[LangRef] Specify NaN behavior more precisely #66579

RalfJung · 2023-09-16T15:43:22Z

The current docs don't say anything about the possible set of values returned by a NaN-producing operation. However, there are some coding patterns where such guarantees are strongly desired, such as NaN boxing (as used e.g. in the Spidermonkey JS engine). And in fact the set of possible payloads that can be observed today when compiling code via LLVM is fairly limited (based on what optimizations and backends and hardware currently do) -- but without a documented guarantee, languages compiling to LLVM cannot really rely on this.

There was a long discussion about the exact shape that the guarantees should take on Discourse. This PR transcribes what I perceived as the consensus, plus some further clarifications: "unless specified otherwise", returned NaNs have a non-deterministic sign, and the payload is generally either the preferred (all-zero) payload or propagates the payload of one of the inputs.

llvmbot · 2023-09-16T15:44:29Z

@llvm/pr-subscribers-llvm-ir

Changes

The current docs don't say anything about the possible set of values returned by a NaN-producing operation. However, there are some coding patterns where such guarantees are strongly desired, such as NaN boxing (as used e.g. in the Spidermonkey JS engine). And in fact the set of possible payloads that can be observed today when compiling code via LLVM is fairly limited (based on what optimizations and backends and hardware currently do) -- but without a documented guarantee, languages compiling to LLVM cannot really rely on this.

There was a long discussion about the exact shape that the guarantees should take on Discourse. This PR transcribes what I perceived as the consensus, plus some further clarifications.

The point I am least sure about is the status of the min/max operations on x86: the operations provided by hardware do not match the IEEE spec, so the backend has to emit a sequence of instructions that implement the IEEE operations; the bitwise operations used by the backend to achieve this could run afoul of the NaN guarantees. Here @jyknight stated that what the backend does might be compatible with the NaN guarantee -- but they didn't say that it affirmatively is compatible, so it would be great if someone with knowledge of the x86 backend could confirm or deny this.

Full diff: https://github.com/llvm/llvm-project/pull/66579.diff

1 Files Affected:

(modified) llvm/docs/LangRef.rst (+30-6)

diff --git a/llvm/docs/LangRef.rst b/llvm/docs/LangRef.rst
index f542e70bcfee810..021e7ba2fb41722 100644
--- a/llvm/docs/LangRef.rst
+++ b/llvm/docs/LangRef.rst
@@ -3394,17 +3394,41 @@ Floating-Point Environment
 The default LLVM floating-point environment assumes that traps are disabled and
 status flags are not observable. Therefore, floating-point math operations do
 not have side effects and may be speculated freely. Results assume the
-round-to-nearest rounding mode.
+round-to-nearest rounding mode, and subnormals are assumed to be preserved.
+Running default LLVM code in an environment where these assumptions are not met
+can lead to undefined behavior.
+
+The representation bits of a floating-point value do not mutate arbitrarily; if
+there is no floating-point operation being performed, the NaN payload (if any)
+is preserved.
+
+When a floating-point math operation produces a NaN value, the result has a
+non-deterministic sign. The payload is non-deterministically chosen from the
+following set:
+
+- The payload that is all-zero except that the ``quiet`` bit is set.
+  ("Preferred NaN" case)
+- The payload of any input operand that is a NaN, bit-wise ORed with a payload that has
+  the ``quiet`` bit set. ("Quieting NaN propagation" case)
+- The payload of any input operand that is a NaN. ("Unchanged NaN propagation" case)
+- A target-specific set of further NaN payloads, that definitely all have their
+  ``quiet`` bit set. The set can depend on the payloads of the input NaNs.
+  This set is empty on x86 and ARM, but can be non-empty on other architectures.
+  (For instance, on wasm, if any input NaN is not the preferred NaN, then
+  this set contains all quiet NaNs; otherwise, it is empty.
+  On SPARC, this set consists of the all-one payload.)
+
+In particular, if all input NaNs are quiet, then the output NaN is definitely
+quiet. Signaling NaN outputs can only occur if they are provided as an input
+value. For example, "fmul SNaN, 1.0" may be simplified to SNaN rather than QNaN.
 
