disagreement in converting float32 result to float64 on 32-bit platform #41

JeffBezanson · 2011-06-11T20:39:14Z

This was seen on a 32-bit Xeon system:

julia> ndigf(n)=float64(log(float32(n)))

julia> ndigf(256)
5.5451774444736657

julia> float64(log(float32(256)))
5.5451774597167969

julia> log(256)
5.5451774444795623

The last answer is correct, the middle one is the float64 conversion of the correctly-rounded float32 answer, and the first one is something strange.

julia> num2hex(5.5451774444795623)
"40162e42fefa39ef"

julia> num2hex(5.5451774444736657)
"40162e42fefa2000"

Looks like the float64 answer with the last 13 bits zero'd.

Happens with some functions, e.g. log and sin, but not with sqrt.

The text was updated successfully, but these errors were encountered:

ViralBShah · 2011-06-12T18:16:46Z

Might this have anything to do with the gcc optimization bug mentioned in the fdlibm readme?

JeffBezanson · 2011-06-14T20:24:10Z

Very likely. I tried the libm version and it doesn't have this issue.

StefanKarpinski · 2011-06-15T01:30:57Z

Jeff, can you check if this is fixed by compiling fdlibm with -O1 instead of -O2?

JeffBezanson · 2011-06-15T03:31:38Z

-O1 did not fix this, but it did fix the other one.

JeffBezanson · 2011-06-15T06:46:37Z

What am I smoking, it can't be the same bug, since this is calling logf, the float32 version, which already has the pointer cast bug fixed.

JeffBezanson · 2011-06-15T07:06:25Z

OK, this is just the classic extended precision issue. I can fix it by replacing

return y;

with

volatile float x;
x=y; return x;

in the fdlibm code.
Inside ndigf's llvm-generated code the result stays in an extended-precision register and so can be converted to float64 without first rounding to float32.
So what should we do? I could modify our code generator to force rounding of results returned by C functions.

StefanKarpinski · 2011-06-15T07:11:17Z

I think always rounding is the right thing to do. Consistency and determinism is more important than keeping extra bits around. This same issue has already caused both Java and PHP to have DOS bugs in the past few months, let's learn the lesson from them.

ViralBShah · 2011-06-15T07:39:19Z

I agree that let's go for consistency. That crlibm presentation discusses this issue.

Best to also check with Alan on this issue. He is probably not reading this thread.

-viral

On Jun 15, 2011, at 12:41 PM, StefanKarpinski wrote:

I think always rounding is the right thing to do. Consistency and determinism is more important than keeping extra bits around. This same issue has already caused both Java and PHP to have DOS bugs in the past few months, let's learn the lesson from them.

Reply to this email directly or view it on GitHub:
#41 (comment)

StefanKarpinski · 2011-06-15T07:52:00Z

He can't actually read this thread — this is private and AFAIK, Alan doesn't have a GitHub account.

ArchRobison · 2014-07-23T18:19:21Z

Thanks for adding the test. Alas on i386 systems, using bitcast instead of the volatile store/load sequence flunks the test. I'll poke around some more.

Keno · 2014-07-23T18:23:27Z

We should bring this up on the llvm list.

ArchRobison · 2014-07-23T18:59:31Z

I concur, and sent a note to llvmdev@cs.uiuc.edu.

JeffBezanson · 2014-07-23T19:28:46Z

Maybe try bitcasting through an int32?

ArchRobison · 2014-07-23T22:27:57Z

I think we're going to end up wanting to take different actions depending upon whether the target has excess precision, with the goal of presenting least obfuscated code to the LLVM optimizer. I'm guessing that we can detect the presence of excess precision from LLVM somehow since Clang must detect it to set FLT_EVAL_METHOD for C99.

[pao: fixed link to FLT_EVAL_METHOD]

JeffBezanson · 2014-07-23T23:45:03Z

It seems this is not an issue at all on 64-bit, so maybe the first move is just to disable this hack on anything x86-64.

…ence only on 32-bit X86.

…g#41 to use volatile store/load on 32-bit x86 only.

Fix issue #7687, by revising original fix for issue #41

Document recent PRs

Fix depwarns on 0.7

Fix atbash-cipher test sets indentation

Fix world age problems when running tests

Without this change, the compiler fails to notice that `env_threads isa Int` in the fall-through case, leading to a union-split with a branch that is in fact unreachable: ``` 43 ┄ %109 = φ (#41 => %105, #42 => %108)::Union{Nothing, Int64} │ %110 = (%109 isa Int64)::Bool └─── goto #45 if not %110 ... 45 ─ %126 = π (%109, Nothing) │ Base.convert(Int64, %126)::Union{} └─── unreachable ``` Co-authored-by: Jeff Bezanson <jeff.bezanson@gmail.com>

Without this change, the compiler fails to notice that `env_threads isa Int` in the fall-through case, leading to a union-split with a branch that is in fact unreachable: ``` 43 ┄ %109 = φ (#41 => %105, #42 => %108)::Union{Nothing, Int64} │ %110 = (%109 isa Int64)::Bool └─── goto #45 if not %110 ... 45 ─ %126 = π (%109, Nothing) │ Base.convert(Int64, %126)::Union{} └─── unreachable ``` After this change, the union-split is eliminated. Co-authored-by: Jeff Bezanson <jeff.bezanson@gmail.com>

