drop llvm 16 support (#56751)

Co-authored-by: Valentin Churavy <[email protected]>

Author: vtjnash

doc/src/devdocs/build/build.md

@@ -249,7 +249,7 @@ The most complicated dependency is LLVM, for which we require additional patches
 For packaging Julia with LLVM, we recommend either:
  - bundling a Julia-only LLVM library inside the Julia package, or
  - adding the patches to the LLVM package of the distribution.
-   * A complete list of patches is available in on [Github](https://github.com/JuliaLang/llvm-project) see the `julia-release/15.x` branch.
+   * A complete list of patches is available in on [Github](https://github.com/JuliaLang/llvm-project) see the `julia-release/18.x` branch.
    * The only Julia-specific patch is the lib renaming (`llvm7-symver-jlprefix.patch`), which should _not_ be applied to a system LLVM.
    * The remaining patches are all upstream bug fixes, and have been contributed into upstream LLVM.
 

src/APInt-C.cpp

@@ -475,7 +475,6 @@ void LLVMTrunc(jl_datatype_t *ty, integerPart *pa, jl_datatype_t *otys, integerP
     memcpy(pr, pa, onumbytes);
 }
 
-#if JL_LLVM_VERSION >= 170000
 extern "C" JL_DLLEXPORT
 unsigned countr_zero_8(uint8_t Val) {
     return countr_zero(Val);
@@ -495,27 +494,6 @@ extern "C" JL_DLLEXPORT
 unsigned countr_zero_64(uint64_t Val) {
     return countr_zero(Val);
 }
-#else
-extern "C" JL_DLLEXPORT
-unsigned countTrailingZeros_8(uint8_t Val) {
-    return countTrailingZeros(Val);
-}
-
-extern "C" JL_DLLEXPORT
-unsigned countTrailingZeros_16(uint16_t Val) {
-    return countTrailingZeros(Val);
-}
-
-extern "C" JL_DLLEXPORT
-unsigned countTrailingZeros_32(uint32_t Val) {
-    return countTrailingZeros(Val);
-}
-
-extern "C" JL_DLLEXPORT
-unsigned countTrailingZeros_64(uint64_t Val) {
-    return countTrailingZeros(Val);
-}
-#endif
 
 extern "C" JL_DLLEXPORT
 void jl_LLVMSMod(unsigned numbits, integerPart *pa, integerPart *pb, integerPart *pr) {
@@ -545,7 +523,6 @@ void jl_LLVMFlipSign(unsigned numbits, integerPart *pa, integerPart *pb, integer
         memcpy(pr, pa, numbytes);
 }
 
-#if JL_LLVM_VERSION >= 170000
 extern "C" JL_DLLEXPORT
 unsigned LLVMPopcount(unsigned numbits, integerPart *pa) {
     CREATE(a)
@@ -575,34 +552,3 @@ unsigned LLVMCountl_zero(unsigned numbits, integerPart *pa) {
     CREATE(a)
     return a.countl_zero();
 }
-#else
-extern "C" JL_DLLEXPORT
-unsigned LLVMCountPopulation(unsigned numbits, integerPart *pa) {
-    CREATE(a)
-    return a.countPopulation();
-}
-
-extern "C" JL_DLLEXPORT
-unsigned LLVMCountTrailingOnes(unsigned numbits, integerPart *pa) {
-    CREATE(a)
-    return a.countTrailingOnes();
-}
-
-extern "C" JL_DLLEXPORT
-unsigned LLVMCountTrailingZeros(unsigned numbits, integerPart *pa) {
-    CREATE(a)
-    return a.countTrailingZeros();
-}
-
-extern "C" JL_DLLEXPORT
-unsigned LLVMCountLeadingOnes(unsigned numbits, integerPart *pa) {
-    CREATE(a)
-    return a.countLeadingOnes();
-}
-
-extern "C" JL_DLLEXPORT
-unsigned LLVMCountLeadingZeros(unsigned numbits, integerPart *pa) {
-    CREATE(a)
-    return a.countLeadingZeros();
-}
-#endif

