Bug #8358: TestSprintf#test_float test failure

• Status: Closed
• Assignee: luislavena (Luis Lavena)
• Target version: 2.2.0
• ruby -v: ruby 2.1.0dev (2013-05-01) [i386-mingw32]
• Backport: 2.0.0: DONE, 2.1: DONE

[phasis68 (Heesob Park)]

I noticed TestSprintf#test_float
http://ci.rubyinstaller.org/job/ruby-trunk-x86-test-all/1287/console

  1) Failure:
TestSprintf#test_float [C:/Users/Worker/Jenkins/workspace/ruby-trunk-x86-build/test/ruby/test_sprintf.rb:193]:
[ruby-dev:42551].
<"0x1p+2"> expected but was
<"0x1p+1">.

This failure is due to r40404.

And Actually, this issue is almost same to bug #8299.
ruby_hdtoa function requires 53-bit precision
but mingw32 compiler is 64-bit precision.

There are 2 possible workarounds.

1. adding -msse2 -mfpmath=sse flag when compiling.
2. adding _control87(_PC_53, _MCW_PC) when running.

[h.shirosaki (Hiroshi Shirosaki)]

The author describes that it is necessary to specify double-precision (53-bit) rounding precision before invoking strtod or dtoa.
https://github.com/ruby/ruby/blob/trunk/util.c#L526

So that I created a patch to set _PC_53 temporally when calling strtod, dtoa and hdtoa on mingw x86.

It seems that SSE2_MATH is defined only if compiling with -msse2 -mfpmath=sse. In the case it avoids to set _PC_53.

This patch would also fix #8299.

https://gist.github.com/shirosaki/5596940

[luislavena (Luis Lavena)]

• Category set to build
• Status changed from Open to Assigned
• Assignee set to naruse (Yui NARUSE)
• Target version set to 2.1.0

naruse: Can we commit the patch?

[naruse (Yui NARUSE)]

• Assignee changed from naruse (Yui NARUSE) to nobu (Nobuyoshi Nakada)

As I wrote in #8299, up to nobu.

Our choice will be one of the following:

• add _control87(_PC_53, _MCW_PC) (or inline assembler) to some functions
• add _control87(_PC_53, _MCW_PC) to Init_Numeric
• remove the test

[phasis68 (Heesob Park)]

Is there any progress about this issue?

This is the only test failure persisted over 2 months.

http://ci.rubyinstaller.org/job/ruby-trunk-x86-test-all/1742/console

[luislavena (Luis Lavena)]

• Priority changed from Normal to 5

nobu: ping?

[vo.x (Vit Ondruch)]

I observe the same issue on Fedora Rawhide i386.

$ ruby -v
ruby 2.1.0dev (2013-09-22 trunk 43011) [i686-linux]

[phasis68 (Heesob Park)]

This is a very long standing test failure since 2013-03-23.
If you think this is a intended behavior and don't want to apply any patch, please remove this test case.

[luislavena (Luis Lavena)]

• Priority changed from 5 to 7

Hello Nobu, Usa,

This is still happening in 2.1.0 release, and the test is blocking me from releasing RubyInstaller.

We need a response on this, is the test valid or not? can be ignored or the issue needs to be fixed?

Thank you.

[nobu (Nobuyoshi Nakada)]

• Status changed from Assigned to Closed
• % Done changed from 0 to 100

This issue was solved with changeset r44474.
Heesob, thank you for reporting this issue.
Your contribution to Ruby is greatly appreciated.
May Ruby be with you.

---

configure.in: use SSE2

• configure.in: use SSE2 instructions for drop unexpected
  precisions. [ruby-core:54738] [Bug #8358]

[vo.x (Vit Ondruch)]

• Subject changed from TestSprintf#test_float test failuer on mingw32 to TestSprintf#test_float test failuer
• Status changed from Closed to Assigned
• Target version changed from 2.1.0 to 2.2.0

Sorry, but this is not Mingw specific issue. I am facing the same issue on Fedora [1]. The test passes on rubyci.org, because there seems to be explicitly specified -msse2 configuration option for some reasons (may be to make this test pass, but that is just speculation), but we are not using this option by default on Fedora i686.

[1] http://kojipkgs.fedoraproject.org//work/tasks/9943/6349943/build.log

[h.shirosaki (Hiroshi Shirosaki)]

In the case of using SSE2, -mstackrealign flag would be required for Windows.

See #8349

[nobu (Nobuyoshi Nakada)]

• Status changed from Assigned to Closed

This issue was solved with changeset r44538.
Heesob, thank you for reporting this issue.
Your contribution to Ruby is greatly appreciated.
May Ruby be with you.

