aeubanks added a comment.
got IR diffs for llvm-test-suite by dumping the IR in clang after running the
optimization pipeline (-O3), IR file name is md5 of the module identifier
@@@ -1093,6 -1093,6 +1096,16 @@@ void EmitAssemblyHelper::EmitAssembly(B
std::unique_ptr<llvm::ToolOutputFile> ThinLinkOS, DwoOS;
RunOptimizationPipeline(Action, OS, ThinLinkOS);
++ if (!TheModule->getModuleIdentifier().empty()) {
++ std::error_code EC;
++ std::string FileName = "/usr/local/google/home/aeubanks/tmp/wow3/";
++ FileName +=
llvm::utohexstr(llvm::MD5Hash(TheModule->getModuleIdentifier()), false, 16);
++ FileName += ".ll";
++ llvm::raw_fd_ostream OS(FileName, EC);
++ if (EC)
++ report_fatal_error(createStringError(EC, "???"));
++ TheModule->print(OS, nullptr);
++ }
RunCodegenPipeline(Action, OS, DwoOS);
all IR diffs:
https://github.com/aeubanks/llvm-ir-diff/commit/ca19575407a94933d10e9b493784cce47751d0aa
diffs with <50 lines changed:
https://github.com/aeubanks/llvm-ir-diff/commit/5592d7e71191501b6029af1daa1ee438076f4913
I'm going to be out for a couple weeks, but a couple initial observations:
4508D45179E820C5.ll: we don't run instcombine anymore before getting to the
function simplification pipeline, and that causes argpromo to fail. perhaps we
should add back in an early instcombine somewhere. (but the forced noinline in
the file prevents obvious inlining)
1A365B06B0D46C86.ll: `select i1 %37, i32 %38, i32 %35` becomes `call i32
@llvm.smin.i32(i32 %36, i32 4095)`
1ACE620D6E75417F.ll: extra `memset`
A4B000DC4289863A.ll: missed used once global -> alloca
will see if running an early instcombine resolves most of these issues
Repository:
rG LLVM Github Monorepo
CHANGES SINCE LAST ACTION
https://reviews.llvm.org/D145265/new/
https://reviews.llvm.org/D145265
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
