_fcmp>`) and :ref:`call <i_call>`
-instructions have the following flags that can be set to enable
-otherwise unsafe floating point transformations.
+may use the following flags to enable otherwise unsafe 
+floating-point transformations.
 
 ``nnan``
    No NaNs - Allow optimizations to assume the arguments and result are not
    NaN. Such optimizations are required to retain defined behavior over
    NaNs, but the value of the result is undefined.
 
 ``ninf``
    No Infs - Allow optimizations to assume the arguments and result are not
    +/-Inf. Such optimizations are required to retain defined behavior over
    +/-Inf, but the value of the result is undefined.
 
 ``nsz``
    No Signed Zeros - Allow optimizations to treat the sign of a zero
    argument or result as insignificant.
 
 ``arcp``
    Allow Reciprocal - Allow optimizations to use the reciprocal of an
    argument rather than perform division.
 
 ``contract``
    Allow floating-point contraction (e.g. fusing a multiply followed by an
    addition into a fused multiply-and-add).
 
+``afn``
+   Approximate functions - Allow substitution of approximate calculations for
+   functions (sin, log, sqrt, etc). See floating-point intrinsic definitions 
+   for places where this can apply to LLVM's intrinsic math functions. 
+
+``reassoc``
+   Allow reassociation transformations for floating-point instructions. 
+   This may dramatically change results in floating point.
+
 ``fast``
-   Fast - Allow algebraically equivalent transformations that may
-   dramatically change results in floating point (e.g. reassociate). This
-   flag implies all the others.
+   This flag implies all of the others.
 
 .. _uselistorder:
 
 Use-list Order Directives
 -------------------------
 
 Use-list directives encode the in-memory order of each use-list, allowing the
 order to be recreated. ``<order-indexes>`` is a comma-separated list of

 """"""""""
 
 The argument and return value are floating point numbers of the same type.
 
 Semantics:
 """"""""""
 
 This function returns the square root of the operand if it is a nonnegative
-floating point number.
+floating point number. When specified with the fast-math-flag 'afn', the
+result may not match the system's libm implementation. 
 
 '``llvm.powi.*``' Intrinsic
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 Syntax:
 """""""
 
 This is an overloaded intrinsic. You can use ``llvm.powi`` on any

 
 The second argument is an integer power, and the first is a value to
 raise to that power.
 
 Semantics:
 """"""""""
 
 This function returns the first value raised to the second power with an
-unspecified sequence of rounding operations.
+unspecified sequence of rounding operations. When specified with the 
+fast-math-flag 'afn', the result may not match the system's libm 
+implementation.
 
 '``llvm.sin.*``' Intrinsic
 ^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 Syntax:
 """""""
 
 This is an overloaded intrinsic. You can use ``llvm.sin`` on any

 
 The argument and return value are floating point numbers of the same type.
 
 Semantics:
 """"""""""
 
 This function returns the sine of the specified operand, returning the
 same values as the libm ``sin`` functions would, and handles error
-conditions in the same way.
+conditions in the same way. When specified with the fast-math-flag 'afn', 
+the result may not match the system's libm implementation.
 
 '``llvm.cos.*``' Intrinsic
 ^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 Syntax:
 """""""
 
 This is an overloaded intrinsic. You can use ``llvm.cos`` on any

 
 The argument and return value are floating point numbers of the same type.
 
 Semantics:
 """"""""""
 
 This function returns the cosine of the specified operand, returning the
 same values as the libm ``cos`` functions would, and handles error
-conditions in the same way.
+conditions in the same way. When specified with the fast-math-flag 'afn', 
+the result may not match the system's libm implementation.
 
 '``llvm.pow.*``' Intrinsic
 ^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 Syntax:
 """""""
 
 This is an overloaded intrinsic. You can use ``llvm.pow`` on any

 
 The '``llvm.pow.*``' intrinsics return the first operand raised to the
 specified (positive or negative) power.
 
 Arguments:
 """"""""""
 
 The second argument is a floating point power, and the first is a value
-to raise to that power.
+to raise to that power. When specified with the fast-math-flag 'afn', the
+result may not match the system's libm implementation.
 
 Semantics:
 """"""""""
 
 This function returns the first value raised to the second power,
 returning the same values as the libm ``pow`` functions would, and
 handles error conditions in the same way.
 

 """"""""""
 
 The argument and return value are floating point numbers of the same type.
 
 Semantics:
 """"""""""
 
 This function returns the same values as the libm ``exp`` functions
-would, and handles error conditions in the same way.
+would, and handles error conditions in the same way. When specified with the 
+fast-math-flag 'afn', the result may not match the system's libm implementation.
 
 '``llvm.exp2.*``' Intrinsic
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 Syntax:
 """""""
 
 This is an overloaded intrinsic. You can use ``llvm.exp2`` on any

 """"""""""
 
 The argument and return value are floating point numbers of the same type.
 
 Semantics:
 """"""""""
 
 This function returns the same values as the libm ``exp2`` functions
-would, and handles error conditions in the same way.
+would, and handles error conditions in the same way. When specified with 
+the fast-math-flag 'afn', the result may not match the system's libm 
+implementation.
 
 '``llvm.log.*``' Intrinsic
 ^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 Syntax:
 """""""
 
 This is an overloaded intrinsic. You can use ``llvm.log`` on any

 """"""""""
 
 The argument and return value are floating point numbers of the same type.
 
 Semantics:
 """"""""""
 
 This function returns the same values as the libm ``log`` functions
-would, and handles error conditions in the same way.
+would, and handles error conditions in the same way. When specified with the 
+fast-math-flag 'afn', the result may not match the system's libm implementation.
 
 '``llvm.log10.*``' Intrinsic
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 Syntax:
 """""""
 
 This is an overloaded intrinsic. You can use ``llvm.log10`` on any

 """"""""""
 
 The argument and return value are floating point numbers of the same type.
 
 Semantics:
 """"""""""
 
 This function returns the same values as the libm ``log10`` functions
-would, and handles error conditions in the same way.
+would, and handles error conditions in the same way. When specified with 
+the fast-math-flag 'afn', the result may not match the system's libm 
+implementation.
 
 '``llvm.log2.*``' Intrinsic
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 Syntax:
 """""""
 
 This is an overloaded intrinsic. You can use ``llvm.log2`` on any

 """"""""""
 
 The argument and return value are floating point numbers of the same type.
 
 Semantics:
 """"""""""
 
 This function returns the same values as the libm ``log2`` functions
-would, and handles error conditions in the same way.
+would, and handles error conditions in the same way. When specified with the
+fast-math-flag 'afn', the result may not match the system's libm implementation.
 
 '``llvm.fma.*``' Intrinsic
 ^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 Syntax:
 """""""
 
 This is an overloaded intrinsic. You can use ``llvm.fma`` on any

 
 The argument and return value are floating point numbers of the same
 type.
 
 Semantics:
 """"""""""
 
 This function returns the same values as the libm ``fma`` functions
-would, and does not set errno.
+would, and does not set errno. When specified with the fast-math-flag 'afn', 
+the result may not match the system's libm implementation.
 
 '``llvm.fabs.*``' Intrinsic
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 Syntax:
 """""""
 
 This is an overloaded intrinsic. You can use ``llvm.fabs`` on any

 Follows the IEEE-754 semantics for minNum, which also match for libm's
 fmin.
 
 If either operand is a NaN, returns the other non-NaN operand. Returns
 NaN only if both operands are NaN. If the operands compare equal,
 returns a value that compares equal to both operands. This means that
 fmin(+/-0.0, +/-0.0) could return either -0.0 or 0.0.
 
+When specified with the fast-math-flag 'afn', the result may not match 
+the system's libm implementation.
+
 '``llvm.maxnum.*``' Intrinsic
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 Syntax:
 """""""
 
 This is an overloaded intrinsic. You can use ``llvm.maxnum`` on any
 floating point or vector of floating point type. Not all targets support

 Follows the IEEE-754 semantics for maxNum, which also match for libm's
 fmax.
 
 If either operand is a NaN, returns the other non-NaN operand. Returns
 NaN only if both operands are NaN. If the operands compare equal,
 returns a value that compares equal to both operands. This means that
 fmax(+/-0.0, +/-0.0) could return either -0.0 or 0.0.
 
+When specified with the fast-math-flag 'afn', the result may not match 
+the system's libm implementation.
+
 '``llvm.copysign.*``' Intrinsic
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 Syntax:
 """""""
 
 This is an overloaded intrinsic. You can use ``llvm.copysign`` on any
 floating point or vector of floating point type. Not all targets support

 
 The argument and return value are floating point numbers of the same
 type.
 
 Semantics:
 """"""""""
 
 This function returns the same values as the libm ``floor`` functions
-would, and handles error conditions in the same way.
+would, and handles error conditions in the same way. When specified with 
+the fast-math-flag 'afn', the result may not match the system's libm 
+implementation.
 
 '``llvm.ceil.*``' Intrinsic
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 Syntax:
 """""""
 
 This is an overloaded intrinsic. You can use ``llvm.ceil`` on any

 
 The argument and return value are floating point numbers of the same
 type.
 
 Semantics:
 """"""""""
 
 This function returns the same values as the libm ``ceil`` functions
-would, and handles error conditions in the same way.
+would, and handles error conditions in the same way. When specified with 
+the fast-math-flag 'afn', the result may not match the system's libm 
+implementation.
 
 '``llvm.trunc.*``' Intrinsic
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 Syntax:
 """""""
 
 This is an overloaded intrinsic. You can use ``llvm.trunc`` on any

 
 The argument and return value are floating point numbers of the same
 type.
 
 Semantics:
 """"""""""
 
 This function returns the same values as the libm ``trunc`` functions
-would, and handles error conditions in the same way.
+would, and handles error conditions in the same way. When specified with 
+the fast-math-flag 'afn', the result may not match the system's libm 
+implementation.
 
 '``llvm.rint.*``' Intrinsic
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 Syntax:
 """""""
 
 This is an overloaded intrinsic. You can use ``llvm.rint`` on any

 
 The argument and return value are floating point numbers of the same
 type.
 
 Semantics:
 """"""""""
 
 This function returns the same values as the libm ``rint`` functions
-would, and handles error conditions in the same way.
+would, and handles error conditions in the same way. When specified with the 
+fast-math-flag 'afn', the result may not match the system's libm implementation.
 
 '``llvm.nearbyint.*``' Intrinsic
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 Syntax:
 """""""
 
 This is an overloaded intrinsic. You can use ``llvm.nearbyint`` on any

 
 The '``llvm.nearbyint.*``' intrinsics returns the operand rounded to the
 nearest integer.
 
 Arguments:
 """"""""""
 
 The argument and return value are floating point numbers of the same
-type.
+type. When specified with the fast-math-flag 'afn', the result may not match 
+the system's libm implementation.
 
 Semantics:
 """"""""""
 
 This function returns the same values as the libm ``nearbyint``
 functions would, and handles error conditions in the same way.
 
 '``llvm.round.*``' Intrinsic

 
 The argument and return value are floating point numbers of the same
 type.
 
 Semantics:
 """"""""""
 
 This function returns the same values as the libm ``round``
-functions would, and handles error conditions in the same way.
+functions would, and handles error conditions in the same way. When 
+specified with the fast-math-flag 'afn', the result may not match 
+the system's libm implementation.
 
 Bit Manipulation Intrinsics
 ---------------------------
 
 LLVM provides intrinsics for a few important bit manipulation
 operations. These allow efficient code generation for some algorithms.
 
