[valgrind] [Bug 442168] Floating-point erroneous behavior with memcheck with regards to class (isfinite(), etc.)
https://bugs.kde.org/show_bug.cgi?id=442168 --- Comment #7 from Xavier Roche --- (In reply to Julian Seward from comment #4) > (In reply to Xavier Roche from comment #2) > > > The difference between the correctly executed code under valgrind and the > > faulty one: > > > - movsd %xmm0, (%rsp) # 8-byte Spill > > + vmovsd %xmm0, (%rsp) # 8-byte Spill > > Can you give some more information about why you think the change from > movsd to vmovsd causes the error? Also, which one gives correct execution > and which doesn't? The correct one is movsd; the faulty one is vmovsd. This only happens with clang-12 [12.0.0-3ubuntu1~20.04.3] (not clang-11) and only when running valgrind. g++-10 (Ubuntu 10.3.0-1ubuntu1~20.04) does not trigger the problem either. So this is a specific clang-12 issue. As for the reason, I'm not fluent enough on x86-64 assembly to make any educated guess unfortunately. This might have nothing to do with vmovsd -- You are receiving this mail because: You are watching all bug changes.
[valgrind] [Bug 442168] Floating-point erroneous behavior with memcheck with regards to class (isfinite(), etc.)
https://bugs.kde.org/show_bug.cgi?id=442168 --- Comment #6 from Xavier Roche --- Created attachment 141820 --> https://bugs.kde.org/attachment.cgi?id=141820=edit ASM diff between working program with valgrind (clang-11) and the faulty one (clang12) clang-12 -S -std=c++20 -stdlib=libc++ -O2 -march=corei7-avx isfinite-bug.cpp -o /tmp/isfinite-bug-12.S clang-11 -S -std=c++20 -stdlib=libc++ -O2 -march=corei7-avx isfinite-bug.cpp -o /tmp/isfinite-bug-11.S diff -udb /tmp/isfinite-bug-11.S /tmp/isfinite-bug-12.S > /tmp/isfinite-bug-clang-11-clang-12.diff -- You are receiving this mail because: You are watching all bug changes.
[valgrind] [Bug 442168] Floating-point erroneous behavior with memecheck with regards to class (isfinite(), etc.)
https://bugs.kde.org/show_bug.cgi?id=442168
--- Comment #2 from Xavier Roche ---
Additional notes: isolating the faulty function
static double (*volatile pvalidate)(double a) = validate; // used in
minimalTest
double validate(double a)
{
return std::isfinite(a) ? a : 1.0f;
}
The difference between the correctly executed code under valgrind and the
faulty one:
_Z8validated: # @_Z8validated
pushq %rax
.cfi_def_cfa_offset 16
- movsd %xmm0, (%rsp) # 8-byte Spill
+ vmovsd %xmm0, (%rsp) # 8-byte Spill
callq
_Z8isfiniteIdENSt3__19enable_ifIXaasr3std13is_arithmeticIT_EE5valuesr3std14numeric_limitsIS2_EE12has_infinityEbE4typeES2_
- movsd (%rsp), %xmm0 # 8-byte Reload
+ vmovsd (%rsp), %xmm0 # 8-byte Reload
# xmm0 = mem[0],zero
testb %al, %al
jne .LBB6_2
# %bb.1:
- movsd .LCPI6_0(%rip), %xmm0 # xmm0 = mem[0],zero
+ vmovsd .LCPI6_0(%rip), %xmm0 # xmm0 = mem[0],zero
.LBB6_2:
popq%rax
retq
--
You are receiving this mail because:
You are watching all bug changes.
[valgrind] [Bug 442168] Floating-point erroneous behavior with memecheck with regards to class (isfinite(), etc.)
https://bugs.kde.org/show_bug.cgi?id=442168 --- Comment #1 from Xavier Roche --- Bug also present with valgrind-3.18.0.GIT (cadf0432290b0bc147c7b5dd54c63bc94986743c) -- You are receiving this mail because: You are watching all bug changes.
[valgrind] [Bug 442168] New: Floating-point erroneous behavior with memecheck with regards to class (isfinite(), etc.)
https://bugs.kde.org/show_bug.cgi?id=442168 Bug ID: 442168 Summary: Floating-point erroneous behavior with memecheck with regards to class (isfinite(), etc.) Product: valgrind Version: 3.15 SVN Platform: Other OS: Linux Status: REPORTED Severity: normal Priority: NOR Component: memcheck Assignee: jsew...@acm.org Reporter: xavier.ro...@algolia.com Target Milestone: --- Created attachment 141383 --> https://bugs.kde.org/attachment.cgi?id=141383=edit Minimal program demonstrating floating-point erroneous behavior with memcheck SUMMARY Floating-point erroneous behavior with memecheck with regards to class (isfinite(), etc.), leading to consider -NaN as a finite number. STEPS TO REPRODUCE 1. Compile the attached reproducible minimal case with clang-12 in at least O2 and -march=corei7-avx clang-12 -std=c++20 -stdlib=libc++ -O2 -march=corei7-avx isfinite-bug-with-clang12-O2-corei7-avx.cpp -lc++ -lm -o isfinite-bug 2. Run without valgrind ./isfinite-bug 3. Run with valgrind (valgrind --tool=memcheck) valgrind --tool=memcheck ./isfinite-bug OBSERVED RESULT Run without valgrind: "All right" is emitted Run with valgrind: "Error: expected 1 and got -nan" is emitted EXPECTED RESULT The valgrind version should always emit "All right" SOFTWARE/OS VERSIONS Linux, Ubuntu, 5.4.0-80-generic ADDITIONAL INFORMATION The issue is reproduced using valgrind-3.15.0 It requires: * clang-12 * At least O2 * -march=corei7-avx It is _not_ reproduced with an earlier version of clang (tested version: clang-11), which hints of a recent special NaN/fp class optimization. At this stage, I can not guarantee that this is a valgrind issue, or a undefined-behavior-generated-code that happen to work without valgrind. I'm available for any additional information. -- You are receiving this mail because: You are watching all bug changes.
