Hi, The following diff fix a bug in LLVM which it triggered when building lang/rust 1.41.0 (the latest version) on i386 (others archs are fine).
LLVM commit: https://github.com/llvm/llvm-project/commit/2153c4b8281c1e5f25887ef9183947198c50a9d2 RedHat has the same problem with rustc on i686 (and I steal the llvm commit from them): - https://bugzilla.redhat.com/show_bug.cgi?id=1797127 - https://github.com/rust-lang/compiler-builtins/issues/338 To reproduce (without rustc, the file is llvm-ir output from rustc): $ cat load.ll ; ModuleID = 'load.3a1fbbbh-cgu.0' source_filename = "load.3a1fbbbh-cgu.0" target datalayout = "e-m:e-p:32:32-f64:32:64-f80:32-n8:16:32-S128" target triple = "i686-unknown-openbsd" ; Function Attrs: nonlazybind uwtable define void @unordered_load(i64* %src) unnamed_addr #0 { start: %0 = alloca i64, align 4 %1 = load atomic i64, i64* %src unordered, align 8 store i64 %1, i64* %0, align 4 %2 = load i64, i64* %0, align 4 br label %bb1 bb1: ; preds = %start ret void } attributes #0 = { nonlazybind uwtable "no-frame-pointer-elim"="true" "probe-stack"="__rust_probestack" "target-cpu"="pentiumpro" } !llvm.module.flags = !{!0} !0 = !{i32 2, !"RtLibUseGOT", i32 1} before (on i386): $ llc load.ll cmpxchg instructions must be atomic. %1 = cmpxchg i64* %src, i64 0, i64 0 unordered not_atomic in function unordered_load LLVM ERROR: Broken function found, compilation aborted! $ after (on i386): $ llc load.ll $ and load.s is correctly generated and buildable. I tested the change only on i386. Comments or OK ? -- Sebastien Marie diff 9a533cbd26c2f04e3fb1c870b1375d429032de85 3a1dd56b086ed90c170aec4e116fbf7e017b86f9 blob - 653f411bb1b6feb3a5a9f2e5ed5ab5e4b7f35f01 (mode 644) blob + aed6bca5abfda531b5e2ebff68215d0210ccd512 (mode 640) --- devel/llvm/Makefile +++ devel/llvm/Makefile @@ -18,7 +18,7 @@ PKGSPEC-main = llvm-=${LLVM_V} PKGNAME-main = llvm-${LLVM_V} PKGNAME-python = py-llvm-${LLVM_V} PKGNAME-lldb = lldb-${LLVM_V} -REVISION-main = 5 +REVISION-main = 6 REVISION-lldb= 0 CATEGORIES = devel blob - /dev/null blob + 3dcf1e3b1e45f8343e33fc3b841760574c14d83c (mode 640) --- /dev/null +++ devel/llvm/patches/patch-lib_CodeGen_AtomicExpandPass_cpp @@ -0,0 +1,16 @@ +$OpenBSD$ +Fix a crash bug when lowering unordered loads to cmpxchg +https://github.com/llvm/llvm-project/commit/2153c4b8281c1e5f25887ef9183947198c50a9d2 + +Index: lib/CodeGen/AtomicExpandPass.cpp +--- lib/CodeGen/AtomicExpandPass.cpp.orig ++++ lib/CodeGen/AtomicExpandPass.cpp +@@ -431,6 +431,8 @@ bool AtomicExpand::expandAtomicLoadToLL(LoadInst *LI) + bool AtomicExpand::expandAtomicLoadToCmpXchg(LoadInst *LI) { + IRBuilder<> Builder(LI); + AtomicOrdering Order = LI->getOrdering(); ++ if (Order == AtomicOrdering::Unordered) ++ Order = AtomicOrdering::Monotonic; + Value *Addr = LI->getPointerOperand(); + Type *Ty = cast<PointerType>(Addr->getType())->getElementType(); + Constant *DummyVal = Constant::getNullValue(Ty);