 '``llvm.bitreverse.*``' Intrinsics

include/llvm/IR/Instruction.h

public:
 
   /// Drops flags that may cause this instruction to evaluate to poison despite
   /// having non-poison inputs.
   void dropPoisonGeneratingFlags();
 
   /// Determine whether the exact flag is set.
   bool isExact() const;
 
-  /// Set or clear the unsafe-algebra flag on this instruction, which must be an
+  /// Set or clear all fast-math-flags on this instruction, which must be an
   /// operator which supports this flag. See LangRef.html for the meaning of
   /// this flag.
-  void setHasUnsafeAlgebra(bool B);
+  void setFast(bool B);
+
+  /// Set or clear the reassociation flag on this instruction, which must be
+  /// an operator which supports this flag. See LangRef.html for the meaning of
+  /// this flag.
+  void setHasAllowReassoc(bool B);
 
   /// Set or clear the no-nans flag on this instruction, which must be an
   /// operator which supports this flag. See LangRef.html for the meaning of
   /// this flag.
   void setHasNoNaNs(bool B);
 
   /// Set or clear the no-infs flag on this instruction, which must be an
   /// operator which supports this flag. See LangRef.html for the meaning of
   /// this flag.
   void setHasNoInfs(bool B);
 
   /// Set or clear the no-signed-zeros flag on this instruction, which must be
   /// an operator which supports this flag. See LangRef.html for the meaning of
   /// this flag.
   void setHasNoSignedZeros(bool B);
 
   /// Set or clear the allow-reciprocal flag on this instruction, which must be
   /// an operator which supports this flag. See LangRef.html for the meaning of
   /// this flag.
   void setHasAllowReciprocal(bool B);
 
+  /// Set or clear the approximate-math-functions flag on this instruction,
+  /// which must be an operator which supports this flag. See LangRef.html for
+  /// the meaning of this flag.
+  void setHasApproxFunc(bool B);
+
   /// Convenience function for setting multiple fast-math flags on this
   /// instruction, which must be an operator which supports these flags. See
   /// LangRef.html for the meaning of these flags.
   void setFastMathFlags(FastMathFlags FMF);
 
   /// Convenience function for transferring all fast-math flag values to this
   /// instruction, which must be an operator which supports these flags. See
   /// LangRef.html for the meaning of these flags.
   void copyFastMathFlags(FastMathFlags FMF);
 
-  /// Determine whether the unsafe-algebra flag is set.
-  bool hasUnsafeAlgebra() const;
+  /// Determine whether all fast-math-flags are set.
+  bool isFast() const;
+
+  /// Determine whether the allow-reassociation flag is set.
+  bool hasAllowReassoc() const;
 
   /// Determine whether the no-NaNs flag is set.
   bool hasNoNaNs() const;
 
   /// Determine whether the no-infs flag is set.
   bool hasNoInfs() const;
 
   /// Determine whether the no-signed-zeros flag is set.
   bool hasNoSignedZeros() const;
 
   /// Determine whether the allow-reciprocal flag is set.
   bool hasAllowReciprocal() const;
 
   /// Determine whether the allow-contract flag is set.
   bool hasAllowContract() const;
 
+  /// Determine whether the approximate-math-functions flag is set.
+  bool hasApproxFunc() const;
+
   /// Convenience function for getting all the fast-math flags, which must be an
   /// operator which supports these flags. See LangRef.html for the meaning of
   /// these flags.
   FastMathFlags getFastMathFlags() const;
 
   /// Copy I's fast-math flags
   void copyFastMathFlags(const Instruction *I);
 

include/llvm/IR/Operator.h

 
 /// Convenience struct for specifying and reasoning about fast-math flags.
 class FastMathFlags {
 private:
   friend class FPMathOperator;
 
   unsigned Flags = 0;
 
-  FastMathFlags(unsigned F) : Flags(F) { }
+  FastMathFlags(unsigned F) {
+    // If all 7 bits are set, turn this into -1. If the number of bits grows,
+    // this must be updated. This is intended to provide some forward binary
+    // compatibility insurance for the meaning of 'fast' in case bits are added.
+    if (F == 0x7F) Flags = ~0U;
+    else Flags = F;
+  }
 
 public:
-  /// This is how the bits are used in Value::SubclassOptionalData so they
-  /// should fit there too.
+  // This is how the bits are used in Value::SubclassOptionalData so they
+  // should fit there too.
+  // WARNING: We're out of space. SubclassOptionalData only has 7 bits. New
+  // functionality will require a change in how this information is stored.
   enum {
-    UnsafeAlgebra   = (1 << 0),
+    AllowReassoc    = (1 << 0),
     NoNaNs          = (1 << 1),
     NoInfs          = (1 << 2),
     NoSignedZeros   = (1 << 3),
     AllowReciprocal = (1 << 4),
-    AllowContract   = (1 << 5)
+    AllowContract   = (1 << 5),
+    ApproxFunc      = (1 << 6)
   };
 
   FastMathFlags() = default;
 
-  /// Whether any flag is set
   bool any() const { return Flags != 0; }
+  bool none() const { return Flags == 0; }
+  bool all() const { return Flags == ~0U; }
 
-  /// Set all the flags to false
   void clear() { Flags = 0; }
+  void set()   { Flags = ~0U; }
 
   /// Flag queries
+  bool allowReassoc() const    { return 0 != (Flags & AllowReassoc); }
   bool noNaNs() const          { return 0 != (Flags & NoNaNs); }
   bool noInfs() const          { return 0 != (Flags & NoInfs); }
   bool noSignedZeros() const   { return 0 != (Flags & NoSignedZeros); }
   bool allowReciprocal() const { return 0 != (Flags & AllowReciprocal); }
-  bool allowContract() const { return 0 != (Flags & AllowContract); }
-  bool unsafeAlgebra() const   { return 0 != (Flags & UnsafeAlgebra); }
+  bool allowContract() const   { return 0 != (Flags & AllowContract); }
+  bool approxFunc() const      { return 0 != (Flags & ApproxFunc); }
+  /// 'Fast' means all bits are set.
+  bool isFast() const          { return all(); }
 
   /// Flag setters
+  void setAllowReassoc()    { Flags |= AllowReassoc; }
   void setNoNaNs()          { Flags |= NoNaNs; }
   void setNoInfs()          { Flags |= NoInfs; }
   void setNoSignedZeros()   { Flags |= NoSignedZeros; }
   void setAllowReciprocal() { Flags |= AllowReciprocal; }
+  // TODO: Change the other set* functions to take a parameter?
   void setAllowContract(bool B) {
     Flags = (Flags & ~AllowContract) | B * AllowContract;
   }
-  void setUnsafeAlgebra() {
-    Flags |= UnsafeAlgebra;
-    setNoNaNs();
-    setNoInfs();
-    setNoSignedZeros();
-    setAllowReciprocal();
-    setAllowContract(true);
-  }
+  void setApproxFunc()      { Flags |= ApproxFunc; }
+  void setFast()            { set(); }
 
   void operator&=(const FastMathFlags &OtherFlags) {
     Flags &= OtherFlags.Flags;
   }
 };
 
 /// Utility class for floating point operations which can have
 /// information about relaxed accuracy requirements attached to them.
 class FPMathOperator : public Operator {
 private:
   friend class Instruction;
 
-  void setHasUnsafeAlgebra(bool B) {
-    SubclassOptionalData =
-      (SubclassOptionalData & ~FastMathFlags::UnsafeAlgebra) |
-      (B * FastMathFlags::UnsafeAlgebra);
-
-    // Unsafe algebra implies all the others
-    if (B) {
-      setHasNoNaNs(true);
-      setHasNoInfs(true);
-      setHasNoSignedZeros(true);
-      setHasAllowReciprocal(true);
-    }
+  /// 'Fast' means all bits are set.
+  void setFast(bool B) {
+    setHasAllowReassoc(B);
+    setHasNoNaNs(B);
+    setHasNoInfs(B);
+    setHasNoSignedZeros(B);
+    setHasAllowReciprocal(B);
+    setHasAllowContract(B);
+    setHasApproxFunc(B);
+
+  void setHasAllowReassoc(bool B) {
+    SubclassOptionalData =
+    (SubclassOptionalData & ~FastMathFlags::AllowReassoc) |
+    (B * FastMathFlags::AllowReassoc);
   }
 
   void setHasNoNaNs(bool B) {
     SubclassOptionalData =
       (SubclassOptionalData & ~FastMathFlags::NoNaNs) |
       (B * FastMathFlags::NoNaNs);
   }
 
private:
   }
 
   void setHasAllowContract(bool B) {
     SubclassOptionalData =
         (SubclassOptionalData & ~FastMathFlags::AllowContract) |
         (B * FastMathFlags::AllowContract);
   }
 
+  void setHasApproxFunc(bool B) {
+    SubclassOptionalData =
+        (SubclassOptionalData & ~FastMathFlags::ApproxFunc) |
+        (B * FastMathFlags::ApproxFunc);
+  }
+
   /// Convenience function for setting multiple fast-math flags.
   /// FMF is a mask of the bits to set.
   void setFastMathFlags(FastMathFlags FMF) {
     SubclassOptionalData |= FMF.Flags;
   }
 
   /// Convenience function for copying all fast-math flags.
   /// All values in FMF are transferred to this operator.
   void copyFastMathFlags(FastMathFlags FMF) {
     SubclassOptionalData = FMF.Flags;
   }
 
 public:
-  /// Test whether this operation is permitted to be
-  /// algebraically transformed, aka the 'A' fast-math property.
-  bool hasUnsafeAlgebra() const {
-    return (SubclassOptionalData & FastMathFlags::UnsafeAlgebra) != 0;
+  /// Test if this operation allows all non-strict floating-point transforms.
+  bool isFast() const {
+    return ((SubclassOptionalData & FastMathFlags::AllowReassoc) != 0 &&
+            (SubclassOptionalData & FastMathFlags::NoNaNs) != 0 &&
+            (SubclassOptionalData & FastMathFlags::NoInfs) != 0 &&
+            (SubclassOptionalData & FastMathFlags::NoSignedZeros) != 0 &&
+            (SubclassOptionalData & FastMathFlags::AllowReciprocal) != 0 &&
+            (SubclassOptionalData & FastMathFlags::AllowContract) != 0 &&
+            (SubclassOptionalData & FastMathFlags::ApproxFunc) != 0);
+  }
+
+  /// Test if this operation may be simplified with reassociative transforms.
+  bool hasAllowReassoc() const {
+    return (SubclassOptionalData & FastMathFlags::AllowReassoc) != 0;
   }
 
-  /// Test whether this operation's arguments and results are to be
-  /// treated as non-NaN, aka the 'N' fast-math property.
+  /// Test if this operation's arguments and results are assumed not-NaN.
   bool hasNoNaNs() const {
     return (SubclassOptionalData & FastMathFlags::NoNaNs) != 0;
   }
 
-  /// Test whether this operation's arguments and results are to be
-  /// treated as NoN-Inf, aka the 'I' fast-math property.
+  /// Test if this operation's arguments and results are assumed not-infinite.
   bool hasNoInfs() const {
     return (SubclassOptionalData & FastMathFlags::NoInfs) != 0;
   }
 
-  /// Test whether this operation can treat the sign of zero
-  /// as insignificant, aka the 'S' fast-math property.
+  /// Test if this operation can ignore the sign of zero.
   bool hasNoSignedZeros() const {
     return (SubclassOptionalData & FastMathFlags::NoSignedZeros) != 0;
   }
 
-  /// Test whether this operation is permitted to use
-  /// reciprocal instead of division, aka the 'R' fast-math property.
+  /// Test if this operation can use reciprocal multiply instead of division.
   bool hasAllowReciprocal() const {
     return (SubclassOptionalData & FastMathFlags::AllowReciprocal) != 0;
   }
 
-  /// Test whether this operation is permitted to
-  /// be floating-point contracted.
+  /// Test if this operation can be floating-point contracted (FMA).
   bool hasAllowContract() const {
     return (SubclassOptionalData & FastMathFlags::AllowContract) != 0;
   }
 
+  /// Test if this operation allows approximations of math library functions or
+  /// intrinsics.
+  bool hasApproxFunc() const {
+    return (SubclassOptionalData & FastMathFlags::ApproxFunc) != 0;
+  }
+
   /// Convenience function for getting all the fast-math flags
   FastMathFlags getFastMathFlags() const {
     return FastMathFlags(SubclassOptionalData);
   }
 