src/APInt-C.h

@@ -54,19 +54,11 @@ JL_DLLEXPORT int LLVMDiv_uov(unsigned numbits, integerPart *pa, integerPart *pb,
 JL_DLLEXPORT int LLVMRem_sov(unsigned numbits, integerPart *pa, integerPart *pb, integerPart *pr);
 JL_DLLEXPORT int LLVMRem_uov(unsigned numbits, integerPart *pa, integerPart *pb, integerPart *pr);
 
-#if JL_LLVM_VERSION >= 170000
 JL_DLLEXPORT unsigned LLVMPopcount(unsigned numbits, integerPart *pa);
 JL_DLLEXPORT unsigned LLVMCountr_one(unsigned numbits, integerPart *pa);
 JL_DLLEXPORT unsigned LLVMCountr_zero(unsigned numbits, integerPart *pa);
 JL_DLLEXPORT unsigned LLVMCountl_one(unsigned numbits, integerPart *pa);
 JL_DLLEXPORT unsigned LLVMCountl_zero(unsigned numbits, integerPart *pa);
-#else
-JL_DLLEXPORT unsigned LLVMCountPopulation(unsigned numbits, integerPart *pa);
-JL_DLLEXPORT unsigned LLVMCountTrailingOnes(unsigned numbits, integerPart *pa);
-JL_DLLEXPORT unsigned LLVMCountTrailingZeros(unsigned numbits, integerPart *pa);
-JL_DLLEXPORT unsigned LLVMCountLeadingOnes(unsigned numbits, integerPart *pa);
-JL_DLLEXPORT unsigned LLVMCountLeadingZeros(unsigned numbits, integerPart *pa);
-#endif
 
 JL_DLLEXPORT void LLVMFPtoSI(jl_datatype_t *ty, integerPart *pa, jl_datatype_t *oty, integerPart *pr);
 JL_DLLEXPORT void LLVMFPtoUI(jl_datatype_t *ty, integerPart *pa, jl_datatype_t *oty, integerPart *pr);
@@ -82,17 +74,10 @@ JL_DLLEXPORT int LLVMFPtoUI_exact(jl_datatype_t *ty, integerPart *pa, jl_datatyp
 JL_DLLEXPORT void jl_LLVMSMod(unsigned numbits, integerPart *pa, integerPart *pb, integerPart *pr);
 JL_DLLEXPORT void jl_LLVMFlipSign(unsigned numbits, integerPart *pa, integerPart *pb, integerPart *pr);
 
-#if JL_LLVM_VERSION >= 170000
 JL_DLLEXPORT unsigned countr_zero_8(uint8_t Val);
 JL_DLLEXPORT unsigned countr_zero_16(uint16_t Val);
 JL_DLLEXPORT unsigned countr_zero_32(uint32_t Val);
 JL_DLLEXPORT unsigned countr_zero_64(uint64_t Val);
-#else
-JL_DLLEXPORT unsigned countTrailingZeros_8(uint8_t Val);
-JL_DLLEXPORT unsigned countTrailingZeros_16(uint16_t Val);
-JL_DLLEXPORT unsigned countTrailingZeros_32(uint32_t Val);
-JL_DLLEXPORT unsigned countTrailingZeros_64(uint64_t Val);
-#endif
 
 //uint8_t getSwappedBytes_8(uint8_t Value); // no-op
 //uint16_t getSwappedBytes_16(uint16_t Value);

src/Makefile

@@ -77,11 +77,7 @@ else
 # JULIACODEGEN != LLVM
 endif
 
-RT_LLVM_LIBS := support
-
-ifeq ($(shell test $(LLVM_VER_MAJ) -ge 16 && echo true),true)
-RT_LLVM_LIBS += targetparser
-endif
+RT_LLVM_LIBS := support targetparser
 
 ifeq ($(OS),WINNT)
 SRCS += win32_ucontext

src/aotcompile.cpp

@@ -4,11 +4,7 @@
 #include "platform.h"
 
