Drop support for LLVM 20

.github/workflows/pip.yml

@@ -8,7 +8,7 @@ on:
 
 env:
   # TODO: detect this from repo somehow: https://github.com/halide/Halide/issues/8406
-  LLVM_VERSION: 21.1.8
+  LLVM_VERSION: 22.1.0
   FLATBUFFERS_VERSION: 23.5.26
   WABT_VERSION: 1.0.39
 

.github/workflows/testing-arm-linux.yml

@@ -20,7 +20,7 @@ jobs:
       fail-fast: false
       matrix:
         bits: [ "64", "32" ]
-        uv_group: [ "ci-llvm-main", "ci-llvm-22", "ci-llvm-21", "ci-llvm-20" ]
+        uv_group: [ "ci-llvm-main", "ci-llvm-22", "ci-llvm-21" ]
         include:
           - bits: 32
             arch: armv7l

.github/workflows/testing-linux.yml

@@ -20,7 +20,7 @@ jobs:
       fail-fast: false
       matrix:
         bits: [ "32" ]  # Intentionally not 64, as we haven't configured self-hosted runners for it yet.
-        uv_group: [ "ci-llvm-main", "ci-llvm-22", "ci-llvm-21", "ci-llvm-20" ]
+        uv_group: [ "ci-llvm-main", "ci-llvm-22", "ci-llvm-21" ]
         include:
           - bits: 32
             arch: i686

.github/workflows/testing-make.yml

@@ -1,7 +1,7 @@
 name: Makefile
 
 env:
-  LLVM_VERSION: '21'
+  LLVM_VERSION: '22'
 
 on:
   pull_request:

.github/workflows/testing-windows.yml

@@ -24,7 +24,7 @@ jobs:
       fail-fast: false
       matrix:
         bits: [ "32" ]  # Intentionally not 64, as we haven't configured self-hosted runners for it yet.
-        uv_group: [ "ci-llvm-main", "ci-llvm-22", "ci-llvm-21", "ci-llvm-20" ]
+        uv_group: [ "ci-llvm-main", "ci-llvm-22", "ci-llvm-21" ]
         include:
           - bits: 32
             arch: x86

CMakeLists.txt

@@ -199,7 +199,7 @@ option(THREADS_PREFER_PTHREAD_FLAG "When enabled, prefer to use the -pthread fla
 find_package(Threads REQUIRED)
 
 ## LLVM
-find_package(Halide_LLVM 20...99 REQUIRED  # Use 99 to fake a minimum-only constraint
+find_package(Halide_LLVM 21...99 REQUIRED  # Use 99 to fake a minimum-only constraint
              COMPONENTS WebAssembly X86
              OPTIONAL_COMPONENTS AArch64 ARM Hexagon NVPTX PowerPC RISCV)
 

README.md

@@ -151,7 +151,7 @@ issue.
 
 At any point in time, building Halide requires either the latest stable version
 of LLVM, the previous stable version of LLVM, or trunk. At the time of writing,
-this means versions 22, 21, and 20 are supported, but 19 is not.
+this means versions 23, 22, and 21 are supported, but 20 is not.
 
 It is simplest to get a binary release of LLVM on macOS by using
 [Homebrew](https://brew.sh). Just run `brew install llvm`. On Debian flavors of

cmake/FindHalide_LLVM.cmake

@@ -17,7 +17,7 @@ foreach (
     "/usr/local/lib/llvm-@VERSION@" # Third-party packages
     "/opt/llvm-@VERSION@" # Third-party packages
 )
-    foreach (VERSION RANGE 20 22) # inclusive!
+    foreach (VERSION RANGE 21 23) # inclusive!
         string(CONFIGURE "${template}" path @ONLY)
         list(APPEND llvm_paths "${path}")
     endforeach ()
@@ -28,7 +28,7 @@ find_package(LLVM PATHS ${llvm_paths})
 
