[llvm] [SPIRV] Start adding support for `int128` (PR #170798)
Alex Voicu via llvm-commits
llvm-commits at lists.llvm.org
Thu Dec 11 05:14:05 PST 2025
AlexVlx wrote:
> > Hi @AlexVlx
> > I noticed that if I build with gcc and -DLLVM_ENABLE_EXPENSIVE_CHECKS=ON many SPIRV lit tests fail with this patch.
> > E.g. `build-all-gcc132-expensive/bin/llc -O0 -verify-machineinstrs -mtriple=spirv64-unknown-unknown --spirv-ext=+SPV_ALTERA_arbitrary_precision_integers,+SPV_KHR_bit_instructions test/CodeGen/SPIRV/llvm-intrinsics/bitreverse_small_type.ll -o -` fails with
> > ```
> > [...]/gcc/13.2.0/include/c++/13.2.0/bits/stl_algo.h:4892:
> > In function:
> > void std::sort(_RAIter, _RAIter, _Compare) [with _RAIter =
> > llvm::StringRef*; _Compare =
> > llvm::SPIRVExtensionsParser::parse(llvm::cl::Option&, llvm::StringRef,
> > llvm::StringRef, std::
> > debug::set<llvm::SPIRV::Extension::Extension>&)::<lambda(auto:4&&,
> > auto:5&&)>]
> >
> > Error: comparison doesn't meet irreflexive requirements, assert(!(a < a)).
> >
> > Objects involved in the operation:
> > instance "functor" @ 0x7ffd92e1a6a0 {
> > }
> > iterator::value_type "ordered type" {
> > }
> > PLEASE submit a bug report to https://github.com/llvm/llvm-project/issues/ and include the crash backtrace and instructions to reproduce the bug.
> > Stack dump:
> > 0. Program arguments: build-all-gcc132-expensive/bin/llc -O0 -verify-machineinstrs -mtriple=spirv64-unknown-unknown --spirv-ext=+SPV_ALTERA_arbitrary_precision_integers,+SPV_KHR_bit_instructions test/CodeGen/SPIRV/llvm-intrinsics/bitreverse_small_type.ll -o -
> > #0 0x0000000004e8fc6e llvm::sys::PrintStackTrace(llvm::raw_ostream&, int) (build-all-gcc132-expensive/bin/llc+0x4e8fc6e)
> > #1 0x0000000004e8c5cb llvm::sys::RunSignalHandlers() (build-all-gcc132-expensive/bin/llc+0x4e8c5cb)
> > #2 0x0000000004e8c70e SignalHandler(int, siginfo_t*, void*) Signals.cpp:0:0
> > #3 0x00007fa6f7f28990 __restore_rt (/lib64/libpthread.so.0+0x12990)
> > #4 0x00007fa6f58bf52f raise (/lib64/libc.so.6+0x4e52f)
> > #5 0x00007fa6f5892e65 abort (/lib64/libc.so.6+0x21e65)
> > #6 0x000000000080410d (build-all-gcc132-expensive/bin/llc+0x80410d)
> > #7 0x000000000246ef29 llvm::SPIRVExtensionsParser::parse(llvm::cl::Option&, llvm::StringRef, llvm::StringRef, std::__debug::set<llvm::SPIRV::Extension::Extension, std::less<llvm::SPIRV::Extension::Extension>, std::allocator<llvm::SPIRV::Extension::Extension>>&) crtstuff.c:0:0
> > #8 0x00000000024257e5 llvm::cl::opt<std::__debug::set<llvm::SPIRV::Extension::Extension, std::less<llvm::SPIRV::Extension::Extension>, std::allocator<llvm::SPIRV::Extension::Extension>>, false, llvm::SPIRVExtensionsParser>::handleOccurrence(unsigned int, llvm::StringRef, llvm::StringRef) crtstuff.c:0:0
> > #9 0x0000000004d4b5c1 llvm::cl::ParseCommandLineOptions(int, char const* const*, llvm::StringRef, llvm::raw_ostream*, llvm::vfs::FileSystem*, char const*, bool) (build-all-gcc132-expensive/bin/llc+0x4d4b5c1)
> > #10 0x0000000000806968 main (build-all-gcc132-expensive/bin/llc+0x806968)
> > #11 0x00007fa6f58ab7e5 __libc_start_main (/lib64/libc.so.6+0x3a7e5)
> > #12 0x00000000009510ee _start (build-all-gcc132-expensive/bin/llc+0x9510ee)
> > Abort (core dumped)
> > ```
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> > I guess it's this change that cause this
> > ```
> > bool SPIRVExtensionsParser::parse(cl::Option &O, StringRef ArgName,
> > StringRef ArgValue,
> > std::set<SPIRV::Extension::Extension> &Vals) {
> > SmallVector<StringRef, 10> Tokens;
> > ArgValue.split(Tokens, ",", -1, false);
> > - std::sort(Tokens.begin(), Tokens.end());
> > + llvm::sort(Tokens, [](auto &&LHS, auto &&RHS) {
> > + // We want to ensure that we handle "all" first, to ensure that any
> > + // subsequent disablement actually behaves as expected i.e. given
> > + // --spv-ext=all,-foo, we first enable all and then disable foo; this should
> > + // be revisited and simplified.
> > + if (LHS == "all")
> > + return true;
> > + if (RHS == "all")
> > + return false;
> > + return !(RHS < LHS);
> > + });
> > ```
>
> @AlexVlx https://en.cppreference.com/w/cpp/algorithm/sort.html says the following:
>
> ```
> comp - comparison function object (i.e. an object that satisfies the requirements of [Compare](https://en.cppreference.com/w/cpp/named_req/Compare.html)) which returns ​true if the first argument is less than (i.e. is ordered before) the second.
> ```
>
> and then https://en.cppreference.com/w/cpp/named_req/Compare.html has the following requirement
>
> ```
> For all a, comp(a, a) == false.
> ```
>
> And with this patch we instead do
>
> ```
> + if (LHS == "all")
> + return true;
> + if (RHS == "all")
> + return false;
> ```
>
> so `comp("all","all") == true` which doesn't follow the requirement.
Thank you for the analysis. #171826 addresses this.
https://github.com/llvm/llvm-project/pull/170798
More information about the llvm-commits
mailing list