   /// Get the maximum error permitted by this operation in ULPs. An accuracy of
   /// 0.0 means that the operation should be performed with the default
   /// precision.

include/llvm/Transforms/Utils/LoopUtils.h

public:
   static bool isInductionPHI(PHINode *Phi, const Loop* L,
                              PredicatedScalarEvolution &PSE,
                              InductionDescriptor &D, bool Assume = false);
 
   /// Returns true if the induction type is FP and the binary operator does
   /// not have the "fast-math" property. Such operation requires a relaxed FP
   /// mode.
   bool hasUnsafeAlgebra() {
-    return InductionBinOp &&
-      !cast<FPMathOperator>(InductionBinOp)->hasUnsafeAlgebra();
+    return InductionBinOp && !cast<FPMathOperator>(InductionBinOp)->isFast();
   }
 
   /// Returns induction operator that does not have "fast-math" property
   /// and requires FP unsafe mode.
   Instruction *getUnsafeAlgebraInst() {
-    if (!InductionBinOp ||
-        cast<FPMathOperator>(InductionBinOp)->hasUnsafeAlgebra())
+    if (!InductionBinOp || cast<FPMathOperator>(InductionBinOp)->isFast())
       return nullptr;
     return InductionBinOp;
   }
 
   /// Returns binary opcode of the induction operator.
   Instruction::BinaryOps getInductionOpcode() const {
     return InductionBinOp ? InductionBinOp->getOpcode() :
       Instruction::BinaryOpsEnd;

lib/AsmParser/LLLexer.cpp

#define KEYWORD(STR)                                                           \
   KEYWORD(seq_cst);
   KEYWORD(syncscope);
 
   KEYWORD(nnan);
   KEYWORD(ninf);
   KEYWORD(nsz);
   KEYWORD(arcp);
   KEYWORD(contract);
+  KEYWORD(reassoc);
+  KEYWORD(afn);
   KEYWORD(fast);
   KEYWORD(nuw);
   KEYWORD(nsw);
   KEYWORD(exact);
   KEYWORD(inbounds);
   KEYWORD(inrange);
   KEYWORD(align);
   KEYWORD(addrspace);

lib/AsmParser/LLParser.h

bool EatIfPresent(lltok::Kind T) {
       Lex.Lex();
       return true;
     }
 
     FastMathFlags EatFastMathFlagsIfPresent() {
       FastMathFlags FMF;
       while (true)
         switch (Lex.getKind()) {
-        case lltok::kw_fast: FMF.setUnsafeAlgebra();   Lex.Lex(); continue;
+        case lltok::kw_fast: FMF.setFast();            Lex.Lex(); continue;
         case lltok::kw_nnan: FMF.setNoNaNs();          Lex.Lex(); continue;
         case lltok::kw_ninf: FMF.setNoInfs();          Lex.Lex(); continue;
         case lltok::kw_nsz:  FMF.setNoSignedZeros();   Lex.Lex(); continue;
         case lltok::kw_arcp: FMF.setAllowReciprocal(); Lex.Lex(); continue;
         case lltok::kw_contract:
           FMF.setAllowContract(true);
           Lex.Lex();
           continue;
+        case lltok::kw_reassoc: FMF.setAllowReassoc(); Lex.Lex(); continue;
+        case lltok::kw_afn:     FMF.setApproxFunc();   Lex.Lex(); continue;
         default: return FMF;
         }
       return FMF;
     }
 
     bool ParseOptionalToken(lltok::Kind T, bool &Present,
                             LocTy *Loc = nullptr) {
       if (Lex.getKind() != T) {

lib/AsmParser/LLToken.h

enum Kind {
   kw_acq_rel,
   kw_seq_cst,
   kw_syncscope,
   kw_nnan,
   kw_ninf,
   kw_nsz,
   kw_arcp,
   kw_contract,
+  kw_reassoc,
+  kw_afn,
   kw_fast,
   kw_nuw,
   kw_nsw,
   kw_exact,
   kw_inbounds,
   kw_inrange,
   kw_align,
   kw_addrspace,

lib/Bitcode/Reader/BitcodeReader.cpp

case bitc::COMDAT_SELECTION_KIND_NO_DUPLICATES:
     return Comdat::NoDuplicates;
   case bitc::COMDAT_SELECTION_KIND_SAME_SIZE:
     return Comdat::SameSize;
   }
 }
 
 static FastMathFlags getDecodedFastMathFlags(unsigned Val) {
   FastMathFlags FMF;
-  if (0 != (Val & FastMathFlags::UnsafeAlgebra))
-    FMF.setUnsafeAlgebra();
+  if (0 != (Val & FastMathFlags::AllowReassoc))
+    FMF.setAllowReassoc();
   if (0 != (Val & FastMathFlags::NoNaNs))
     FMF.setNoNaNs();
   if (0 != (Val & FastMathFlags::NoInfs))
     FMF.setNoInfs();
   if (0 != (Val & FastMathFlags::NoSignedZeros))
     FMF.setNoSignedZeros();
   if (0 != (Val & FastMathFlags::AllowReciprocal))
     FMF.setAllowReciprocal();
   if (0 != (Val & FastMathFlags::AllowContract))
     FMF.setAllowContract(true);
+  if (0 != (Val & FastMathFlags::ApproxFunc))
+    FMF.setApproxFunc();
   return FMF;
 }
 
 static void upgradeDLLImportExportLinkage(GlobalValue *GV, unsigned Val) {
   switch (Val) {
   case 5: GV->setDLLStorageClass(GlobalValue::DLLImportStorageClass); break;
   case 6: GV->setDLLStorageClass(GlobalValue::DLLExportStorageClass); break;
   }

lib/Bitcode/Writer/BitcodeWriter.cpp

if (const auto *OBO = dyn_cast<OverflowingBinaryOperator>(V)) {
     if (OBO->hasNoSignedWrap())
       Flags |= 1 << bitc::OBO_NO_SIGNED_WRAP;
     if (OBO->hasNoUnsignedWrap())
       Flags |= 1 << bitc::OBO_NO_UNSIGNED_WRAP;
   } else if (const auto *PEO = dyn_cast<PossiblyExactOperator>(V)) {
     if (PEO->isExact())
       Flags |= 1 << bitc::PEO_EXACT;
   } else if (const auto *FPMO = dyn_cast<FPMathOperator>(V)) {
-    if (FPMO->hasUnsafeAlgebra())
-      Flags |= FastMathFlags::UnsafeAlgebra;
+    if (FPMO->hasAllowReassoc())
+      Flags |= FastMathFlags::AllowReassoc;
     if (FPMO->hasNoNaNs())
       Flags |= FastMathFlags::NoNaNs;
     if (FPMO->hasNoInfs())
       Flags |= FastMathFlags::NoInfs;
     if (FPMO->hasNoSignedZeros())
       Flags |= FastMathFlags::NoSignedZeros;
     if (FPMO->hasAllowReciprocal())
       Flags |= FastMathFlags::AllowReciprocal;
     if (FPMO->hasAllowContract())
       Flags |= FastMathFlags::AllowContract;
+    if (FPMO->hasApproxFunc())
+      Flags |= FastMathFlags::ApproxFunc;
   }
 
   return Flags;
 }
 
 void ModuleBitcodeWriter::writeValueAsMetadata(
     const ValueAsMetadata *MD, SmallVectorImpl<uint64_t> &Record) {
   // Mimic an MDNode with a value as one operand.

lib/CodeGen/ExpandReductions.cpp

for (auto *II : Worklist) {
     auto MRK = RecurrenceDescriptor::MRK_Invalid;
     switch (ID) {
     case Intrinsic::experimental_vector_reduce_fadd:
     case Intrinsic::experimental_vector_reduce_fmul:
       // FMFs must be attached to the call, otherwise it's an ordered reduction
       // and it can't be handled by generating this shuffle sequence.
       // TODO: Implement scalarization of ordered reductions here for targets
       // without native support.
-      if (!II->getFastMathFlags().unsafeAlgebra())
+      if (!II->getFastMathFlags().isFast())
         continue;
       Vec = II->getArgOperand(1);
       break;
     case Intrinsic::experimental_vector_reduce_add:
     case Intrinsic::experimental_vector_reduce_mul:
     case Intrinsic::experimental_vector_reduce_and:
     case Intrinsic::experimental_vector_reduce_or:
     case Intrinsic::experimental_vector_reduce_xor:

lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

static bool isVectorReductionOp(const User *I) {
   case Instruction::Mul:
   case Instruction::And:
   case Instruction::Or:
   case Instruction::Xor:
     break;
   case Instruction::FAdd:
   case Instruction::FMul:
     if (const FPMathOperator *FPOp = dyn_cast<const FPMathOperator>(Inst))
-      if (FPOp->getFastMathFlags().unsafeAlgebra())
+      if (FPOp->getFastMathFlags().isFast())
         break;
     LLVM_FALLTHROUGH;
   default:
     return false;
   }
 
   unsigned ElemNum = Inst->getType()->getVectorNumElements();
   unsigned ElemNumToReduce = ElemNum;
while (!UsersToVisit.empty()) {
 
     for (const auto &U : User->users()) {
       auto Inst = dyn_cast<Instruction>(U);
       if (!Inst)
         return false;
 
       if (Inst->getOpcode() == OpCode || isa<PHINode>(U)) {
         if (const FPMathOperator *FPOp = dyn_cast<const FPMathOperator>(Inst))
-          if (!isa<PHINode>(FPOp) && !FPOp->getFastMathFlags().unsafeAlgebra())
+          if (!isa<PHINode>(FPOp) && !FPOp->getFastMathFlags().isFast())
             return false;
         UsersToVisit.push_back(U);
       } else if (const ShuffleVectorInst *ShufInst =
                      dyn_cast<ShuffleVectorInst>(U)) {
         // Detect the following pattern: A ShuffleVector instruction together
         // with a reduction that do partial reduction on the first and second
         // ElemNumToReduce / 2 elements, and store the result in
         // ElemNumToReduce / 2 elements in another vector.
void SelectionDAGBuilder::visitBinary(const User &I, unsigned OpCode) {
   Flags.setNoSignedWrap(nsw);
   Flags.setNoUnsignedWrap(nuw);
   Flags.setVectorReduction(vec_redux);
   Flags.setAllowReciprocal(FMF.allowReciprocal());
   Flags.setAllowContract(FMF.allowContract());
   Flags.setNoInfs(FMF.noInfs());
   Flags.setNoNaNs(FMF.noNaNs());
   Flags.setNoSignedZeros(FMF.noSignedZeros());
-  Flags.setUnsafeAlgebra(FMF.unsafeAlgebra());
+  Flags.setUnsafeAlgebra(FMF.isFast());
 