---

configure.in: use SSE2

• configure.in: use SSE2 instructions to drop unexpected precisions on
  other than mingw. [ruby-core:59472] [Bug #8358]

[vo.x (Vit Ondruch)]

• File 0001-Properly-detect-platform-for-SSE2-instructions.patch 0001-Properly-detect-platform-for-SSE2-instructions.patch added
• Status changed from Closed to Assigned

It still does not for for me. It seems that there are two reasons:

1. The [i[4-6]86] if wrongly expanded to i4-686 which does not exist. It should be enclosed in additional brackets.
2. For Fedora, the $target is acutally "i686-redhat-linux-gnu", so there should be wildcard to match the rest of the target string.

The attached patch fixes both issues.

[nobu (Nobuyoshi Nakada)]

• Status changed from Assigned to Closed

Applied in changeset r44890.

---

configure.in: Properly detect platform for SSE2 instructions.

• configure.in: add qouting brackets and append wildcard for the
  rest after target_cpu, to properly detect platform for SSE2
  instructions. [ruby-core:60576] [Bug #8358]

[vo.x (Vit Ondruch)]

Thanks Nobu.

[naruse (Yui NARUSE)]

r44890 breaks 32bit linux
http://c5632.rubyci.org/~chkbuild/ruby-trunk/log/20140208T150302Z.diff.html.gz

[naruse (Yui NARUSE)]

• Status changed from Closed to Assigned

[nobu (Nobuyoshi Nakada)]

Can't you show its config.log file?

[akr (Akira Tanaka)]

I committed r44896 to fix the compilation error.

[zzak (zzak _)]

• Assignee changed from nobu (Nobuyoshi Nakada) to naruse (Yui NARUSE)

@naruse (Yui NARUSE) could you verify that r44896 fixed this on 32bit linux?

[anatolik (Anatol Pomozov)]

I maintain ruby package in Linux Arch and our users reported an issue with it https://bugs.archlinux.org/task/39470

We compile 32-bit binaries for old machines (PentiumPro+ CPU). The old CPUs do not have SSE2 support, but we've found that our 32-bit kernels contain SSE2 operations.

$ objdump -d /usr/lib/libruby.so | grep movapd | head
2c743: 66 0f 28 c8 movapd %xmm0,%xmm1
2c7e5: 66 0f 28 9b 10 89 f4 movapd -0xb76f0(%ebx),%xmm3
2c800: 66 0f 28 f3 movapd %xmm3,%xmm6
2c808: 66 0f 28 fb movapd %xmm3,%xmm7
32676: 66 0f 28 c8 movapd %xmm0,%xmm1
765d5: 66 0f 28 8b 10 89 f4 movapd -0xb76f0(%ebx),%xmm1

After I reverted the whole sse2 block

-    AS_CASE(["$target"],
-       [*-darwin*], [
-           # doesn't seem necessary on Mac OS X
-       ],
-       [[i[4-6]86*]], [
-           RUBY_TRY_CFLAGS(-msse2 -mfpmath=sse, [
-               RUBY_APPEND_OPTION(XCFLAGS, -msse2 -mfpmath=sse)
-           ])
-            AS_CASE(["$XCFLAGS"],
-                [[*-msse2*]], [
-                    RUBY_TRY_CFLAGS(-mstackrealign, [
-                        RUBY_APPEND_OPTION(XCFLAGS, -mstackrealign)
-                    ])
-                ])
-       ]
-    )

i686 binaries do not contain sse2 operations anymore, x86_64 (64 bit binaries for newer CPU) contain sse2 as expected. Tests "make test" are passed.

Here are the compilation flags uses on Arch:

config.status: creating x86_64-linux-fake.rb
	CC = gcc
	LD = ld
	LDSHARED = gcc -shared
	CFLAGS = -march=x86-64 -mtune=generic -O2 -pipe -fstack-protector --param=ssp-buffer-size=4 -fPIC 
	XCFLAGS = -D_FORTIFY_SOURCE=2 -fstack-protector -fno-strict-overflow -fvisibility=hidden -DRUBY_EXPORT
	CPPFLAGS = -D_FORTIFY_SOURCE=2   -I. -I.ext/include/x86_64-linux -I./include -I.
	DLDFLAGS = -Wl,-soname,libruby.so.2.1  -fstack-protector  
	SOLIBS = -lpthread -lgmp -ldl -lcrypt -lm  

Note to "-march=x86-64" it is for 64-bit builds, on 32-bit builds it is "-march=i686". I believe this flags are enough to tell gcc whether sse2 should be enabled. compiler (at least on Linux Arch) does not need additional sse2 compilation flags.

