Valgrind does not currently support the "compare and signal" floating-point instructions (KEBR, KDBR, KXBR, KEB, and KDB). This didn't matter much since they weren't emitted by GCC or LLVM until recently. But that has changed; newer GCC and LLVM versions *do* emit these instructions. Thus Valgrind should support them now.
Andreas, given that newer GCC and LLVM do emit these now, it sounds like supporting them is potentially of high priority. What's your view?
(In reply to Julian Seward from comment #1) > Andreas, given that newer GCC and LLVM do emit these now, it sounds like > supporting them is potentially of high priority. What's your view? Absolutely. I have a patch ready, although it doesn't handle the "signalling" part correctly yet. Then again, signalling doesn't seem to be handled correctly for s390x at all yet. Do you know how important the correct handling of FP signals is considered to be on other platforms?
Created attachment 125487 [details] Support compare-and-signal instructions This is the patch I was referring to. It should be sufficient unless accurate FP signal handling is needed.
I tried the patch on koji (remote builder) on Fedora rawhide with GCC10. It seems to cause: op name: CmpF32 op type is (F32,F32) -> I32 arg tys are (F64,F64) IR SANITY CHECK FAILURE in a couple of tests. Unfortunately I don't have direct access to the machine to test better. Also note that this is a straight backport to 3.15.0, not a full git master. https://kojipkgs.fedoraproject.org//work/tasks/1760/41221760/build.log
(In reply to Mark Wielaard from comment #4) > I tried the patch on koji (remote builder) on Fedora rawhide with GCC10. It > seems to cause: > > op name: CmpF32 > op type is (F32,F32) -> I32 > arg tys are (F64,F64) > > IR SANITY CHECK FAILURE > > in a couple of tests. I think, but haven't verified yet, that this is caused by s390_irgen_CxB () taking an IRType for which two IRTemps are created (op1 and op2) which are then used with binop(Iop_CmpF32, mkexpr(op1), mkexpr(op2)). The Iop_CmpF32 is used unconditionally, but the type given to s390_irgen_CxB () can be Ity_F32 or Ity_F64.
Created attachment 125541 [details] updated patch with IROp cmp_op argument for s390_irgen_CxB () Testing this updated patch that adds an IrOp cmp_op argument to s390_irgen_CxB which takes an Iop_CmpF32 or Iop_CmpF64 and uses that for the binop.
Test with the revised patch look good. There are a couple of failures, but none seem related to this issue: == 746 tests, 4 stderr failures, 0 stdout failures, 0 stderrB failures, 0 stdoutB failures, 2 post failures == memcheck/tests/leak_cpp_interior (stderr) memcheck/tests/memcmptest (stderr) memcheck/tests/vbit-test/vbit-test (stderr) drd/tests/tc04_free_lock (stderr) massif/tests/new-cpp (post) massif/tests/overloaded-new (post) https://kojipkgs.fedoraproject.org//work/tasks/6548/41246548/build.log
Created attachment 125542 [details] Updated patch with fixed CxB and added test case Functionally equivalent to the updated patch by Mark. This version also adds testing for the instructions CEB, CDB, KEB, and KDB, which were untested before. The new test would have caught the problem.
Pushed as e83c28e10c99d52b22ee69e25857dac8bf3d5240 after approval via IRC.