[llvm-bugs] [Bug 37860] New: Inline asm fails to notify on bad inline asm eventually crashing

via llvm-bugs llvm-bugs at lists.llvm.org
Tue Jun 19 15:45:42 PDT 2018


https://bugs.llvm.org/show_bug.cgi?id=37860

            Bug ID: 37860
           Summary: Inline asm fails to notify on bad inline asm
                    eventually crashing
           Product: clang
           Version: unspecified
          Hardware: PC
                OS: All
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: -New Bugs
          Assignee: unassignedclangbugs at nondot.org
          Reporter: echristo at gmail.com
                CC: llvm-bugs at lists.llvm.org

Found via a crash on invalid, there was an inline asm that looked like:

__attribute__((target("avx,avx2,fma"))) void Accumulate(__m256 a, __m256 b,
                                                        __m256 *acc) {

  asm("vfmadd132ps %0 %1 (%2)" : : "r"(a), "r"(b), "r"(acc) :);
}

The somewhat more correct version would look like:

__attribute__((target("avx,avx2,fma"))) void Accumulate(__m256 a, __m256 b,
                                                        __m256 *acc) {

  asm("vfmadd132ps %0,%1,%2" : "+x"(a) : "x"(b), "x"(*acc) :);
}


Filing this bug to get better diagnostics and so I don't forget it.

The register constraints are wrong and we should have identified them,
unfortunately the code that looks at that isn't looking at the attributes on
the target to make those determinations and so wouldn't identify either them or
what's going on in a number of places.

The inline asm checker would have failed even on the correct code because it's
using the -msse4.2 command line to do the verification thus thinking that the
'x' constraint corresponds to 128 bits rather than 256 bits.

This means that we need subtarget/verification checking here for target
attribute code in addition to the code that verifies for intrinsics so we don't
incorrectly warn/error on valid code.

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20180619/980f8fc0/attachment.html>


More information about the llvm-bugs mailing list