 # Neither LLVM_VERSION nor LLVM_PACKAGE_VERSION work as find_package arguments
 # in git/development builds as they include a "git" suffix. This applies at
-# time of writing to versions 18-21, inclusive.
+# time of writing to versions 21-23, inclusive.
 if (LLVM_FOUND)
     set(Halide_LLVM_VERSION "${LLVM_VERSION_MAJOR}.${LLVM_VERSION_MINOR}.${LLVM_VERSION_PATCH}")
 endif ()

pyproject.toml

@@ -90,7 +90,6 @@ ci-base = [
 ci-llvm-main = [{ include-group = "ci-base" }, "halide-llvm~=23.0.0.dev0"]
 ci-llvm-22 = [{ include-group = "ci-base" }, "halide-llvm~=22.1.0"]
 ci-llvm-21 = [{ include-group = "ci-base" }, "halide-llvm~=21.1.0"]
-ci-llvm-20 = [{ include-group = "ci-base" }, "halide-llvm~=20.1.0"]
 
 [project.urls]
 Homepage = "https://halide-lang.org"

src/CodeGen_Internal.cpp

@@ -612,9 +612,6 @@ void get_target_options(const llvm::Module &module, llvm::TargetOptions &options
 
     options = llvm::TargetOptions();
     options.AllowFPOpFusion = per_instruction_fast_math_flags ? llvm::FPOpFusion::Strict : llvm::FPOpFusion::Fast;
-#if LLVM_VERSION < 210
-    options.UnsafeFPMath = !per_instruction_fast_math_flags;
-#endif
 #if LLVM_VERSION < 230
     options.NoInfsFPMath = !per_instruction_fast_math_flags;
 #endif
@@ -676,11 +673,7 @@ std::unique_ptr<llvm::TargetMachine> make_target_machine(const llvm::Module &mod
     std::string mattrs = get_modflag_string(module, "halide_mattrs");
 
     auto *tm = llvm_target->createTargetMachine(
-#if LLVM_VERSION >= 210
         triple,
-#else
-        triple.str(),
-#endif
         mcpu_target,
         mattrs,
         options,

src/CodeGen_LLVM.cpp

@@ -407,11 +407,7 @@ void CodeGen_LLVM::init_codegen(const std::string &name) {
 
     internal_assert(module && context);
 
-#if LLVM_VERSION >= 210
     debug(1) << "Target triple of initial module: " << module->getTargetTriple().str() << "\n";
-#else
-    debug(1) << "Target triple of initial module: " << module->getTargetTriple() << "\n";
-#endif
 
     module->setModuleIdentifier(name);
 
@@ -3397,11 +3393,7 @@ void CodeGen_LLVM::visit(const Call *op) {
             llvm::BasicBlock *block = llvm::BasicBlock::Create(module->getContext(), "entry", fn);
             IRBuilderBase::InsertPoint here = builder->saveIP();
             builder->SetInsertPoint(block);
-#if LLVM_VERSION >= 210
             Value *ret = builder->CreateVScale(i32_t);
-#else
-            Value *ret = builder->CreateVScale(ConstantInt::get(i32_t, 1));
-#endif
             builder->CreateRet(ret);
 
             // To avoid vscale_range(n,n) added in CodeGen_Internal

src/CodeGen_PTX_Dev.cpp

@@ -628,9 +628,6 @@ vector<char> CodeGen_PTX_Dev::compile_to_src() {
 
     TargetOptions options;
     options.AllowFPOpFusion = CodeGen_GPU_Dev::any_strict_float ? llvm::FPOpFusion::Strict : llvm::FPOpFusion::Fast;
-#if LLVM_VERSION < 210
-    options.UnsafeFPMath = !CodeGen_GPU_Dev::any_strict_float;
-#endif
 #if LLVM_VERSION < 230
     options.NoInfsFPMath = !CodeGen_GPU_Dev::any_strict_float;
 #endif
@@ -641,11 +638,7 @@ vector<char> CodeGen_PTX_Dev::compile_to_src() {
 