 Floating-point math operations are allowed to treat all NaNs as if they were
-quiet NaNs. For example, "pow(1.0, SNaN)" may be simplified to 1.0. This also
-means that SNaN may be passed through a math operation without quieting. For
-example, "fmul SNaN, 1.0" may be simplified to SNaN rather than QNaN. However,
-SNaN values are never created by math operations. They may only occur when
-provided as a program input value.
+quiet NaNs. For example, "pow(1.0, SNaN)" may be simplified to 1.0.
 
 Code that requires different behavior than this should use the
 :ref:`Constrained Floating-Point Intrinsics <constrainedfp>`.
+In particular, constrained intrinsics rule out the "Unchanged NaN propagation" case;
+they are guaranteed to return a QNaN.
 
 .. _fastmath:

programmerjake · 2023-09-17T05:10:44Z

llvm/docs/LangRef.rst

+
+The representation bits of a floating-point value do not mutate arbitrarily; if
+there is no floating-point operation being performed, the NaN payload (if any)
+is preserved.


floating point bits do mutate arbitrarily on 32-bit x86:
here, all I do is bitcast 0x7f800001 to float, return it from a function, then bitcast back to an integer -- that changes it to 0x7fc00001.
https://clang.godbolt.org/z/sjaq55q3G

Yes, that's a bug in the x86-32 backend / calling convention.

Given that this bug is hard-coded in the calling convention, I wonder if it should be called out here. But the main point of this sentence is that LLVM itself will not do anything that would mutate FP values.

In fact I assume one can get LLVM to miscompile code by assuming that return values don't change their bits, similar to #44218.

I do think we should at least mention the major ways in which LLVM backends violate the spec:

x86-32 with SSE2 enabled may implicitly convert floating-point values returned from a function to x86_fp80 and back; such conversions may signal a floating-point exception, or modify the value.

x86-32 without SSE2 may do such conversions just about anywhere.

older MIPS versions use the opposite polarity for the sNaN vs qNaN bit, and LLVM does not correctly represent this (this is at least potentially fixable, unlike the above).

I added a paragraph with backend problems and linked to relevant issues. Is there an issue for the "x86-32 return value problem"?

I'm not aware of one in LLVM, only in rust, so I filed a new one, #66803.

llvm/docs/LangRef.rst

jcranmer-intel

It is worth calling out somewhere that qNaN/sNaN are defined using the convention that leading bit = 0 is sNaN, 1 is qNaN, and systems that do not adhere to this convention (e.g., old MIPS) are unsupported.

Also probably worth calling out that things like bitcast do not count as floating-point operations.

llvm/docs/LangRef.rst

jcranmer-intel · 2023-09-18T15:14:02Z

llvm/docs/LangRef.rst

@@ -3394,17 +3394,41 @@ Floating-Point Environment
 The default LLVM floating-point environment assumes that traps are disabled and
 status flags are not observable. Therefore, floating-point math operations do
 not have side effects and may be speculated freely. Results assume the
-round-to-nearest rounding mode.
+round-to-nearest rounding mode, and subnormals are assumed to be preserved.
+Running default LLVM code in an environment where these assumptions are not met


"default LLVM code" is not a well-defined phrase :-)

You should explicitly refer to the role of the strictfp attribute here.

I'm mentioning strictfp now, I hope what I am saying makes sense. :)

llvm/docs/LangRef.rst

jcranmer-intel · 2023-09-18T15:52:14Z

llvm/docs/LangRef.rst

+there is no floating-point operation being performed, the NaN payload (if any)
+is preserved.
+
+When a floating-point math operation produces a NaN value, the result has a


I would prepend "Unless otherwise specified here", and add notes to fneg, llvm.fabs, and llvm.copysign that they are guaranteed to not affect the NaN payload or qNaN/sNaN status.

Those 3 shouldn't be considered floating point math operations (as the term is used here and before). We definitely need to make it clear in their definitions -- it has implications beyond just the NaN behavior.