   SDValue BinNodeValue = DAG.getNode(OpCode, getCurSDLoc(), Op1.getValueType(),
                                      Op1, Op2, Flags);
   setValue(&I, BinNodeValue);
 }
 
 void SelectionDAGBuilder::visitShift(const User &I, unsigned Opcode) {
   SDValue Op1 = getValue(I.getOperand(0));
void SelectionDAGBuilder::visitVectorReduce(const CallInst &I,
   FastMathFlags FMF;
   if (isa<FPMathOperator>(I))
     FMF = I.getFastMathFlags();
   SDNodeFlags SDFlags;
   SDFlags.setNoNaNs(FMF.noNaNs());
 
   switch (Intrinsic) {
   case Intrinsic::experimental_vector_reduce_fadd:
-    if (FMF.unsafeAlgebra())
+    if (FMF.isFast())
       Res = DAG.getNode(ISD::VECREDUCE_FADD, dl, VT, Op2);
     else
       Res = DAG.getNode(ISD::VECREDUCE_STRICT_FADD, dl, VT, Op1, Op2);
     break;
   case Intrinsic::experimental_vector_reduce_fmul:
-    if (FMF.unsafeAlgebra())
+    if (FMF.isFast())
       Res = DAG.getNode(ISD::VECREDUCE_FMUL, dl, VT, Op2);
     else
       Res = DAG.getNode(ISD::VECREDUCE_STRICT_FMUL, dl, VT, Op1, Op2);
     break;
   case Intrinsic::experimental_vector_reduce_add:
     Res = DAG.getNode(ISD::VECREDUCE_ADD, dl, VT, Op1);
     break;
   case Intrinsic::experimental_vector_reduce_mul:

lib/IR/AsmWriter.cpp

static void writeAtomicRMWOperation(raw_ostream &Out,
   case AtomicRMWInst::Min:  Out << " min"; break;
   case AtomicRMWInst::UMax: Out << " umax"; break;
   case AtomicRMWInst::UMin: Out << " umin"; break;
   }
 }
 
 static void WriteOptimizationInfo(raw_ostream &Out, const User *U) {
   if (const FPMathOperator *FPO = dyn_cast<const FPMathOperator>(U)) {
-    // Unsafe algebra implies all the others, no need to write them all out
-    if (FPO->hasUnsafeAlgebra())
+    // 'Fast' is an abbreviation for all fast-math-flags.
+    if (FPO->isFast())
       Out << " fast";
     else {
+      if (FPO->hasAllowReassoc())
+        Out << " reassoc";
       if (FPO->hasNoNaNs())
         Out << " nnan";
       if (FPO->hasNoInfs())
         Out << " ninf";
       if (FPO->hasNoSignedZeros())
         Out << " nsz";
       if (FPO->hasAllowReciprocal())
         Out << " arcp";
       if (FPO->hasAllowContract())
         Out << " contract";
+      if (FPO->hasApproxFunc())
+        Out << " afn";
     }
   }
 
   if (const OverflowingBinaryOperator *OBO =
         dyn_cast<OverflowingBinaryOperator>(U)) {
     if (OBO->hasNoUnsignedWrap())
       Out << " nuw";
     if (OBO->hasNoSignedWrap())

lib/IR/Instruction.cpp

case Instruction::GetElementPtr:
     break;
   }
 }
 
 bool Instruction::isExact() const {
   return cast<PossiblyExactOperator>(this)->isExact();
 }
 
-void Instruction::setHasUnsafeAlgebra(bool B) {
+void Instruction::setFast(bool B) {
   assert(isa<FPMathOperator>(this) && "setting fast-math flag on invalid op");
-  cast<FPMathOperator>(this)->setHasUnsafeAlgebra(B);
+  cast<FPMathOperator>(this)->setFast(B);
+}
+
+void Instruction::setHasAllowReassoc(bool B) {
+  assert(isa<FPMathOperator>(this) && "setting fast-math flag on invalid op");
+  cast<FPMathOperator>(this)->setHasAllowReassoc(B);
 }
 
 void Instruction::setHasNoNaNs(bool B) {
   assert(isa<FPMathOperator>(this) && "setting fast-math flag on invalid op");
   cast<FPMathOperator>(this)->setHasNoNaNs(B);
 }
 
 void Instruction::setHasNoInfs(bool B) {
   assert(isa<FPMathOperator>(this) && "setting fast-math flag on invalid op");
   cast<FPMathOperator>(this)->setHasNoInfs(B);
 }
 
 void Instruction::setHasNoSignedZeros(bool B) {
   assert(isa<FPMathOperator>(this) && "setting fast-math flag on invalid op");
   cast<FPMathOperator>(this)->setHasNoSignedZeros(B);
 }
 
 void Instruction::setHasAllowReciprocal(bool B) {
   assert(isa<FPMathOperator>(this) && "setting fast-math flag on invalid op");
   cast<FPMathOperator>(this)->setHasAllowReciprocal(B);
 }
 
+void Instruction::setHasApproxFunc(bool B) {
+  assert(isa<FPMathOperator>(this) && "setting fast-math flag on invalid op");
+  cast<FPMathOperator>(this)->setHasApproxFunc(B);
+}
+
 void Instruction::setFastMathFlags(FastMathFlags FMF) {
   assert(isa<FPMathOperator>(this) && "setting fast-math flag on invalid op");
   cast<FPMathOperator>(this)->setFastMathFlags(FMF);
 }
 
 void Instruction::copyFastMathFlags(FastMathFlags FMF) {
   assert(isa<FPMathOperator>(this) && "copying fast-math flag on invalid op");
   cast<FPMathOperator>(this)->copyFastMathFlags(FMF);
 }
 
-bool Instruction::hasUnsafeAlgebra() const {
+bool Instruction::isFast() const {
   assert(isa<FPMathOperator>(this) && "getting fast-math flag on invalid op");
-  return cast<FPMathOperator>(this)->hasUnsafeAlgebra();
+  return cast<FPMathOperator>(this)->isFast();
+}
+
+bool Instruction::hasAllowReassoc() const {
+  assert(isa<FPMathOperator>(this) && "getting fast-math flag on invalid op");
+  return cast<FPMathOperator>(this)->hasAllowReassoc();
 }
 
 bool Instruction::hasNoNaNs() const {
   assert(isa<FPMathOperator>(this) && "getting fast-math flag on invalid op");
   return cast<FPMathOperator>(this)->hasNoNaNs();
 }
 
 bool Instruction::hasNoInfs() const {
bool Instruction::hasAllowReciprocal() const {
   return cast<FPMathOperator>(this)->hasAllowReciprocal();
 }
 
 bool Instruction::hasAllowContract() const {
   assert(isa<FPMathOperator>(this) && "getting fast-math flag on invalid op");
   return cast<FPMathOperator>(this)->hasAllowContract();
 }
 
+bool Instruction::hasApproxFunc() const {
+  assert(isa<FPMathOperator>(this) && "getting fast-math flag on invalid op");
+  return cast<FPMathOperator>(this)->hasApproxFunc();
+}
+
 FastMathFlags Instruction::getFastMathFlags() const {
   assert(isa<FPMathOperator>(this) && "getting fast-math flag on invalid op");
   return cast<FPMathOperator>(this)->getFastMathFlags();
 }
 
 void Instruction::copyFastMathFlags(const Instruction *I) {
   copyFastMathFlags(I->getFastMathFlags());
 }

 bool Instruction::isAssociative() const {
   unsigned Opcode = getOpcode();
   if (isAssociative(Opcode))
     return true;
 
   switch (Opcode) {
   case FMul:
   case FAdd:
-    return cast<FPMathOperator>(this)->hasUnsafeAlgebra();
+    return cast<FPMathOperator>(this)->isFast();
   default:
     return false;
   }
 }
 
 Instruction *Instruction::cloneImpl() const {
   llvm_unreachable("Subclass of Instruction failed to implement cloneImpl");
 }

lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp

if (!FPMath)
     return false;
 
   const FPMathOperator *FPOp = cast<const FPMathOperator>(&FDiv);
   float ULP = FPOp->getFPAccuracy();
   if (ULP < 2.5f)
     return false;
 
   FastMathFlags FMF = FPOp->getFastMathFlags();
-  bool UnsafeDiv = HasUnsafeFPMath || FMF.unsafeAlgebra() ||
+  bool UnsafeDiv = HasUnsafeFPMath || FMF.isFast() ||
                                       FMF.allowReciprocal();
 
   // With UnsafeDiv node will be optimized to just rcp and mul.
   if (ST->hasFP32Denormals() || UnsafeDiv)
     return false;
 
   IRBuilder<> Builder(FDiv.getParent(), std::next(FDiv.getIterator()), FPMath);
   Builder.setFastMathFlags(FMF);

lib/Target/AMDGPU/AMDGPULibCalls.cpp

 
 bool AMDGPULibCalls::parseFunctionName(const StringRef& FMangledName,
                                     FuncInfo *FInfo) {
   return AMDGPULibFunc::parse(FMangledName, *FInfo);
 }
 
 bool AMDGPULibCalls::isUnsafeMath(const CallInst *CI) const {
   if (auto Op = dyn_cast<FPMathOperator>(CI))
-    if (Op->hasUnsafeAlgebra())
+    if (Op->isFast())
       return true;
   const Function *F = CI->getParent()->getParent();
   Attribute Attr = F->getFnAttribute("unsafe-fp-math");
   return Attr.getValueAsString() == "true";
 }
 
 bool AMDGPULibCalls::useNativeFunc(const StringRef F) const {
   return AllNative ||

lib/Transforms/InstCombine/InstCombineAddSub.cpp

if (isMpy) {
     AddSub0 = Opnd0_0;
     AddSub1 = Opnd1_0;
   }
 
   if (!Factor)
     return nullptr;
 
   FastMathFlags Flags;
-  Flags.setUnsafeAlgebra();
+  Flags.setFast();
   if (I0) Flags &= I->getFastMathFlags();
   if (I1) Flags &= I->getFastMathFlags();
 
   // Create expression "NewAddSub = AddSub0 +/- AddsSub1"
   Value *NewAddSub = (I->getOpcode() == Instruction::FAdd) ?
                       createFAdd(AddSub0, AddSub1) :
                       createFSub(AddSub0, AddSub1);
   if (ConstantFP *CFP = dyn_cast<ConstantFP>(NewAddSub)) {
Value *FAddCombine::performFactorization(Instruction *I) {
 
   Value *RI = createFDiv(NewAddSub, Factor);
   if (Instruction *II = dyn_cast<Instruction>(RI))
     II->setFastMathFlags(Flags);
   return RI;
 }
 
 Value *FAddCombine::simplify(Instruction *I) {
-  assert(I->hasUnsafeAlgebra() && "Should be in unsafe mode");
+  assert(I->isFast() && "Should be in unsafe mode");
 
   // Currently we are not able to handle vector type.
   if (I->getType()->isVectorTy())
     return nullptr;
 
   assert((I->getOpcode() == Instruction::FAdd ||
           I->getOpcode() == Instruction::FSub) && "Expect add/sub");
 
if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
       }
     }
   }
 
   // Handle specials cases for FAdd with selects feeding the operation
   if (Value *V = SimplifySelectsFeedingBinaryOp(I, LHS, RHS))
     return replaceInstUsesWith(I, V);
 
-  if (I.hasUnsafeAlgebra()) {
+  if (I.isFast()) {
     if (Value *V = FAddCombine(Builder).simplify(&I))
       return replaceInstUsesWith(I, V);
   }
 
   return Changed ? &I : nullptr;
 }
 
 /// Optimize pointer differences into the same array into a size.  Consider:
if (Value *V = dyn_castFNegVal(FPEI->getOperand(0))) {
       return NewI;
     }
   }
 
   // Handle specials cases for FSub with selects feeding the operation
   if (Value *V = SimplifySelectsFeedingBinaryOp(I, Op0, Op1))
     return replaceInstUsesWith(I, V);
 
-  if (I.hasUnsafeAlgebra()) {
+  if (I.isFast()) {
     if (Value *V = FAddCombine(Builder).simplify(&I))
       return replaceInstUsesWith(I, V);
   }
 
   return nullptr;
 }

lib/Transforms/InstCombine/InstCombineCalls.cpp

if (isa<ConstantFP>(Arg0) && !isa<ConstantFP>(Arg1)) {
       return II;
     }
     if (Value *V = simplifyMinnumMaxnum(*II))
       return replaceInstUsesWith(*II, V);
     break;
   }
   case Intrinsic::fmuladd: {
     // Canonicalize fast fmuladd to the separate fmul + fadd.
-    if (II->hasUnsafeAlgebra()) {
+    if (II->isFast()) {
       BuilderTy::FastMathFlagGuard Guard(Builder);
       Builder.setFastMathFlags(II->getFastMathFlags());
       Value *Mul = Builder.CreateFMul(II->getArgOperand(0),
                                       II->getArgOperand(1));
       Value *Add = Builder.CreateFAdd(Mul, II->getArgOperand(2));
       Add->takeName(II);
       return replaceInstUsesWith(*II, Add);
     }

lib/Transforms/InstCombine/InstCombineMulDivRem.cpp

 /// Detect pattern log2(Y * 0.5) with corresponding fast math flags.
 static void detectLog2OfHalf(Value *&Op, Value *&Y, IntrinsicInst *&Log2) {
   if (!Op->hasOneUse())
     return;
 