     std::unique_ptr<TargetMachine>
         target_machine(llvm_target->createTargetMachine(
-#if LLVM_VERSION >= 210
             triple,
-#else
-            triple.str(),
-#endif
             mcpu_target(), mattrs(), options,
             llvm::Reloc::PIC_,
             llvm::CodeModel::Small,

src/JITModule.cpp

@@ -364,11 +364,7 @@ void compile_module_impl(
 
     // Make the execution engine
     debug(2) << "Creating new execution engine\n";
-#if LLVM_VERSION >= 210
     debug(2) << "Target triple: " << m->getTargetTriple().str() << "\n";
-#else
-    debug(2) << "Target triple: " << m->getTargetTriple() << "\n";
-#endif
     string error_string;
 
     llvm::for_each(*m, set_function_attributes_from_halide_target_options);
@@ -412,29 +408,15 @@ void compile_module_impl(
         // i386: "JIT session error: Unsupported i386 relocation:4" (R_386_PLT32)
         // ARM 32bit: Unsupported target machine architecture in ELF object shared runtime-jitted-objectbuffer
         // Windows 64-bit: JIT session error: could not register eh-frame: __register_frame function not found
-#if LLVM_VERSION >= 210
         linkerBuilder = [&](llvm::orc::ExecutionSession &session) {
             return std::make_unique<llvm::orc::RTDyldObjectLinkingLayer>(session, [&](const llvm::MemoryBuffer &) {
                 return std::make_unique<HalideJITMemoryManager>(dependencies);
             });
         };
-#else
-        linkerBuilder = [&](llvm::orc::ExecutionSession &session, const llvm::Triple &) {
-            return std::make_unique<llvm::orc::RTDyldObjectLinkingLayer>(session, [&]() {
-                return std::make_unique<HalideJITMemoryManager>(dependencies);
-            });
-        };
-#endif
     } else {
-#if LLVM_VERSION >= 210
         linkerBuilder = [](llvm::orc::ExecutionSession &session) {
             return std::make_unique<llvm::orc::ObjectLinkingLayer>(session);
         };
-#else
-        linkerBuilder = [](llvm::orc::ExecutionSession &session, const llvm::Triple &) {
-            return std::make_unique<llvm::orc::ObjectLinkingLayer>(session);
-        };
-#endif
     }
 
     auto JIT = llvm::cantFail(llvm::orc::LLJITBuilder()

src/LLVM_Headers.h

@@ -1,10 +1,10 @@
 #ifndef HALIDE_LLVM_HEADERS_H
 #define HALIDE_LLVM_HEADERS_H
 
-#if LLVM_VERSION >= 200
+#if LLVM_VERSION >= 210
 // We're good to go
 #else
-#error "Compiling Halide requires LLVM 20.0 or newer"
+#error "Compiling Halide requires LLVM 21.0 or newer"
 #endif
 
 // No msvc warnings from llvm headers please

src/LLVM_Output.cpp

@@ -353,11 +353,7 @@ void emit_file(const llvm::Module &module_in, Internal::LLVMOStream &out,
                llvm::CodeGenFileType file_type) {
     // Make sure to run this with a large stack!
     debug(1) << "emit_file.Compiling to native code...\n";
-#if LLVM_VERSION >= 210
     debug(2) << "Target triple: " << module_in.getTargetTriple().str() << "\n";
-#else
-    debug(2) << "Target triple: " << module_in.getTargetTriple() << "\n";
-#endif
 
     auto time_start = std::chrono::high_resolution_clock::now();
 

src/LLVM_Runtime_Linker.cpp

@@ -630,11 +630,7 @@ void link_modules(std::vector<std::unique_ptr<llvm::Module>> &modules, Target t,
             }
         }
         module->setDataLayout(data_layout);
-#if LLVM_VERSION >= 210
         module->setTargetTriple(triple);
-#else
-        module->setTargetTriple(triple.str());
-#endif
     }
 