In particular, they all ought to be specified as: "This operation never raises a floating-point exception, and the result is an exact copy of the input, other than the sign bit. It is not considered a floating-point math operation, but rather, bit-manipulation which operates on floating-point values."

The paragraph above says that FP exceptions cannot be observed anyway, so it seems confusing/misleading to now talk about some operations not raising FP exceptions.

I am explicitly calling out bitcasts and fnet, fabs, copysign as not being affected by the NaN non-determinism now.

I'd like to see some text added to the specification sections for the fneg/fabs/copysign operations. Noting there that exceptions cannot be signaled (in addition to them only touching the sign bit) is meaningful -- it means that you may use them in a "strictfp" function. Unlike other operations, they do not need a "constrained" variant, because their behavior is fully defined regardless.

I would still add "unless otherwise specified", because you have cases like llvm.canonicalize which is an FP operation that can arbitrarily mutate NaN payloads but requires the output to be qNaN, even if the input is sNaN.

Fair. That can then also cover fmin/fmax, should we need to add an exception for them.

I'd like to see some text added to the specification sections for the fneg/fabs/copysign operations.

Done.

…ed; add list of known-broken cases

RalfJung · 2023-09-19T06:59:51Z

llvm/docs/LangRef.rst

+.. _floatnan:
+
+Behavior of Floating-Point NaN values
+-------------------------------------


This wasn't really about the "FP environment" any more so I moved this all into a new section on NaN values.

RalfJung · 2023-09-19T07:04:24Z

Regarding fmin/fmax on x86, @Muon pointed out that LLVM currently simply calls the underlying libm functions to implement these operations. The C standard has nothing to say about the case where both inputs are a NaN. My local manpage says

If both arguments are NaN, a NaN is returned.

This is not very specific and could permit arbitrary NaN payloads. I don't know if that actually happens in practice though.

So I'm still unsure about whether we need to say that fmin/fmax on x86 can produce arbitrary NaN payloads. LLVM assumes things about memcpy that are not guaranteed by the C standard, so it could conceivably do the same for fmin/fmax -- but of course if real implementations violate the assumptions then that's not an option.

I feel like for now it is better to not make a guarantee that we aren't sure can be held up. Though if LLVM just uses libm, couldn't this cause other NaNs even on non-x86?

llvm/docs/LangRef.rst

jcranmer-intel · 2023-09-19T14:36:34Z

llvm/docs/LangRef.rst

+there is no floating-point operation being performed, the NaN payload (if any)
+is preserved.
+
+When a floating-point math operation produces a NaN value, the result has a


I would still add "unless otherwise specified", because you have cases like llvm.canonicalize which is an FP operation that can arbitrarily mutate NaN payloads but requires the output to be qNaN, even if the input is sNaN.

nunoplopes · 2023-09-19T16:00:39Z

FWIW, I'm trying to implement this new semantics in Alive2. I'll report back the changes in failed/passed tests later.

llvm/docs/LangRef.rst

…what this achieves

llvm/docs/LangRef.rst

nunoplopes · 2023-09-20T11:48:09Z

I've a question regarding fpext & fptrunc: what's the payload we want there? We can't just copy the inputs.

nunoplopes · 2023-09-20T14:26:04Z

I've a question regarding fpext & fptrunc: what's the payload we want there? We can't just copy the inputs.

Ok, I did the easy thing, which was to ignore the rules about copying the NaN payload from the input (in Alive2).
The good news is that there are only 4 tests failing. But they are related with fptrunc & fpext.

They are:

; Transforms/InstCombine/maximum.ll
define float @reduce_precision(float %x, float %y) {
  %x.ext = fpext float %x to double
  %y.ext = fpext float %y to double
  %maximum = fmaximum double %x.ext, %y.ext
  %trunc = fptrunc double %maximum to float
  ret float %trunc
}
=>
define float @reduce_precision(float %x, float %y) {
  %maximum1 = fmaximum float %x, %y
  ret float %maximum1
}
Transformation doesn't verify!