   IntrinsicInst *II = dyn_cast<IntrinsicInst>(Op);
   if (!II)
     return;
-  if (II->getIntrinsicID() != Intrinsic::log2 || !II->hasUnsafeAlgebra())
+  if (II->getIntrinsicID() != Intrinsic::log2 || !II->isFast())
     return;
   Log2 = II;
 
   Value *OpLog2Of = II->getArgOperand(0);
   if (!OpLog2Of->hasOneUse())
     return;
 
   Instruction *I = dyn_cast<Instruction>(OpLog2Of);
   if (!I)
     return;
-  if (I->getOpcode() != Instruction::FMul || !I->hasUnsafeAlgebra())
+
+  if (I->getOpcode() != Instruction::FMul || !I->isFast())
     return;
 
   if (match(I->getOperand(0), m_SpecificFP(0.5)))
     Y = I->getOperand(1);
   else if (match(I->getOperand(1), m_SpecificFP(0.5)))
     Y = I->getOperand(0);
 }
 
if (C0) {
         Constant *F = ConstantExpr::getFDiv(C1, C);
         if (isNormalFp(F))
           R = BinaryOperator::CreateFDiv(Opnd0, F);
       }
     }
   }
 
   if (R) {
-    R->setHasUnsafeAlgebra(true);
+    R->setFast(true);
     InsertNewInstWith(R, *InsertBefore);
   }
 
   return R;
 }
 
 Instruction *InstCombiner::visitFMul(BinaryOperator &I) {
   bool Changed = SimplifyAssociativeOrCommutative(I);
   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
 
   if (Value *V = SimplifyVectorOp(I))
     return replaceInstUsesWith(I, V);
 
   if (isa<Constant>(Op0))
     std::swap(Op0, Op1);
 
   if (Value *V = SimplifyFMulInst(Op0, Op1, I.getFastMathFlags(),
                                   SQ.getWithInstruction(&I)))
     return replaceInstUsesWith(I, V);
 
-  bool AllowReassociate = I.hasUnsafeAlgebra();
+  bool AllowReassociate = I.isFast();
 
   // Simplify mul instructions with a constant RHS.
   if (isa<Constant>(Op1)) {
     if (Instruction *FoldedMul = foldOpWithConstantIntoOperand(I))
       return FoldedMul;
 
     // (fmul X, -1.0) --> (fsub -0.0, X)
     if (match(Op1, m_SpecificFP(-1.0))) {
if (Value *V = SimplifyFDivInst(Op0, Op1, I.getFastMathFlags(),
                                   SQ.getWithInstruction(&I)))
     return replaceInstUsesWith(I, V);
 
   if (isa<Constant>(Op0))
     if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
       if (Instruction *R = FoldOpIntoSelect(I, SI))
         return R;
 
-  bool AllowReassociate = I.hasUnsafeAlgebra();
+  bool AllowReassociate = I.isFast();
   bool AllowReciprocal = I.hasAllowReciprocal();
 
   if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
     if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
       if (Instruction *R = FoldOpIntoSelect(I, SI))
         return R;
 
     if (AllowReassociate) {

lib/Transforms/Scalar/Reassociate.cpp

XorOpnd::XorOpnd(Value *V) {
   isOr = true;
 }
 
 /// Return true if V is an instruction of the specified opcode and if it
 /// only has one use.
 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) {
   if (V->hasOneUse() && isa<Instruction>(V) &&
       cast<Instruction>(V)->getOpcode() == Opcode &&
-      (!isa<FPMathOperator>(V) ||
-       cast<Instruction>(V)->hasUnsafeAlgebra()))
+      (!isa<FPMathOperator>(V) || cast<Instruction>(V)->isFast()))
     return cast<BinaryOperator>(V);
   return nullptr;
 }
 
 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode1,
                                         unsigned Opcode2) {
   if (V->hasOneUse() && isa<Instruction>(V) &&
       (cast<Instruction>(V)->getOpcode() == Opcode1 ||
        cast<Instruction>(V)->getOpcode() == Opcode2) &&
-      (!isa<FPMathOperator>(V) ||
-       cast<Instruction>(V)->hasUnsafeAlgebra()))
+      (!isa<FPMathOperator>(V) || cast<Instruction>(V)->isFast()))
     return cast<BinaryOperator>(V);
   return nullptr;
 }
 
 void ReassociatePass::BuildRankMap(Function &F,
                                    ReversePostOrderTraversal<Function*> &RPOT) {
   unsigned Rank = 2;
 
#endif
 
       // At this point we have a value which, first of all, is not a binary
       // expression of the right kind, and secondly, is only used inside the
       // expression.  This means that it can safely be modified.  See if we
       // can usefully morph it into an expression of the right kind.
       assert((!isa<Instruction>(Op) ||
               cast<Instruction>(Op)->getOpcode() != Opcode
               || (isa<FPMathOperator>(Op) &&
-                  !cast<Instruction>(Op)->hasUnsafeAlgebra())) &&
+                  !cast<Instruction>(Op)->isFast())) &&
              "Should have been handled above!");
       assert(Op->hasOneUse() && "Has uses outside the expression tree!");
 
       // If this is a multiply expression, turn any internal negations into
       // multiplies by -1 so they can be reassociated.
       if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op))
         if ((Opcode == Instruction::Mul && BinaryOperator::isNeg(BO)) ||
             (Opcode == Instruction::FMul && BinaryOperator::isFNeg(BO))) {
if (Instruction *Res = canonicalizeNegConstExpr(I))
     I = Res;
 
   // Commute binary operators, to canonicalize the order of their operands.
   // This can potentially expose more CSE opportunities, and makes writing other
   // transformations simpler.
   if (I->isCommutative())
     canonicalizeOperands(I);
 
   // Don't optimize floating point instructions that don't have unsafe algebra.
-  if (I->getType()->isFPOrFPVectorTy() && !I->hasUnsafeAlgebra())
+  if (I->getType()->isFPOrFPVectorTy() && !I->isFast())
     return;
 
   // Do not reassociate boolean (i1) expressions.  We want to preserve the
   // original order of evaluation for short-circuited comparisons that
   // SimplifyCFG has folded to AND/OR expressions.  If the expression
   // is not further optimized, it is likely to be transformed back to a
   // short-circuited form for code gen, and the source order may have been
   // optimized for the most likely conditions.

lib/Transforms/Utils/LoopUtils.cpp

RecurrenceDescriptor::isMinMaxSelectCmpPattern(Instruction *I, InstDesc &Prev) {
   return InstDesc(false, I);
 }
 
 RecurrenceDescriptor::InstDesc
 RecurrenceDescriptor::isRecurrenceInstr(Instruction *I, RecurrenceKind Kind,
                                         InstDesc &Prev, bool HasFunNoNaNAttr) {
   bool FP = I->getType()->isFloatingPointTy();
   Instruction *UAI = Prev.getUnsafeAlgebraInst();
-  if (!UAI && FP && !I->hasUnsafeAlgebra())
+  if (!UAI && FP && !I->isFast())
     UAI = I; // Found an unsafe (unvectorizable) algebra instruction.
 
   switch (I->getOpcode()) {
   default:
     return InstDesc(false, I);
   case Instruction::PHI:
     return InstDesc(I, Prev.getMinMaxKind(), Prev.getUnsafeAlgebraInst());
   case Instruction::Sub:
case MRK_FloatMax:
     P = CmpInst::FCMP_OGT;
     break;
   }
 
   // We only match FP sequences with unsafe algebra, so we can unconditionally
   // set it on any generated instructions.
   IRBuilder<>::FastMathFlagGuard FMFG(Builder);
   FastMathFlags FMF;
-  FMF.setUnsafeAlgebra();
+  FMF.setFast();
   Builder.setFastMathFlags(FMF);
 
   Value *Cmp;
   if (RK == MRK_FloatMin || RK == MRK_FloatMax)
     Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
   else
     Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
 
assert(InductionBinOp &&
            (InductionBinOp->getOpcode() == Instruction::FAdd ||
             InductionBinOp->getOpcode() == Instruction::FSub) &&
            "Original bin op should be defined for FP induction");
 
     Value *StepValue = cast<SCEVUnknown>(Step)->getValue();
 
     // Floating point operations had to be 'fast' to enable the induction.
     FastMathFlags Flags;
-    Flags.setUnsafeAlgebra();
+    Flags.setFast();
 
     Value *MulExp = B.CreateFMul(StepValue, Index);
     if (isa<Instruction>(MulExp))
       // We have to check, the MulExp may be a constant.
       cast<Instruction>(MulExp)->setFastMathFlags(Flags);
 
     Value *BOp = B.CreateBinOp(InductionBinOp->getOpcode() , StartValue,
                                MulExp, "induction");
Optional<unsigned> llvm::getLoopEstimatedTripCount(Loop *L) {
   else
     return (FalseVal + (TrueVal / 2)) / TrueVal;
 }
 
 /// \brief Adds a 'fast' flag to floating point operations.
 static Value *addFastMathFlag(Value *V) {
   if (isa<FPMathOperator>(V)) {
     FastMathFlags Flags;
-    Flags.setUnsafeAlgebra();
+    Flags.setFast();
     cast<Instruction>(V)->setFastMathFlags(Flags);
   }
   return V;
 }
 
 // Helper to generate a log2 shuffle reduction.
 Value *
 llvm::getShuffleReduction(IRBuilder<> &Builder, Value *Src, unsigned Op,
Value *llvm::createSimpleTargetReduction(
   assert(isa<VectorType>(Src->getType()) && "Type must be a vector");
 
   Value *ScalarUdf = UndefValue::get(Src->getType()->getVectorElementType());
   std::function<Value*()> BuildFunc;
   using RD = RecurrenceDescriptor;
   RD::MinMaxRecurrenceKind MinMaxKind = RD::MRK_Invalid;
   // TODO: Support creating ordered reductions.
   FastMathFlags FMFUnsafe;
-  FMFUnsafe.setUnsafeAlgebra();
+  FMFUnsafe.setFast();
 
   switch (Opcode) {
   case Instruction::Add:
     BuildFunc = [&]() { return Builder.CreateAddReduce(Src); };
     break;
   case Instruction::Mul:
     BuildFunc = [&]() { return Builder.CreateMulReduce(Src); };
     break;

lib/Transforms/Utils/SimplifyLibCalls.cpp

Value *LibCallSimplifier::optimizePow(CallInst *CI, IRBuilder<> &B) {
 
   // pow(exp(x), y) -> exp(x * y)
   // pow(exp2(x), y) -> exp2(x * y)
   // We enable these only with fast-math. Besides rounding differences, the
   // transformation changes overflow and underflow behavior quite dramatically.
   // Example: x = 1000, y = 0.001.
   // pow(exp(x), y) = pow(inf, 0.001) = inf, whereas exp(x*y) = exp(1).
   auto *OpC = dyn_cast<CallInst>(Op1);
-  if (OpC && OpC->hasUnsafeAlgebra() && CI->hasUnsafeAlgebra()) {
+  if (OpC && OpC->isFast() && CI->isFast()) {
     LibFunc Func;
     Function *OpCCallee = OpC->getCalledFunction();
     if (OpCCallee && TLI->getLibFunc(OpCCallee->getName(), Func) &&
         TLI->has(Func) && (Func == LibFunc_exp || Func == LibFunc_exp2)) {
       IRBuilder<>::FastMathFlagGuard Guard(B);
       B.setFastMathFlags(CI->getFastMathFlags());
       Value *FMul = B.CreateFMul(OpC->getArgOperand(0), Op2, "mul");
       return emitUnaryFloatFnCall(FMul, OpCCallee->getName(), B,
                                   OpCCallee->getAttributes());
     }
   }
 