     // Link them all together
@@ -1420,11 +1416,7 @@ std::unique_ptr<llvm::Module> get_initial_module_for_ptx_device(Target target, l
     }
 
     llvm::Triple triple("nvptx64--");
-#if LLVM_VERSION >= 210
     modules[0]->setTargetTriple(triple);
-#else
-    modules[0]->setTargetTriple(triple.str());
-#endif
 
     llvm::DataLayout dl("e-i64:64-v16:16-v32:32-n16:32:64");
     modules[0]->setDataLayout(dl);

test/autoschedulers/mullapudi2016/histogram.cpp

@@ -126,15 +126,6 @@ int main(int argc, char **argv) {
         return 0;
     }
 
-    // LLVM 20 generates poor SVE2 code for the histogram autoscheduler test,
-    // causing it to exceed the performance threshold. Fixed in LLVM 21+.
-    if (Internal::get_llvm_version() < 210 &&
-        get_jit_target_from_environment().has_feature(Target::SVE2)) {
-        printf("[SKIP] LLVM %d generates suboptimal SVE2 code for this test.\n",
-               Internal::get_llvm_version());
-        return 0;
-    }
-
     if (argc != 2) {
         fprintf(stderr, "Usage: %s <autoscheduler-lib>\n", argv[0]);
         return 1;

test/autoschedulers/mullapudi2016/max_filter.cpp

@@ -131,8 +131,7 @@ int main(int argc, char **argv) {
     // LLVM 21 calls getFixedValue() on scalable TypeSize objects in the
     // AArch64 backend, triggering an assertion. Fixed in LLVM 22 by:
     // https://github.com/llvm/llvm-project/commit/d1500d12be60 (PR #169764)
-    if (Internal::get_llvm_version() >= 210 &&
-        Internal::get_llvm_version() < 220 &&
+    if (Internal::get_llvm_version() < 220 &&
         get_jit_target_from_environment().has_feature(Target::SVE2)) {
         printf("[SKIP] LLVM 21 has known getFixedValue() assertion failures on SVE scalable types.\n");
         return 0;

test/correctness/align_bounds.cpp

@@ -26,8 +26,7 @@ int main(int argc, char **argv) {
     // LLVM 21 calls getFixedValue() on scalable TypeSize objects in the
     // AArch64 backend, triggering an assertion. Fixed in LLVM 22 by:
     // https://github.com/llvm/llvm-project/commit/d1500d12be60 (PR #169764)
-    if (Internal::get_llvm_version() >= 210 &&
-        Internal::get_llvm_version() < 220 &&
+    if (Internal::get_llvm_version() < 220 &&
         get_jit_target_from_environment().has_feature(Target::SVE2)) {
         printf("[SKIP] LLVM 21 has known getFixedValue() assertion failures on SVE scalable types.\n");
         return 0;

test/correctness/bound_small_allocations.cpp

@@ -13,8 +13,7 @@ int main(int argc, char **argv) {
     // LLVM 21 calls getFixedValue() on scalable TypeSize objects in the
     // AArch64 backend, triggering an assertion. Fixed in LLVM 22 by:
     // https://github.com/llvm/llvm-project/commit/d1500d12be60 (PR #169764)
-    if (Internal::get_llvm_version() >= 210 &&
-        Internal::get_llvm_version() < 220 &&
+    if (Internal::get_llvm_version() < 220 &&
         get_jit_target_from_environment().has_feature(Target::SVE2)) {
         printf("[SKIP] LLVM 21 has known getFixedValue() assertion failures on SVE scalable types.\n");
         return 0;

test/correctness/fit_function.cpp

@@ -10,8 +10,7 @@ int main(int argc, char **argv) {
     // LLVM 21 calls getFixedValue() on scalable TypeSize objects in the
     // AArch64 backend, triggering an assertion. Fixed in LLVM 22 by:
     // https://github.com/llvm/llvm-project/commit/d1500d12be60 (PR #169764)
-    if (Internal::get_llvm_version() >= 210 &&
-        Internal::get_llvm_version() < 220 &&
+    if (Internal::get_llvm_version() < 220 &&
         get_jit_target_from_environment().has_feature(Target::SVE2)) {
         printf("[SKIP] LLVM 21 has known getFixedValue() assertion failures on SVE scalable types.\n");
         return 0;