 // target support
-#if JL_LLVM_VERSION >= 170000
 #include <llvm/TargetParser/Triple.h>
-#else
-#include <llvm/ADT/Triple.h>
-#endif
 #include "llvm/Support/CodeGen.h"
 #include <llvm/ADT/Statistic.h>
 #include <llvm/Analysis/TargetLibraryInfo.h>
@@ -963,11 +959,7 @@ static FunctionInfo getFunctionWeight(const Function &F)
         auto val = F.getFnAttribute("julia.mv.clones").getValueAsString();
         // base16, so must be at most 4 * length bits long
         // popcount gives number of clones
-        #if JL_LLVM_VERSION >= 170000
         info.clones = APInt(val.size() * 4, val, 16).popcount() + 1;
-        #else
-        info.clones = APInt(val.size() * 4, val, 16).countPopulation() + 1;
-        #endif
     }
     info.weight += info.insts;
     // more basic blocks = more complex than just sum of insts,
@@ -1372,7 +1364,7 @@ static AOTOutputs add_output_impl(Module &M, TargetMachine &SourceTM, ShardTimer
             // So for now we inject a definition of these functions that calls our runtime
             // functions. We do so after optimization to avoid cloning these functions.
             // Float16 conversion routines
-#if defined(_CPU_X86_64_) && defined(_OS_DARWIN_) && JL_LLVM_VERSION >= 160000
+#if defined(_CPU_X86_64_) && defined(_OS_DARWIN_)
             // LLVM 16 reverted to soft-float ABI for passing half on x86_64 Darwin
             // https://github.com/llvm/llvm-project/commit/2bcf51c7f82ca7752d1bba390a2e0cb5fdd05ca9
             injectCRTAlias(M, "__gnu_h2f_ieee", "julia_half_to_float",
@@ -1721,9 +1713,6 @@ static SmallVector<AOTOutputs, 16> add_output(Module &M, TargetMachine &TM, Stri
             std::function<void()> func = [&, i]() {
                 LLVMContext ctx;
                 ctx.setDiscardValueNames(true);
-                #if JL_LLVM_VERSION < 170000
-                SetOpaquePointer(ctx);
-                #endif
                 // Lazily deserialize the entire module
                 timers[i].deserialize.startTimer();
                 auto EM = getLazyBitcodeModule(MemoryBufferRef(StringRef(serialized.data(), serialized.size()), "Optimized"), ctx);
@@ -1888,7 +1877,7 @@ void jl_dump_native_impl(void *native_code,
             Str += "10.14.0";
         TheTriple.setOSName(Str);
     }
-    Optional<Reloc::Model> RelocModel;
+    std::optional<Reloc::Model> RelocModel;
     if (TheTriple.isOSLinux() || TheTriple.isOSFreeBSD() || TheTriple.isOSOpenBSD()) {
         RelocModel = Reloc::PIC_;
     }
@@ -1933,9 +1922,6 @@ void jl_dump_native_impl(void *native_code,
         JL_TIMING(NATIVE_AOT, NATIVE_Sysimg);
         LLVMContext Context;
         Context.setDiscardValueNames(true);
-        #if JL_LLVM_VERSION < 170000
-        SetOpaquePointer(Context);
-        #endif
         Module sysimgM("sysimg", Context);
         sysimgM.setTargetTriple(TheTriple.str());
         sysimgM.setDataLayout(DL);
@@ -2081,9 +2067,6 @@ void jl_dump_native_impl(void *native_code,
         JL_TIMING(NATIVE_AOT, NATIVE_Metadata);
         LLVMContext Context;
         Context.setDiscardValueNames(true);
-        #if JL_LLVM_VERSION < 170000
-        SetOpaquePointer(Context);
-        #endif
         Module metadataM("metadata", Context);
         metadataM.setTargetTriple(TheTriple.str());
         metadataM.setDataLayout(DL);
@@ -2299,16 +2282,7 @@ void jl_get_llvmf_defn_impl(jl_llvmf_dump_t* dump, jl_method_instance_t *mi, jl_
                 }
                 else {
                     auto p = literal_static_pointer_val(global.first, global.second->getValueType());
-                    #if JL_LLVM_VERSION >= 170000
                     Type *elty = PointerType::get(output.getContext(), 0);
-                    #else
-                    Type *elty;
-                    if (p->getType()->isOpaquePointerTy()) {
-                        elty = PointerType::get(output.getContext(), 0);
-                    } else {
-                        elty = p->getType()->getNonOpaquePointerElementType();
-                    }
-                    #endif
                     // For pretty printing, when LLVM inlines the global initializer into its loads
                     auto alias = GlobalAlias::create(elty, 0, GlobalValue::PrivateLinkage, global.second->getName() + ".jit", p, global.second->getParent());
                     global.second->setInitializer(ConstantExpr::getBitCast(alias, global.second->getValueType()));