   ConstantFP *Op2C = dyn_cast<ConstantFP>(Op2);
   if (!Op2C)
     return Ret;
 
   if (Op2C->getValueAPF().isZero()) // pow(x, 0.0) -> 1.0
     return ConstantFP::get(CI->getType(), 1.0);
 
   if (Op2C->isExactlyValue(-0.5) &&
       hasUnaryFloatFn(TLI, Op2->getType(), LibFunc_sqrt, LibFunc_sqrtf,
                       LibFunc_sqrtl)) {
     // If -ffast-math:
     // pow(x, -0.5) -> 1.0 / sqrt(x)
-    if (CI->hasUnsafeAlgebra()) {
+    if (CI->isFast()) {
       IRBuilder<>::FastMathFlagGuard Guard(B);
       B.setFastMathFlags(CI->getFastMathFlags());
 
       // TODO: If the pow call is an intrinsic, we should lower to the sqrt
       // intrinsic, so we match errno semantics.  We also should check that the
       // target can in fact lower the sqrt intrinsic -- we currently have no way
       // to ask this question other than asking whether the target has a sqrt
       // libcall, which is a sufficient but not necessary condition.
       Value *Sqrt = emitUnaryFloatFnCall(Op1, TLI->getName(LibFunc_sqrt), B,
                                          Callee->getAttributes());
 
       return B.CreateFDiv(ConstantFP::get(CI->getType(), 1.0), Sqrt, "sqrtrecip");
     }
   }
 
   if (Op2C->isExactlyValue(0.5) &&
       hasUnaryFloatFn(TLI, Op2->getType(), LibFunc_sqrt, LibFunc_sqrtf,
                       LibFunc_sqrtl)) {
 
     // In -ffast-math, pow(x, 0.5) -> sqrt(x).
-    if (CI->hasUnsafeAlgebra()) {
+    if (CI->isFast()) {
       IRBuilder<>::FastMathFlagGuard Guard(B);
       B.setFastMathFlags(CI->getFastMathFlags());
 
       // TODO: As above, we should lower to the sqrt intrinsic if the pow is an
       // intrinsic, to match errno semantics.
       return emitUnaryFloatFnCall(Op1, TLI->getName(LibFunc_sqrt), B,
                                   Callee->getAttributes());
     }
Value *LibCallSimplifier::optimizePow(CallInst *CI, IRBuilder<> &B) {
   if (Op2C->isExactlyValue(1.0)) // pow(x, 1.0) -> x
     return Op1;
   if (Op2C->isExactlyValue(2.0)) // pow(x, 2.0) -> x*x
     return B.CreateFMul(Op1, Op1, "pow2");
   if (Op2C->isExactlyValue(-1.0)) // pow(x, -1.0) -> 1.0/x
     return B.CreateFDiv(ConstantFP::get(CI->getType(), 1.0), Op1, "powrecip");
 
   // In -ffast-math, generate repeated fmul instead of generating pow(x, n).
-  if (CI->hasUnsafeAlgebra()) {
+  if (CI->isFast()) {
     APFloat V = abs(Op2C->getValueAPF());
     // We limit to a max of 7 fmul(s). Thus max exponent is 32.
     // This transformation applies to integer exponents only.
     if (V.compare(APFloat(V.getSemantics(), 32.0)) == APFloat::cmpGreaterThan ||
         !V.isInteger())
       return nullptr;
 
     // Propagate fast math flags.
Value *LibCallSimplifier::optimizeFMinFMax(CallInst *CI, IRBuilder<> &B) {
   // function, do that first.
   StringRef Name = Callee->getName();
   if ((Name == "fmin" || Name == "fmax") && hasFloatVersion(Name))
     if (Value *Ret = optimizeBinaryDoubleFP(CI, B))
       return Ret;
 
   IRBuilder<>::FastMathFlagGuard Guard(B);
   FastMathFlags FMF;
-  if (CI->hasUnsafeAlgebra()) {
+  if (CI->isFast()) {
     // Unsafe algebra sets all fast-math-flags to true.
-    FMF.setUnsafeAlgebra();
+    FMF.setFast();
   } else {
     // At a minimum, no-nans-fp-math must be true.
     if (!CI->hasNoNaNs())
       return nullptr;
     // No-signed-zeros is implied by the definitions of fmax/fmin themselves:
     // "Ideally, fmax would be sensitive to the sign of zero, for example
     // fmax(-0. 0, +0. 0) would return +0; however, implementation in software
     // might be impractical."

 
 Value *LibCallSimplifier::optimizeLog(CallInst *CI, IRBuilder<> &B) {
   Function *Callee = CI->getCalledFunction();
   Value *Ret = nullptr;
   StringRef Name = Callee->getName();
   if (UnsafeFPShrink && hasFloatVersion(Name))
     Ret = optimizeUnaryDoubleFP(CI, B, true);
 
-  if (!CI->hasUnsafeAlgebra())
+  if (!CI->isFast())
     return Ret;
   Value *Op1 = CI->getArgOperand(0);
   auto *OpC = dyn_cast<CallInst>(Op1);
 
   // The earlier call must also be unsafe in order to do these transforms.
-  if (!OpC || !OpC->hasUnsafeAlgebra())
+  if (!OpC || !OpC->isFast())
     return Ret;
 
   // log(pow(x,y)) -> y*log(x)
   // This is only applicable to log, log2, log10.
   if (Name != "log" && Name != "log2" && Name != "log10")
     return Ret;
 
   IRBuilder<>::FastMathFlagGuard Guard(B);
   FastMathFlags FMF;
-  FMF.setUnsafeAlgebra();
+  FMF.setFast();
   B.setFastMathFlags(FMF);
 
   LibFunc Func;
   Function *F = OpC->getCalledFunction();
   if (F && ((TLI->getLibFunc(F->getName(), Func) && TLI->has(Func) &&
       Func == LibFunc_pow) || F->getIntrinsicID() == Intrinsic::pow))
     return B.CreateFMul(OpC->getArgOperand(1),
       emitUnaryFloatFnCall(OpC->getOperand(0), Callee->getName(), B,
Value *LibCallSimplifier::optimizeSqrt(CallInst *CI, IRBuilder<> &B) {
   Value *Ret = nullptr;
   // TODO: Once we have a way (other than checking for the existince of the
   // libcall) to tell whether our target can lower @llvm.sqrt, relax the
   // condition below.
   if (TLI->has(LibFunc_sqrtf) && (Callee->getName() == "sqrt" ||
                                   Callee->getIntrinsicID() == Intrinsic::sqrt))
     Ret = optimizeUnaryDoubleFP(CI, B, true);
 
-  if (!CI->hasUnsafeAlgebra())
+  if (!CI->isFast())
     return Ret;
 
   Instruction *I = dyn_cast<Instruction>(CI->getArgOperand(0));
-  if (!I || I->getOpcode() != Instruction::FMul || !I->hasUnsafeAlgebra())
+  if (!I || I->getOpcode() != Instruction::FMul || !I->isFast())
     return Ret;
 
   // We're looking for a repeated factor in a multiplication tree,
   // so we can do this fold: sqrt(x * x) -> fabs(x);
   // or this fold: sqrt((x * x) * y) -> fabs(x) * sqrt(y).
   Value *Op0 = I->getOperand(0);
   Value *Op1 = I->getOperand(1);
   Value *RepeatOp = nullptr;
   Value *OtherOp = nullptr;
   if (Op0 == Op1) {
     // Simple match: the operands of the multiply are identical.
     RepeatOp = Op0;
   } else {
     // Look for a more complicated pattern: one of the operands is itself
     // a multiply, so search for a common factor in that multiply.
     // Note: We don't bother looking any deeper than this first level or for
     // variations of this pattern because instcombine's visitFMUL and/or the
     // reassociation pass should give us this form.
     Value *OtherMul0, *OtherMul1;
     if (match(Op0, m_FMul(m_Value(OtherMul0), m_Value(OtherMul1)))) {
       // Pattern: sqrt((x * y) * z)
-      if (OtherMul0 == OtherMul1 &&
-          cast<Instruction>(Op0)->hasUnsafeAlgebra()) {
+      if (OtherMul0 == OtherMul1 && cast<Instruction>(Op0)->isFast()) {
         // Matched: sqrt((x * x) * z)
         RepeatOp = OtherMul0;
         OtherOp = Op1;
       }
     }
   }
   if (!RepeatOp)
     return Ret;
if (UnsafeFPShrink && Name == "tan" && hasFloatVersion(Name))
     Ret = optimizeUnaryDoubleFP(CI, B, true);
 
   Value *Op1 = CI->getArgOperand(0);
   auto *OpC = dyn_cast<CallInst>(Op1);
   if (!OpC)
     return Ret;
 
   // Both calls must allow unsafe optimizations in order to remove them.
-  if (!CI->hasUnsafeAlgebra() || !OpC->hasUnsafeAlgebra())
+  if (!CI->isFast() || !OpC->isFast())
     return Ret;
 
   // tan(atan(x)) -> x
   // tanf(atanf(x)) -> x
   // tanl(atanl(x)) -> x
   LibFunc Func;
   Function *F = OpC->getCalledFunction();
   if (F && TLI->getLibFunc(F->getName(), Func) && TLI->has(Func) &&
Value *LibCallSimplifier::optimizeCall(CallInst *CI) {
 
   SmallVector<OperandBundleDef, 2> OpBundles;
   CI->getOperandBundlesAsDefs(OpBundles);
   IRBuilder<> Builder(CI, /*FPMathTag=*/nullptr, OpBundles);
   bool isCallingConvC = isCallingConvCCompatible(CI);
 
   // Command-line parameter overrides instruction attribute.
   // This can't be moved to optimizeFloatingPointLibCall() because it may be
-  // used by the intrinsic optimizations. 
+  // used by the intrinsic optimizations.
   if (EnableUnsafeFPShrink.getNumOccurrences() > 0)
     UnsafeFPShrink = EnableUnsafeFPShrink;
-  else if (isa<FPMathOperator>(CI) && CI->hasUnsafeAlgebra())
+  else if (isa<FPMathOperator>(CI) && CI->isFast())
     UnsafeFPShrink = true;
 
   // First, check for intrinsics.
   if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
     if (!isCallingConvC)
       return nullptr;
     // The FP intrinsics have corresponding constrained versions so we don't
     // need to check for the StrictFP attribute here.

lib/Transforms/Vectorize/LoopVectorize.cpp

 /// TODO: We should use actual block probability here, if available. Currently,
 ///       we always assume predicated blocks have a 50% chance of executing.
 static unsigned getReciprocalPredBlockProb() { return 2; }
 
 /// A helper function that adds a 'fast' flag to floating-point operations.
 static Value *addFastMathFlag(Value *V) {
   if (isa<FPMathOperator>(V)) {
     FastMathFlags Flags;
-    Flags.setUnsafeAlgebra();
+    Flags.setFast();
     cast<Instruction>(V)->setFastMathFlags(Flags);
   }
   return V;
 }
 
 /// A helper function that returns an integer or floating-point constant with
 /// value C.
 static Constant *getSignedIntOrFpConstant(Type *Ty, int64_t C) {
Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx, Value *Step,
 
   // Add the consecutive indices to the vector value.
   Constant *Cv = ConstantVector::get(Indices);
 
   Step = Builder.CreateVectorSplat(VLen, Step);
 
   // Floating point operations had to be 'fast' to enable the induction.
   FastMathFlags Flags;
-  Flags.setUnsafeAlgebra();
+  Flags.setFast();
 
   Value *MulOp = Builder.CreateFMul(Cv, Step);
   if (isa<Instruction>(MulOp))
     // Have to check, MulOp may be a constant
     cast<Instruction>(MulOp)->setFastMathFlags(Flags);
 
   Value *BOp = Builder.CreateBinOp(BinOp, Val, MulOp, "induction");
   if (isa<Instruction>(BOp))
for (Instruction &I : *BB) {
         }
 