So my question: why to enable sse2 if the binary is compiled for i686 microarchitecture? "i686" is PentiumPro family that does not support sse2, sse2 is supported only starting from pentium4. See http://gcc.gnu.org/onlinedocs/gcc-4.8.2/gcc/i386-and-x86-64-Options.html#i386-and-x86-64-Options

[anatolik (Anatol Pomozov)]

We've decided to revert the whole sse2 block from configure.in in official Linux Arch ruby package. https://www.archlinux.org/packages/extra/x86_64/ruby/

That change breaks our users who uses 32-bit packages on old hardware that does not support SSE2.

[zzak (zzak _)]

• Subject changed from TestSprintf#test_float test failuer to TestSprintf#test_float test failure
• Status changed from Assigned to Feedback
• Assignee changed from naruse (Yui NARUSE) to zzak (zzak _)
• Priority changed from 7 to Normal

[luislavena (Luis Lavena)]

@Zachary

This has been waiting review by Nobu/usa in a long time, why is assigned to you? are you going to revert the SSE2 block?

[hansdegraaff (Hans de Graaff)]

This change is also causing issues for our Gentoo users that still use x86-based systems. See https://bugs.gentoo.org/show_bug.cgi?id=503804 for our downstream bug. We've also dropped the sse2 configure.in block for now to allow compilation on x86 hardware.

[zzak (zzak _)]

• Assignee changed from zzak (zzak _) to nobu (Nobuyoshi Nakada)

@luis (Luis Lopez) I was using assign as a reminder to check back on this ticket, when I have more time to review Anatol's response.

I think nobu will be able to decide about removing the SSE2 code. Admittedly it sucks that downstream package maintainers have to maintain separate configuration.

[h.shirosaki (Hiroshi Shirosaki)]

I found SSE2 is not enabled on mingw because of target='i386-pc-mingw32'.
Here is a patch.

diff --git a/configure.in b/configure.in
index 17ed3ed..cc684b5 100644
--- a/configure.in
+++ b/configure.in
@@ -860,7 +860,7 @@ if test "$GCC" = yes; then
        [*-darwin*], [
            # doesn't seem necessary on Mac OS X
        ],
-       [[i[4-6]86*]], [
+       [[i[4-6]86*|i386*mingw*]], [
            RUBY_TRY_CFLAGS(-msse2 -mfpmath=sse, [
                RUBY_APPEND_OPTION(XCFLAGS, -msse2 -mfpmath=sse)
            ])

[luislavena (Luis Lavena)]

• Status changed from Feedback to Assigned

Hello nobu,

Ping on this? this is still an issue on trunk and ruby_2_1 branch:

TestSprintf#test_float [C:/Users/Luis/Code/ruby/ruby/test/ruby/test_sprintf.rb:198]:
[ruby-dev:42551].
<"0x1p+2"> expected but was
<"0x1p+1">.

(ruby 2.2.0dev (2014-05-11 trunk 45921) [i386-mingw32])

If you don't object, going to apply Shirosaki's patch and request backport to 2.1.

Thank you.

[luislavena (Luis Lavena)]

An update.

Before this patch:

https://gist.github.com/luislavena/beb9bff73fdca800debc

With patch applied:

https://gist.github.com/luislavena/fbb28cac6c53e5bc6256

[nobu (Nobuyoshi Nakada)]

Luis Lavena wrote:

If you don't object, going to apply Shirosaki's patch and request backport to 2.1.

Go ahead.

[nobu (Nobuyoshi Nakada)]

• Description updated (diff)

[zzak (zzak _)]

• Assignee changed from nobu (Nobuyoshi Nakada) to luislavena (Luis Lavena)

• Status changed from Assigned to Closed

Applied in changeset r45954.

---

configure.in: enable SSE2 on mingw

• configure.in: enable SSE2 on mingw. target='i386-pc-mingw32'.
  [ruby-core:62095] [Bug #8358]

[nobu (Nobuyoshi Nakada)]

• Backport changed from 1.9.3: UNKNOWN, 2.0.0: UNKNOWN to 2.0.0: REQUIRED, 2.1: REQUIRED

[nagachika (Tomoyuki Chikanaga)]

• Related to Backport #9495: Bacport #8358 - TestSprintf#test_float test failuer added

[nagachika (Tomoyuki Chikanaga)]

• Backport changed from 2.0.0: REQUIRED, 2.1: REQUIRED to 2.0.0: REQUIRED, 2.1: DONE

r44474, r44538, r44539, r44890 and r44896 were already backported into ruby_2_1 at r44993.
And r45954 was backported at r46467.

[usa (Usaku NAKAMURA)]

• Backport changed from 2.0.0: REQUIRED, 2.1: DONE to 2.0.0: DONE, 2.1: DONE

r44474, r44538, r44539, r44890, r44896 and r45954 were backported into ruby_2_0_0 at r46514.

[vo.x (Vit Ondruch)]

• Related to Bug #10120: TestSprintf#test_float still an issue added