src/ccall.cpp

@@ -264,9 +264,6 @@ static GlobalVariable *emit_plt_thunk(
     SmallVector<Value*, 16> args;
     for (auto &arg : plt->args())
         args.push_back(&arg);
-    #if JL_LLVM_VERSION < 170000
-    assert(cast<PointerType>(ptr->getType())->isOpaqueOrPointeeTypeMatches(functype));
-    #endif
     CallInst *ret = irbuilder.CreateCall(
         functype,
         ptr, ArrayRef<Value*>(args));

src/cgutils.cpp

@@ -60,11 +60,7 @@ static Value *decay_derived(jl_codectx_t &ctx, Value *V)
     if (T->getPointerAddressSpace() == AddressSpace::Derived)
         return V;
     // Once llvm deletes pointer element types, we won't need it here any more either.
-    #if JL_LLVM_VERSION >= 170000
     Type *NewT = PointerType::get(T, AddressSpace::Derived);
-    #else
-    Type *NewT = PointerType::getWithSamePointeeType(cast<PointerType>(T), AddressSpace::Derived);
-    #endif
     return ctx.builder.CreateAddrSpaceCast(V, NewT);
 }
 
@@ -74,11 +70,7 @@ static Value *maybe_decay_tracked(jl_codectx_t &ctx, Value *V)
     Type *T = V->getType();
     if (T->getPointerAddressSpace() != AddressSpace::Tracked)
         return V;
-    #if JL_LLVM_VERSION >= 170000
     Type *NewT = PointerType::get(T, AddressSpace::Derived);
-    #else
-    Type *NewT = PointerType::getWithSamePointeeType(cast<PointerType>(T), AddressSpace::Derived);
-    #endif
     return ctx.builder.CreateAddrSpaceCast(V, NewT);
 }
 