ERROR: Value mismatch

Example:
float %x = #x00000000 (+0.0)
float %y = #x7f808000 (SNaN)

Source:
double %x.ext = #x0000000000000000 (+0.0)
double %y.ext = #x7ff8000000000000 (QNaN)
double %maximum = #x7ff8000000000000 (QNaN)
float %trunc = #x7fc00000 (QNaN)

Target:
float %maximum1 = #x7f808000 (SNaN)
Source value: #x7fc00000 (QNaN)
Target value: #x7f808000 (SNaN)

; Transforms/InstCombine/select-binop-foldable-floating-point.ll
define float @select_fsub(i1 %cond, float %A, float %B) {
  %C = fsub float %A, %B
  %D = select i1 %cond, float %C, float %A
  ret float %D
}
=>
define float @select_fsub(i1 %cond, float %A, float %B) {
  %C = select i1 %cond, float %B, float 0.000000
  %D = fsub float %A, %C
  ret float %D
}
Transformation doesn't verify! (unsound)
ERROR: Target's return value is more undefined

Example:
i1 %cond = #x0 (0)
float %A = #x7f804000 (SNaN)
float %B = poison

Source:
float %C = poison
float %D = #x7f804000 (SNaN)

Target:
float %C = #x00000000 (+0.0)
float %D = #x7fd00000 (QNaN)
Source value: #x7f804000 (SNaN)
Target value: #x7fd00000 (QNaN)

; Transforms/InstCombine/fpcast.ll
define half @test4-fast(float %a) {
  %b = fsub fast float -0.000000, %a
  %c = fptrunc float %b to half
  ret half %c
}
=>
define half @test4-fast(float %a) {
  %1 = fptrunc float %a to half
  %c = fneg fast half %1
  ret half %c
}
Transformation doesn't verify!

ERROR: Target is more poisonous than source

Example:
float %a = #xd8804000 (-1128098930098176)

Source:
float %b = #x7fc00000 (QNaN)
half %c = #x7e00 (QNaN)

Target:
half %1 = #xfc00 (-oo)
half %c = poison
Source value: #x7e00 (QNaN)
Target value: poison

; Transforms/InstCombine/fpextend.ll
define float @test4(float %x) {
  %t1 = fpext float %x to double
  %t2 = fsub double -0.000000, %t1
  %t34 = fptrunc double %t2 to float
  ret float %t34
}
=>
define float @test4(float %x) {
  %t34 = fneg float %x
  ret float %t34
}
Transformation doesn't verify! (unsound)
ERROR: Value mismatch

Example:
float %x = #xff829ffe (SNaN)

Source:
double %t1 = #x7ff8000000000000 (QNaN)
double %t2 = #x7ff8000000000000 (QNaN)
float %t34 = #x7fc00000 (QNaN)

Target:
float %t34 = #x7f829ffe (SNaN)
Source value: #x7fc00000 (QNaN)
Target value: #x7f829ffe (SNaN)

nunoplopes · 2023-09-20T14:26:58Z

We need a fix for these instructions, so that SNaN can passthrough.

jcranmer-intel · 2023-09-20T16:24:21Z

I've a question regarding fpext & fptrunc: what's the payload we want there? We can't just copy the inputs.

fpext and fptrunc correspond to the IEEE 754 operation convertFormat, so the rules related to SNaN handling need to apply (i.e., it may or may not be quieting input SNaN). IEEE 754 furthermore recommends NaN propagation as follows:

Conversion of a quiet NaN from a narrower format to a wider format in the same radix, and then back to
the same narrower format, should not change the quiet NaN payload in any way except to make it
canonical.