         // FP instructions can allow unsafe algebra, thus vectorizable by
         // non-IEEE-754 compliant SIMD units.
         // This applies to floating-point math operations and calls, not memory
         // operations, shuffles, or casts, as they don't change precision or
         // semantics.
       } else if (I.getType()->isFloatingPointTy() && (CI || I.isBinaryOp()) &&
-                 !I.hasUnsafeAlgebra()) {
+                 !I.isFast()) {
         DEBUG(dbgs() << "LV: Found FP op with unsafe algebra.\n");
         Hints->setPotentiallyUnsafe();
       }
 
       // Reduction instructions are allowed to have exit users.
       // All other instructions must not have external users.
       if (hasOutsideLoopUser(TheLoop, &I, AllowedExit)) {
         ORE->emit(createMissedAnalysis("ValueUsedOutsideLoop", &I)

lib/Transforms/Vectorize/SLPVectorizer.cpp

bool isAssociative(Instruction *I) const {
       assert(Kind != RK_None && *this && LHS && RHS &&
              "Expected reduction operation.");
       switch (Kind) {
       case RK_Arithmetic:
         return I->isAssociative();
       case RK_Min:
       case RK_Max:
         return Opcode == Instruction::ICmp ||
-               cast<Instruction>(I->getOperand(0))->hasUnsafeAlgebra();
+               cast<Instruction>(I->getOperand(0))->isFast();
       case RK_UMin:
       case RK_UMax:
         assert(Opcode == Instruction::ICmp &&
                "Only integer compare operation is expected.");
         return true;
       case RK_None:
         break;
       }
bool tryToReduce(BoUpSLP &V, TargetTransformInfo *TTI) {
     if (NumReducedVals < 4)
       return false;
 
     unsigned ReduxWidth = PowerOf2Floor(NumReducedVals);
 
     Value *VectorizedTree = nullptr;
     IRBuilder<> Builder(ReductionRoot);
     FastMathFlags Unsafe;
-    Unsafe.setUnsafeAlgebra();
+    Unsafe.setFast();
     Builder.setFastMathFlags(Unsafe);
     unsigned i = 0;
 
     BoUpSLP::ExtraValueToDebugLocsMap ExternallyUsedValues;
     // The same extra argument may be used several time, so log each attempt
     // to use it.
     for (auto &Pair : ExtraArgs)
       ExternallyUsedValues[Pair.second].push_back(Pair.first);

test/Assembler/fast-math-flags.ll

 ; RUN: llvm-as < %s | llvm-dis | FileCheck %s
 ; RUN: opt -S < %s | FileCheck %s
 ; RUN: verify-uselistorder %s
 
 @addr   = external global i64
 @select = external global i1
 @vec    = external global <3 x float>
 @arr    = external global [3 x float]
 
+declare float @foo(float)
+
 define float @none(float %x, float %y) {
 entry:
 ; CHECK:  %vec = load  <3 x float>,  <3 x float>* @vec
   %vec    = load  <3 x float>,  <3 x float>* @vec
 ; CHECK:  %select = load i1, i1* @select
   %select = load i1, i1* @select
 ; CHECK:  %arr    = load [3 x float], [3 x float]* @arr
   %arr    = load [3 x float], [3 x float]* @arr
; CHECK: %a = fsub contract float %x, %y
   %a = fsub contract float %x, %y
 ; CHECK: %b = fadd contract float %x, %y
   %b = fadd contract float %x, %y
 ; CHECK: %c = fmul contract float %a, %b
   %c = fmul contract float %a, %b
   ret float %c
 }
 
+; CHECK: @reassoc(
+define float @reassoc(float %x, float %y) {
+; CHECK: %a = fsub reassoc float %x, %y
+  %a = fsub reassoc float %x, %y
+; CHECK: %b = fmul reassoc float %x, %y
+  %b = fmul reassoc float %x, %y
+; CHECK: %c = call reassoc float @foo(float %b)
+  %c = call reassoc float @foo(float %b)
+  ret float %c
+}
+
+; CHECK: @afn(
+define float @afn(float %x, float %y) {
+; CHECK: %a = fdiv afn float %x, %y
+  %a = fdiv afn float %x, %y
+; CHECK: %b = frem afn float %x, %y
+  %b = frem afn float %x, %y
+; CHECK: %c = call afn float @foo(float %b)
+  %c = call afn float @foo(float %b)
+  ret float %c
+}
+
 ; CHECK: no_nan_inf
 define float @no_nan_inf(float %x, float %y) {
 entry:
 ; CHECK:  %vec = load <3 x float>, <3 x float>* @vec
   %vec    = load <3 x float>, <3 x float>* @vec
 ; CHECK:  %select = load i1, i1* @select
   %select = load i1, i1* @select
 ; CHECK:  %arr = load [3 x float], [3 x float]* @arr

 entry:
 ; CHECK:  %vec = load <3 x float>, <3 x float>* @vec
   %vec    = load <3 x float>, <3 x float>* @vec
 ; CHECK:  %select = load i1, i1* @select
   %select = load i1, i1* @select
 ; CHECK:  %arr    = load [3 x float], [3 x float]* @arr
   %arr    = load [3 x float], [3 x float]* @arr
 
-; CHECK:  %a = fadd nnan ninf float %x, %y
-  %a = fadd ninf nnan float %x, %y
-; CHECK:  %a_vec = fadd nnan <3 x float> %vec, %vec
-  %a_vec = fadd nnan <3 x float> %vec, %vec
+; CHECK:  %a = fadd nnan ninf afn float %x, %y
+  %a = fadd ninf nnan afn float %x, %y
+; CHECK:  %a_vec = fadd reassoc nnan <3 x float> %vec, %vec
+  %a_vec = fadd reassoc nnan <3 x float> %vec, %vec
 ; CHECK:  %b = fsub fast float %x, %y
   %b = fsub nnan nsz fast float %x, %y
 ; CHECK:  %b_vec = fsub nnan <3 x float> %vec, %vec
   %b_vec = fsub nnan <3 x float> %vec, %vec
 ; CHECK:  %c = fmul fast float %x, %y
   %c = fmul nsz fast arcp float %x, %y
 ; CHECK:  %c_vec = fmul nsz <3 x float> %vec, %vec
   %c_vec = fmul nsz <3 x float> %vec, %vec

test/Bitcode/compatibility-3.6.ll

define void @fastmathflags(float %op1, float %op2) {
   ; CHECK: %f.nnan = fadd nnan float %op1, %op2
   %f.ninf = fadd ninf float %op1, %op2
   ; CHECK: %f.ninf = fadd ninf float %op1, %op2
   %f.nsz = fadd nsz float %op1, %op2
   ; CHECK: %f.nsz = fadd nsz float %op1, %op2
   %f.arcp = fadd arcp float %op1, %op2
   ; CHECK: %f.arcp = fadd arcp float %op1, %op2
   %f.fast = fadd fast float %op1, %op2
-  ; CHECK: %f.fast = fadd fast float %op1, %op2
+  ; 'fast' used to be its own bit, but this changed in Oct 2017.
+  ; The binary test file does not have the newer 'contract' and 'afn' bits set, so this is not fully 'fast'.
+  ; CHECK: %f.fast = fadd reassoc nnan ninf nsz arcp float %op1, %op2
   ret void
 }
 
 ;; Type System
 %opaquety = type opaque
 define void @typesystem() {
   %p0 = bitcast i8* null to i32 (i32)*
   ; CHECK: %p0 = bitcast i8* null to i32 (i32)*

test/Bitcode/compatibility-3.7.ll

define void @fastmathflags(float %op1, float %op2) {
   ; CHECK: %f.nnan = fadd nnan float %op1, %op2
   %f.ninf = fadd ninf float %op1, %op2
   ; CHECK: %f.ninf = fadd ninf float %op1, %op2
   %f.nsz = fadd nsz float %op1, %op2
   ; CHECK: %f.nsz = fadd nsz float %op1, %op2
   %f.arcp = fadd arcp float %op1, %op2
   ; CHECK: %f.arcp = fadd arcp float %op1, %op2
   %f.fast = fadd fast float %op1, %op2
-  ; CHECK: %f.fast = fadd fast float %op1, %op2
+  ; 'fast' used to be its own bit, but this changed in Oct 2017.
+  ; The binary test file does not have the newer 'contract' and 'afn' bits set, so this is not fully 'fast'.
+  ; CHECK: %f.fast = fadd reassoc nnan ninf nsz arcp float %op1, %op2
   ret void
 }
 
 ;; Type System
 %opaquety = type opaque
 define void @typesystem() {
   %p0 = bitcast i8* null to i32 (i32)*
   ; CHECK: %p0 = bitcast i8* null to i32 (i32)*

test/Bitcode/compatibility-3.8.ll

define void @fastmathflags(float %op1, float %op2) {
   ; CHECK: %f.nnan = fadd nnan float %op1, %op2
   %f.ninf = fadd ninf float %op1, %op2
   ; CHECK: %f.ninf = fadd ninf float %op1, %op2
   %f.nsz = fadd nsz float %op1, %op2
   ; CHECK: %f.nsz = fadd nsz float %op1, %op2
   %f.arcp = fadd arcp float %op1, %op2
   ; CHECK: %f.arcp = fadd arcp float %op1, %op2
   %f.fast = fadd fast float %op1, %op2
-  ; CHECK: %f.fast = fadd fast float %op1, %op2
+  ; 'fast' used to be its own bit, but this changed in Oct 2017.
+  ; The binary test file does not have the newer 'contract' and 'afn' bits set, so this is not fully 'fast'.
+  ; CHECK: %f.fast = fadd reassoc nnan ninf nsz arcp float %op1, %op2
   ret void
 }
 
 ; Check various fast math flags and floating-point types on calls.
 
 declare float @fmf1()
 declare double @fmf2()
 declare <4 x double> @fmf3()
 
 ; CHECK-LABEL: fastMathFlagsForCalls(
 define void @fastMathFlagsForCalls(float %f, double %d1, <4 x double> %d2) {
   %call.fast = call fast float @fmf1()
-  ; CHECK: %call.fast = call fast float @fmf1()
+  ; 'fast' used to be its own bit, but this changed in Oct 2017.
+  ; The binary test file does not have the newer 'contract' and 'aml' bits set, so this is not fully 'fast'.
+  ; CHECK: %call.fast = call reassoc nnan ninf nsz arcp float @fmf1()
 
   ; Throw in some other attributes to make sure those stay in the right places.
 
   %call.nsz.arcp = notail call nsz arcp double @fmf2()
   ; CHECK: %call.nsz.arcp = notail call nsz arcp double @fmf2()
 
   %call.nnan.ninf = tail call nnan ninf fastcc <4 x double> @fmf3()
   ; CHECK: %call.nnan.ninf = tail call nnan ninf fastcc <4 x double> @fmf3()

test/Bitcode/compatibility-3.9.ll

define void @fastmathflags(float %op1, float %op2) {
   ; CHECK: %f.nnan = fadd nnan float %op1, %op2
   %f.ninf = fadd ninf float %op1, %op2
   ; CHECK: %f.ninf = fadd ninf float %op1, %op2
   %f.nsz = fadd nsz float %op1, %op2
   ; CHECK: %f.nsz = fadd nsz float %op1, %op2
   %f.arcp = fadd arcp float %op1, %op2
   ; CHECK: %f.arcp = fadd arcp float %op1, %op2
   %f.fast = fadd fast float %op1, %op2
-  ; CHECK: %f.fast = fadd fast float %op1, %op2
+  ; 'fast' used to be its own bit, but this changed in Oct 2017.
+  ; The binary test file does not have the newer 'contract' and 'afn' bits set, so this is not fully 'fast'.
+  ; CHECK: %f.fast = fadd reassoc nnan ninf nsz arcp float %op1, %op2
   ret void
 }
 
 ; Check various fast math flags and floating-point types on calls.
 
 declare float @fmf1()
 declare double @fmf2()
 declare <4 x double> @fmf3()
 
 ; CHECK-LABEL: fastMathFlagsForCalls(
 define void @fastMathFlagsForCalls(float %f, double %d1, <4 x double> %d2) {
   %call.fast = call fast float @fmf1()
-  ; CHECK: %call.fast = call fast float @fmf1()
+  ; 'fast' used to be its own bit, but this changed in Oct 2017.
+  ; The binary test file does not have the newer 'contract' and 'afn' bits set, so this is not fully 'fast'.
+  ; CHECK: %call.fast = call reassoc nnan ninf nsz arcp float @fmf1()
 
   ; Throw in some other attributes to make sure those stay in the right places.
 