@@ -1010,54 +1002,6 @@ static void emit_memcpy_llvm(jl_codectx_t &ctx, Value *dst, jl_aliasinfo_t const
 {
     if (sz == 0)
         return;
-    #if JL_LLVM_VERSION < 170000
-    // If the types are small and simple, use load and store directly.
-    // Going through memcpy can cause LLVM (e.g. SROA) to create bitcasts between float and int
-    // that interferes with other optimizations.
-    // TODO: Restore this for opaque pointers? Needs extra type information from the caller.
-    if (ctx.builder.getContext().supportsTypedPointers() && sz <= 64) {
-        // The size limit is arbitrary but since we mainly care about floating points and
-        // machine size vectors this should be enough.
-        const DataLayout &DL = jl_Module->getDataLayout();
-        auto srcty = cast<PointerType>(src->getType());
-        //TODO unsafe nonopaque pointer
-        auto srcel = srcty->getNonOpaquePointerElementType();
-        auto dstty = cast<PointerType>(dst->getType());
-        //TODO unsafe nonopaque pointer
-        auto dstel = dstty->getNonOpaquePointerElementType();
-        while (srcel->isArrayTy() && srcel->getArrayNumElements() == 1) {
-            src = ctx.builder.CreateConstInBoundsGEP2_32(srcel, src, 0, 0);
-            srcel = srcel->getArrayElementType();
-            srcty = srcel->getPointerTo();
-        }
-        while (dstel->isArrayTy() && dstel->getArrayNumElements() == 1) {
-            dst = ctx.builder.CreateConstInBoundsGEP2_32(dstel, dst, 0, 0);
-            dstel = dstel->getArrayElementType();
-            dstty = dstel->getPointerTo();
-        }
-
-        llvm::Type *directel = nullptr;
-        if (srcel->isSized() && srcel->isSingleValueType() && DL.getTypeStoreSize(srcel) == sz) {
-            directel = srcel;
-            dst = emit_bitcast(ctx, dst, srcty);
-        }
-        else if (dstel->isSized() && dstel->isSingleValueType() &&
-                DL.getTypeStoreSize(dstel) == sz) {
-            directel = dstel;
-            src = emit_bitcast(ctx, src, dstty);
-        }
-        if (directel) {
-            if (isa<Instruction>(src) && !src->hasName())
-                setName(ctx.emission_context, src, "memcpy_refined_src");
-            if (isa<Instruction>(dst) && !dst->hasName())
-                setName(ctx.emission_context, dst, "memcpy_refined_dst");
-            auto val = src_ai.decorateInst(ctx.builder.CreateAlignedLoad(directel, src, MaybeAlign(align_src), is_volatile));
-            dst_ai.decorateInst(ctx.builder.CreateAlignedStore(val, dst, align_dst, is_volatile));
-            ++SkippedMemcpys;
-            return;
-        }
-    }
-    #endif
     ++EmittedMemcpys;
 
     // the memcpy intrinsic does not allow to specify different alias tags
@@ -1940,7 +1884,7 @@ static std::pair<Value*, bool> emit_isa(jl_codectx_t &ctx, const jl_cgval_t &x,
     // actual `isa` calls, this optimization should already have been performed upstream
     // anyway, but having this optimization in codegen might still be beneficial for
     // `typeassert`s if we can make it correct.
-    Optional<bool> known_isa;
+    std::optional<bool> known_isa;
     jl_value_t *intersected_type = type;
     if (x.constant)
         known_isa = jl_isa(x.constant, type);
@@ -3786,11 +3730,7 @@ static void recursively_adjust_ptr_type(llvm::Value *Val, unsigned FromAS, unsig
     for (auto *User : Val->users()) {
         if (isa<GetElementPtrInst>(User)) {
             GetElementPtrInst *Inst = cast<GetElementPtrInst>(User);
-            #if JL_LLVM_VERSION >= 170000
             Inst->mutateType(PointerType::get(Inst->getType(), ToAS));
-            #else
-            Inst->mutateType(PointerType::getWithSamePointeeType(cast<PointerType>(Inst->getType()), ToAS));
-            #endif
             recursively_adjust_ptr_type(Inst, FromAS, ToAS);
         }
         else if (isa<IntrinsicInst>(User)) {
@@ -3799,11 +3739,7 @@ static void recursively_adjust_ptr_type(llvm::Value *Val, unsigned FromAS, unsig
         }
         else if (isa<BitCastInst>(User)) {
             BitCastInst *Inst = cast<BitCastInst>(User);
-            #if JL_LLVM_VERSION >= 170000
             Inst->mutateType(PointerType::get(Inst->getType(), ToAS));
-            #else
-            Inst->mutateType(PointerType::getWithSamePointeeType(cast<PointerType>(Inst->getType()), ToAS));
-            #endif
             recursively_adjust_ptr_type(Inst, FromAS, ToAS);
         }
     }