The natural way to implement this is to designate certain bits to chop, but I don't know what the full suite of implementations here is. In x86, the afflicted bits are the LSB, so fpext on 0xffffffff results in 0xffffffffe0000000, and I think APFloat is following the same rules here, but I'm not certain all NaN-propagating implementations would follow the same rules.

nunoplopes · 2023-09-20T18:52:05Z

I've a question regarding fpext & fptrunc: what's the payload we want there? We can't just copy the inputs.

fpext and fptrunc correspond to the IEEE 754 operation convertFormat, so the rules related to SNaN handling need to apply (i.e., it may or may not be quieting input SNaN). IEEE 754 furthermore recommends NaN propagation as follows:

Conversion of a quiet NaN from a narrower format to a wider format in the same radix, and then back to
the same narrower format, should not change the quiet NaN payload in any way except to make it
canonical.

The natural way to implement this is to designate certain bits to chop, but I don't know what the full suite of implementations here is. In x86, the afflicted bits are the LSB, so fpext on 0xffffffff results in 0xffffffffe0000000, and I think APFloat is following the same rules here, but I'm not certain all NaN-propagating implementations would follow the same rules.

I've implemented what you suggested and the 4 cases above now verify.
I have zero regressions on the LLVM test suite with the proposed semantics! 🚀

Muon · 2023-09-21T04:59:21Z

I'm pretty sure the significand lives in the least significant bits on every single architecture. For format conversions I'm pretty sure everyone just truncates/zero extends. EDIT: I mean only for NaN handling.

programmerjake · 2023-09-21T05:37:00Z

I'm pretty sure the significand lives in the least significant bits on every single architecture. For format conversions I'm pretty sure everyone just truncates/zero extends.

IEEE754 specifies fp->fp format conversions should use the programmer-chosen rounding mode, for LLVM for f32 <-> f64 that's always round to nearest, ties to even

Muon · 2023-09-21T05:40:29Z

Ah, I meant for NaN handling.

… are preserved

RalfJung · 2023-09-21T07:58:19Z

So for fpext/fptruncate, what does this mean? Should we add something like this?

fpext:

If value is a NaN, then the result will be a NaN with the same sign, with the quiet/signaling bit non-deterministically either preserved or reset to 1, and with the payload extended with trailing (least-significant) zeroes.

(I wasn't sure how to best say "zero-extend on the right", since normal zero-extension happens on the left.)

fptruncate:

If value is a NaN, then the result will be a NaN with the same sign, with the quiet/signaling bit non-deterministically either preserved or reset to 1, and with the payload truncated by removing the least significant bits.

Or is that not correct and the payload may be reset to the preferred payload as well?
Are we sufficiently sure that this is correct for all hardware? It seems to be stricter than what IEEE mandates. The alternative would be to say that the payload is arbitrary.

programmerjake · 2023-09-21T08:11:02Z

IIRC RISC-V converts NaN payloads to the canonical qNaN when converting between fp formats.

RalfJung · 2023-09-21T08:52:10Z

So, like this?

fpext:

If value is a NaN, then the result will be a NaN with the same sign, with the quiet/signaling bit non-deterministically either preserved or reset to 1, and with the payload non-deterministically being all-zero or the payload of value extended with trailing (least-significant) zeroes.

fptruncate:

If value is a NaN, then the result will be a NaN with the same sign, with the quiet/signaling bit non-deterministically either preserved or reset to 1, and with the payload non-deterministically being all-zero or the payload of value truncated by removing the least significant bits.

Though maybe it's easier to just say

fpext+fptruncate:

If value is a NaN, then the result will be a NaN with the same sign, with the quiet/signaling bit non-deterministically either preserved or reset to 1, and with the payload non-deterministically chosen.

Or even

fpext+fptruncate:

If value is a NaN, then the result will be a NaN with the quiet/signaling bit non-deterministically either preserved or reset to 1, and with the sign and payload non-deterministically chosen.

jyknight · 2023-09-21T14:30:30Z

The section "What happens to the payload when the precision is converted?" in This 2018 paper claimed that all known implementations preserve the upper NaN bits when truncating, and append zeros when extending.

Of course, that's only the case for the ISAs which propagate NaN payloads at all (so, not RISC-V and not WebAssembly). On RISCV, the sign is always positive and the NaN canonical. On Wasm, the sign is non-deterministic, and the payload follows Wasm's usual "nondeterministic result if given non-canonical input NaNs, canonical NaN with nondeterministic sign otherwise" rule.