Without this change, the compiler fails to notice that `env_threads isa Int` in the fall-through case, leading to a union-split with a branch that is in fact unreachable: ``` 43 ┄ %109 = φ (JuliaLang#41 => %105, JuliaLang#42 => %108)::Union{Nothing, Int64} │ %110 = (%109 isa Int64)::Bool └─── goto JuliaLang#45 if not %110 ... 45 ─ %126 = π (%109, Nothing) │ Base.convert(Int64, %126)::Union{} └─── unreachable ``` After this change, the union-split is eliminated. Co-authored-by: Jeff Bezanson <jeff.bezanson@gmail.com>

JeffBezanson added a commit that referenced this issue Jun 15, 2011

fixing issue #41, force rounding of Float32 where needed

13fc340

JeffBezanson closed this as completed Jun 15, 2011

ViralBShah mentioned this issue Mar 19, 2012

WIP: libuv rewrite and Windows port #521

Merged

timholy mentioned this issue Jun 30, 2012

Reliable segfault with >4d array #993

Closed

This was referenced Jul 13, 2012

Function access on parallel processors different between interactive and script #1046

Closed

Segfault on parallel program #1054

Closed

dslate mentioned this issue Jul 24, 2012

printf: "%.0f" formatting bug #1081

Closed

vtjnash added a commit that referenced this issue Jan 8, 2013

fix open_any_tcp_port. fixes #41

998facc

ArchRobison mentioned this issue Jul 21, 2014

Hungry for more SIMD vectorization #7687

Closed

JeffBezanson added a commit that referenced this issue Jul 22, 2014

add test for issue #41

bcfa49f

ArchRobison pushed a commit to ArchRobison/julia that referenced this issue Jul 24, 2014

Revise solution to issue JuliaLang#41 to use volatile store/load sequ…

fe4e3c3

…ence only on 32-bit X86.

ArchRobison pushed a commit to ArchRobison/julia that referenced this issue Jul 24, 2014

Fix issue JuliaLang#7687, by revising original fix for issue JuliaLan…

08767a4

…g#41 to use volatile store/load on 32-bit x86 only.

ArchRobison pushed a commit to ArchRobison/julia that referenced this issue Jul 24, 2014

Fix issue JuliaLang#7687, by revising original fix for issue JuliaLan…

448f86d

…g#41 to use volatile store/load on 32-bit x86 only.

ArchRobison pushed a commit to ArchRobison/julia that referenced this issue Jul 24, 2014

Fix issue JuliaLang#7687, by revising original fix for issue JuliaLan…

3f77ae0

…g#41 to use volatile store/load on 32-bit x86 only.

ArchRobison mentioned this issue Jul 24, 2014

Fix issue #7687, by revising original fix for issue #41 #7719

Merged

JeffBezanson added a commit that referenced this issue Jul 24, 2014

Merge pull request #7719 from ArchRobison/adr/fpext

9d922bc

Fix issue #7687, by revising original fix for issue #41

ViralBShah added the 32-bit label Jul 27, 2014

StefanKarpinski pushed a commit that referenced this issue Feb 8, 2018

Merge pull request #41 from garborg/patch-1

a292237

Document recent PRs

StefanKarpinski pushed a commit that referenced this issue Feb 8, 2018

Merge pull request #41 from yuyichao/0.7

e5f8a5c

Fix depwarns on 0.7

cmcaine pushed a commit to cmcaine/julia that referenced this issue Sep 24, 2020

Merge pull request JuliaLang#41 from aawray/patch-1

8d0d1f2

Fix atbash-cipher test sets indentation

chflood mentioned this issue Mar 9, 2022

OOM not always signaled properly in Fedora 33 #44539

Closed

Keno pushed a commit that referenced this issue Oct 9, 2023

Merge pull request #41 from JuliaDebug/teh/world_age

7705307

Fix world age problems when running tests

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

disagreement in converting float32 result to float64 on 32-bit platform #41

disagreement in converting float32 result to float64 on 32-bit platform #41

JeffBezanson commented Jun 11, 2011

ViralBShah commented Jun 12, 2011

JeffBezanson commented Jun 14, 2011

StefanKarpinski commented Jun 15, 2011

JeffBezanson commented Jun 15, 2011

JeffBezanson commented Jun 15, 2011

JeffBezanson commented Jun 15, 2011

StefanKarpinski commented Jun 15, 2011

ViralBShah commented Jun 15, 2011

StefanKarpinski commented Jun 15, 2011

ArchRobison commented Jul 23, 2014

Keno commented Jul 23, 2014

ArchRobison commented Jul 23, 2014

JeffBezanson commented Jul 23, 2014

ArchRobison commented Jul 23, 2014

JeffBezanson commented Jul 23, 2014

disagreement in converting float32 result to float64 on 32-bit platform #41

disagreement in converting float32 result to float64 on 32-bit platform #41

Comments

JeffBezanson commented Jun 11, 2011

ViralBShah commented Jun 12, 2011

JeffBezanson commented Jun 14, 2011

StefanKarpinski commented Jun 15, 2011

JeffBezanson commented Jun 15, 2011

JeffBezanson commented Jun 15, 2011

JeffBezanson commented Jun 15, 2011

StefanKarpinski commented Jun 15, 2011

ViralBShah commented Jun 15, 2011

StefanKarpinski commented Jun 15, 2011

ArchRobison commented Jul 23, 2014

Keno commented Jul 23, 2014

ArchRobison commented Jul 23, 2014

JeffBezanson commented Jul 23, 2014

ArchRobison commented Jul 23, 2014

JeffBezanson commented Jul 23, 2014