   %call.nsz.arcp = notail call nsz arcp double @fmf2()
   ; CHECK: %call.nsz.arcp = notail call nsz arcp double @fmf2()
 
   %call.nnan.ninf = tail call nnan ninf fastcc <4 x double> @fmf3()
   ; CHECK: %call.nnan.ninf = tail call nnan ninf fastcc <4 x double> @fmf3()

test/Bitcode/compatibility-4.0.ll

define void @fastmathflags(float %op1, float %op2) {
   %f.nnan = fadd nnan float %op1, %op2
   ; CHECK: %f.nnan = fadd nnan float %op1, %op2
   %f.ninf = fadd ninf float %op1, %op2
   ; CHECK: %f.ninf = fadd ninf float %op1, %op2
   %f.nsz = fadd nsz float %op1, %op2
   ; CHECK: %f.nsz = fadd nsz float %op1, %op2
   %f.arcp = fadd arcp float %op1, %op2
   ; CHECK: %f.arcp = fadd arcp float %op1, %op2
+  ; 'fast' used to be its own bit, but this changed in Oct 2017.
+  ; The binary test file does not have the newer 'contract' and 'afn' bits set, so this is not fully 'fast'.
   %f.fast = fadd fast float %op1, %op2
-  ; CHECK: %f.fast = fadd fast float %op1, %op2
+  ; CHECK: %f.fast = fadd reassoc nnan ninf nsz arcp float %op1, %op2
   ret void
 }
 
 ; Check various fast math flags and floating-point types on calls.
 
 declare float @fmf1()
 declare double @fmf2()
 declare <4 x double> @fmf3()
 
 ; CHECK-LABEL: fastMathFlagsForCalls(
 define void @fastMathFlagsForCalls(float %f, double %d1, <4 x double> %d2) {
   %call.fast = call fast float @fmf1()
-  ; CHECK: %call.fast = call fast float @fmf1()
+  ; 'fast' used to be its own bit, but this changed in Oct 2017.
+  ; The binary test file does not have the newer 'contract' and 'afn' bits set, so this is not fully 'fast'.
+  ; CHECK: %call.fast = call reassoc nnan ninf nsz arcp float @fmf1()
 
   ; Throw in some other attributes to make sure those stay in the right places.
 
   %call.nsz.arcp = notail call nsz arcp double @fmf2()
   ; CHECK: %call.nsz.arcp = notail call nsz arcp double @fmf2()
 
   %call.nnan.ninf = tail call nnan ninf fastcc <4 x double> @fmf3()
   ; CHECK: %call.nnan.ninf = tail call nnan ninf fastcc <4 x double> @fmf3()

test/Bitcode/compatibility-5.0.ll

define void @fastmathflags(float %op1, float %op2) {
   ; CHECK: %f.ninf = fadd ninf float %op1, %op2
   %f.nsz = fadd nsz float %op1, %op2
   ; CHECK: %f.nsz = fadd nsz float %op1, %op2
   %f.arcp = fadd arcp float %op1, %op2
   ; CHECK: %f.arcp = fadd arcp float %op1, %op2
   %f.contract = fadd contract float %op1, %op2
   ; CHECK: %f.contract = fadd contract float %op1, %op2
   %f.fast = fadd fast float %op1, %op2
-  ; CHECK: %f.fast = fadd fast float %op1, %op2
+  ; 'fast' used to be its own bit, but this changed in Oct 2017.
+  ; The binary test file does not have the newer 'afn' bit set, so this is not fully 'fast'.
+  ; CHECK: %f.fast = fadd reassoc nnan ninf nsz arcp contract float %op1, %op2
   ret void
 }
 
 ; Check various fast math flags and floating-point types on calls.
 
 declare float @fmf1()
 declare double @fmf2()
 declare <4 x double> @fmf3()
 
 ; CHECK-LABEL: fastMathFlagsForCalls(
 define void @fastMathFlagsForCalls(float %f, double %d1, <4 x double> %d2) {
   %call.fast = call fast float @fmf1()
-  ; CHECK: %call.fast = call fast float @fmf1()
+  ; 'fast' used to be its own bit, but this changed in Oct 2017.
+  ; The binary test file does not have the newer 'afn' bit set, so this is not fully 'fast'.
+  ; CHECK: %call.fast = call reassoc nnan ninf nsz arcp contract float @fmf1()
 
   ; Throw in some other attributes to make sure those stay in the right places.
 
   %call.nsz.arcp = notail call nsz arcp double @fmf2()
   ; CHECK: %call.nsz.arcp = notail call nsz arcp double @fmf2()
 
   %call.nnan.ninf = tail call nnan ninf fastcc <4 x double> @fmf3()
   ; CHECK: %call.nnan.ninf = tail call nnan ninf fastcc <4 x double> @fmf3()

test/Bitcode/compatibility.ll

define void @fastmathflags(float %op1, float %op2) {
   %f.ninf = fadd ninf float %op1, %op2
   ; CHECK: %f.ninf = fadd ninf float %op1, %op2
   %f.nsz = fadd nsz float %op1, %op2
   ; CHECK: %f.nsz = fadd nsz float %op1, %op2
   %f.arcp = fadd arcp float %op1, %op2
   ; CHECK: %f.arcp = fadd arcp float %op1, %op2
   %f.contract = fadd contract float %op1, %op2
   ; CHECK: %f.contract = fadd contract float %op1, %op2
+  %f.afn = fadd afn float %op1, %op2
+  ; CHECK: %f.afn = fadd afn float %op1, %op2
+  %f.reassoc = fadd reassoc float %op1, %op2
+  ; CHECK: %f.reassoc = fadd reassoc float %op1, %op2
   %f.fast = fadd fast float %op1, %op2
   ; CHECK: %f.fast = fadd fast float %op1, %op2
   ret void
 }
 
 ; Check various fast math flags and floating-point types on calls.
 
 declare float @fmf1()

unittests/IR/IRBuilderTest.cpp

TEST_F(IRBuilderTest, FastMathFlags) {
   EXPECT_FALSE(Builder.getFastMathFlags().any());
   ASSERT_TRUE(isa<Instruction>(F));
   FAdd = cast<Instruction>(F);
   EXPECT_FALSE(FAdd->hasNoNaNs());
 
   FastMathFlags FMF;
   Builder.setFastMathFlags(FMF);
 
+  // By default, no flags are set.
   F = Builder.CreateFAdd(F, F);
   EXPECT_FALSE(Builder.getFastMathFlags().any());
+  ASSERT_TRUE(isa<Instruction>(F));
+  FAdd = cast<Instruction>(F);
+  EXPECT_FALSE(FAdd->hasNoNaNs());
+  EXPECT_FALSE(FAdd->hasNoInfs());
+  EXPECT_FALSE(FAdd->hasNoSignedZeros());
+  EXPECT_FALSE(FAdd->hasAllowReciprocal());
+  EXPECT_FALSE(FAdd->hasAllowContract());
+  EXPECT_FALSE(FAdd->hasAllowReassoc());
+  EXPECT_FALSE(FAdd->hasApproxFunc());
 
-  FMF.setUnsafeAlgebra();
+  // Set all flags in the instruction.
+  FAdd->setFast(true);
+  EXPECT_TRUE(FAdd->hasNoNaNs());
+  EXPECT_TRUE(FAdd->hasNoInfs());
+  EXPECT_TRUE(FAdd->hasNoSignedZeros());
+  EXPECT_TRUE(FAdd->hasAllowReciprocal());
+  EXPECT_TRUE(FAdd->hasAllowContract());
+  EXPECT_TRUE(FAdd->hasAllowReassoc());
+  EXPECT_TRUE(FAdd->hasApproxFunc());
+
+  // All flags are set in the builder.
+  FMF.setFast();
   Builder.setFastMathFlags(FMF);
 
   F = Builder.CreateFAdd(F, F);
   EXPECT_TRUE(Builder.getFastMathFlags().any());
+  EXPECT_TRUE(Builder.getFastMathFlags().all());
   ASSERT_TRUE(isa<Instruction>(F));
   FAdd = cast<Instruction>(F);
   EXPECT_TRUE(FAdd->hasNoNaNs());
+  EXPECT_TRUE(FAdd->isFast());
 
   // Now, try it with CreateBinOp
   F = Builder.CreateBinOp(Instruction::FAdd, F, F);
   EXPECT_TRUE(Builder.getFastMathFlags().any());
   ASSERT_TRUE(isa<Instruction>(F));
   FAdd = cast<Instruction>(F);
   EXPECT_TRUE(FAdd->hasNoNaNs());
+  EXPECT_TRUE(FAdd->isFast());
 
   F = Builder.CreateFDiv(F, F);
-  EXPECT_TRUE(Builder.getFastMathFlags().any());
-  EXPECT_TRUE(Builder.getFastMathFlags().UnsafeAlgebra);
+  EXPECT_TRUE(Builder.getFastMathFlags().all());
   ASSERT_TRUE(isa<Instruction>(F));
   FDiv = cast<Instruction>(F);
   EXPECT_TRUE(FDiv->hasAllowReciprocal());
 
+  // Clear all FMF in the builder.
   Builder.clearFastMathFlags();
 
   F = Builder.CreateFDiv(F, F);
   ASSERT_TRUE(isa<Instruction>(F));
   FDiv = cast<Instruction>(F);
   EXPECT_FALSE(FDiv->hasAllowReciprocal());
  
+  // Try individual flags.
   FMF.clear();
   FMF.setAllowReciprocal();
   Builder.setFastMathFlags(FMF);
 
   F = Builder.CreateFDiv(F, F);
   EXPECT_TRUE(Builder.getFastMathFlags().any());
   EXPECT_TRUE(Builder.getFastMathFlags().AllowReciprocal);
   ASSERT_TRUE(isa<Instruction>(F));
TEST_F(IRBuilderTest, FastMathFlags) {
 
   FC = Builder.CreateFAdd(F, F);
   EXPECT_TRUE(Builder.getFastMathFlags().any());
   EXPECT_TRUE(Builder.getFastMathFlags().AllowContract);
   ASSERT_TRUE(isa<Instruction>(FC));
   FAdd = cast<Instruction>(FC);
   EXPECT_TRUE(FAdd->hasAllowContract());
 
+  FMF.setApproxFunc();
   Builder.clearFastMathFlags();
+  Builder.setFastMathFlags(FMF);
+  // Now 'aml' and 'contract' are set.
+  F = Builder.CreateFMul(F, F);
+  FAdd = cast<Instruction>(F);
+  EXPECT_TRUE(FAdd->hasApproxFunc());
+  EXPECT_TRUE(FAdd->hasAllowContract());
+  EXPECT_FALSE(FAdd->hasAllowReassoc());
+  
+  FMF.setAllowReassoc();
+  Builder.clearFastMathFlags();
+  Builder.setFastMathFlags(FMF);
+  // Now 'aml' and 'contract' and 'reassoc' are set.
+  F = Builder.CreateFMul(F, F);
+  FAdd = cast<Instruction>(F);
+  EXPECT_TRUE(FAdd->hasApproxFunc());
+  EXPECT_TRUE(FAdd->hasAllowContract());
+  EXPECT_TRUE(FAdd->hasAllowReassoc());
 
   // Test a call with FMF.
   auto CalleeTy = FunctionType::get(Type::getFloatTy(Ctx),
                                     /*isVarArg=*/false);
   auto Callee =
       Function::Create(CalleeTy, Function::ExternalLinkage, "", M.get());
 
   FCall = Builder.CreateCall(Callee, None);