Given that these are the exact same issues that arise for all other FP instructions, I think these instructions should use our normal rules, just clarified to say how the bits are copied if they happen to be copied.

Thus, I suggest the spec should say something like this for fptrunc,

NaN values follow the usual :ref:`NaN behaviors <floatnan>`, except that _if_ a NaN payload is copied from the input, then the low order bits of the NaN payload which cannot fit in the resulting type are discarded.

And for fpext:

NaN values follow the usual :ref:`NaN behaviors <floatnan>`, except that _if_ a NaN payload is copied from the input, then it is copied to the high order bits of the resulting payload, and the remaining low order bits are zero.

RalfJung · 2023-09-21T15:38:47Z

I have adjusted this a bit to clarify what is mean by "if a NaN payload is copied from the input", and pushed that to the PR.

RalfJung · 2023-10-02T15:59:05Z

@nikic @jcranmer-intel @efriedma-quic what are the next steps towards landing this PR?

nunoplopes · 2023-10-02T16:03:38Z

@nikic @jcranmer-intel @efriedma-quic what are the next steps towards landing this PR?

It's approved; just click the button "squash and merge" (or ask us to do it for you) 🚀

RalfJung · 2023-10-02T16:07:40Z

A bunch of changes happened since it got approved though. I'm not familiar with how LLVM handles that situation.

Also I don't have that button. :)

efriedma-quic · 2023-10-03T19:24:19Z

A bunch of changes happened since it got approved though. I'm not familiar with how LLVM handles that situation.

If you think the changes are significant enough to require another review, please point out the specific parts. Otherwise, we can consider the patch ready.

Also I don't have that button. :)

From the developer policy: "If you do not have commit access, please let people know during the review and someone should commit it on your behalf."

jyknight · 2023-10-04T16:24:55Z

Before we merge, though: can you rewrite the commit description (by editing the title&first comment) to accurately describe the current situation?

Thx!

RalfJung · 2023-10-04T18:32:29Z

If you think the changes are significant enough to require another review, please point out the specific parts. Otherwise, we can consider the patch ready.

I find it hard to judge whether the reviewers would consider the changes significant enough. I am surprised that as an outsider to the project, I am expected to make such a call.

I was about to produce a summary of the changes since the approval, but thankfully @jyknight was faster. :)

can you rewrite the commit description (by editing the title&first comment) to accurately describe the current situation?

I've updated the PR description. The PR title still seems fine?

Should I squash the commits?

jyknight · 2023-10-04T19:06:25Z

I've updated the PR description. The PR title still seems fine?

Thanks. I tweaked the title slightly.

Should I squash the commits?

No, you shouldn't -- squash happens during the merge (via the button labeled "Squash and Merge"). I will press said button now.

RalfJung · 2023-10-04T19:38:43Z

Great, thanks a lot. :)

Follow-up to #66579: while implementing those semantics in Miri I realized there's a special case to be considered in truncating float casts.

specify NaN behavior more precisely

ad615db

llvmbot added the llvm:ir label Sep 16, 2023

programmerjake reviewed Sep 17, 2023

View reviewed changes

Muon reviewed Sep 18, 2023

View reviewed changes

llvm/docs/LangRef.rst Outdated Show resolved Hide resolved

nikic requested review from jcranmer-intel and efriedma-quic September 18, 2023 07:00

jcranmer-intel reviewed Sep 18, 2023

View reviewed changes

clarify which part is the payload and which operations are not affect…

91f5076

…ed; add list of known-broken cases

RalfJung force-pushed the float-nanana branch from d4a0c76 to 68a32c0 Compare September 19, 2023 06:57

RalfJung commented Sep 19, 2023

View reviewed changes

programmerjake reviewed Sep 19, 2023

View reviewed changes

llvm/docs/LangRef.rst Outdated Show resolved Hide resolved

mention strictfp and denormal-fp-math

7029109

RalfJung force-pushed the float-nanana branch from 68a32c0 to 7029109 Compare September 19, 2023 07:13

jcranmer-intel reviewed Sep 19, 2023

View reviewed changes

fix nits

e27e3ea

jcranmer-intel approved these changes Sep 19, 2023

View reviewed changes

nunoplopes reviewed Sep 19, 2023

View reviewed changes

llvm/docs/LangRef.rst Show resolved Hide resolved

the target 'extra' NaNs can depend on the inputs in general; clarify …

62e022b

…what this achieves

nunoplopes reviewed Sep 19, 2023

View reviewed changes

llvm/docs/LangRef.rst Show resolved Hide resolved

s-barannikov requested a review from arsenm September 19, 2023 17:11

RalfJung added 2 commits September 21, 2023 09:48

link to issue for x86-32 with SSE calling convention problem

83f182a

fneg, fabs, copysign: explicitly state that NaN quietness and payload…

670b08c

… are preserved

fptrunc and fpext

c14faf3

jyknight approved these changes Oct 4, 2023

View reviewed changes

jyknight changed the title ~~specify NaN behavior more precisely~~ [LangRef] Specify NaN behavior more precisely Oct 4, 2023

jyknight merged commit d688816 into llvm:main Oct 4, 2023
1 check passed

RalfJung deleted the float-nanana branch October 4, 2023 19:38

tnv01 mentioned this pull request Oct 4, 2023

try 2 #68258

Merged

RalfJung mentioned this pull request Oct 9, 2023

clarify NaN propagation in fptrunc #68554

Merged

stepthomas mentioned this pull request Oct 10, 2023

AMDGPU stepthomas atomic csub no rtn forms ver2 stepthomas/llvm-project#1

Closed

jyknight pushed a commit that referenced this pull request Oct 26, 2023

clarify NaN propagation in fptrunc (#68554)

fceb719

Follow-up to #66579: while implementing those semantics in Miri I realized there's a special case to be considered in truncating float casts.

quaternic mentioned this pull request Feb 2, 2024

Clarify documentation on propagating signalling NaNs through (non-strict) fp-operations #80472

Open

RalfJung mentioned this pull request May 2, 2024

Difference between compile-time and runtime float precision on 32-bit x86 without SSE can cause miscompilation leading to segfault #89885

Open

[LangRef] Specify NaN behavior more precisely #66579

[LangRef] Specify NaN behavior more precisely #66579

Conversation

RalfJung commented Sep 16, 2023 • edited Loading

llvmbot commented Sep 16, 2023

Choose a reason for hiding this comment

RalfJung Sep 17, 2023 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

jcranmer-intel left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

RalfJung commented Sep 19, 2023 • edited Loading

Choose a reason for hiding this comment

nunoplopes commented Sep 19, 2023

nunoplopes commented Sep 20, 2023

nunoplopes commented Sep 20, 2023

nunoplopes commented Sep 20, 2023

jcranmer-intel commented Sep 20, 2023

nunoplopes commented Sep 20, 2023

Muon commented Sep 21, 2023 • edited Loading

programmerjake commented Sep 21, 2023

Muon commented Sep 21, 2023

RalfJung commented Sep 21, 2023 • edited Loading

programmerjake commented Sep 21, 2023

RalfJung commented Sep 21, 2023

jyknight commented Sep 21, 2023

RalfJung commented Sep 21, 2023

RalfJung commented Oct 2, 2023

nunoplopes commented Oct 2, 2023

RalfJung commented Oct 2, 2023

efriedma-quic commented Oct 3, 2023

jyknight commented Oct 4, 2023

RalfJung commented Oct 4, 2023

jyknight commented Oct 4, 2023

RalfJung commented Oct 4, 2023

RalfJung commented Sep 16, 2023 •

edited

Loading

RalfJung Sep 17, 2023 •

edited

Loading

RalfJung commented Sep 19, 2023 •

edited

Loading

Muon commented Sep 21, 2023 •

edited

Loading

RalfJung commented Sep 21, 2023 •

edited

Loading