test/correctness/logical.cpp

@@ -15,8 +15,7 @@ int main(int argc, char **argv) {
     // LLVM 21 calls getFixedValue() on scalable TypeSize objects in the
     // AArch64 backend, triggering an assertion. Fixed in LLVM 22 by:
     // https://github.com/llvm/llvm-project/commit/d1500d12be60 (PR #169764)
-    if (Internal::get_llvm_version() >= 210 &&
-        Internal::get_llvm_version() < 220 &&
+    if (Internal::get_llvm_version() < 220 &&
         get_jit_target_from_environment().has_feature(Target::SVE2)) {
         printf("[SKIP] LLVM 21 has known getFixedValue() assertion failures on SVE scalable types.\n");
         return 0;

test/correctness/predicated_store_load.cpp

@@ -490,8 +490,7 @@ int main(int argc, char **argv) {
     // LLVM 21 calls getFixedValue() on scalable TypeSize objects in the
     // AArch64 backend, triggering an assertion. Fixed in LLVM 22 by:
     // https://github.com/llvm/llvm-project/commit/d1500d12be60 (PR #169764)
-    if (Internal::get_llvm_version() >= 210 &&
-        Internal::get_llvm_version() < 220 &&
+    if (Internal::get_llvm_version() < 220 &&
         t.has_feature(Target::SVE2)) {
         printf("[SKIP] LLVM 21 has known getFixedValue() assertion failures on SVE scalable types.\n");
         return 0;

test/correctness/truncated_pyramid.cpp

@@ -9,8 +9,7 @@ int main(int argc, char **argv) {
     // LLVM 21 calls getFixedValue() on scalable TypeSize objects in the
     // AArch64 backend, triggering an assertion. Fixed in LLVM 22 by:
     // https://github.com/llvm/llvm-project/commit/d1500d12be60 (PR #169764)
-    if (Internal::get_llvm_version() >= 210 &&
-        Internal::get_llvm_version() < 220 &&
+    if (Internal::get_llvm_version() < 220 &&
         get_jit_target_from_environment().has_feature(Target::SVE2)) {
         printf("[SKIP] LLVM 21 has known getFixedValue() assertion failures on SVE scalable types.\n");
         return 0;

test/correctness/vectorize_nested.cpp

@@ -280,8 +280,7 @@ int main(int argc, char **argv) {
     // LLVM 21 calls getFixedValue() on scalable TypeSize objects in the
     // AArch64 backend, triggering an assertion. Fixed in LLVM 22 by:
     // https://github.com/llvm/llvm-project/commit/d1500d12be60 (PR #169764)
-    if (Internal::get_llvm_version() >= 210 &&
-        Internal::get_llvm_version() < 220 &&
+    if (Internal::get_llvm_version() < 220 &&
         get_jit_target_from_environment().has_feature(Target::SVE2)) {
         printf("[SKIP] LLVM 21 has known getFixedValue() assertion failures on SVE scalable types.\n");
         return 0;

test/performance/block_transpose.cpp

@@ -118,14 +118,6 @@ int main(int argc, char **argv) {
         return 0;
     }
 
-    // LLVM 20 is missing llvm.vector.interleave4/deinterleave4 for SVE.
-    if (Internal::get_llvm_version() < 210 &&
-        target.has_feature(Target::SVE2)) {
-        printf("[SKIP] LLVM %d is missing SVE vector interleave/deinterleave intrinsics.\n",
-               Internal::get_llvm_version());
-        return 0;
-    }
-
     test_transpose(scalar_trans);
     test_transpose_wrap(scalar_trans);
     test_transpose(vec_y_trans);

test/performance/const_division.cpp

@@ -128,8 +128,7 @@ int main(int argc, char **argv) {
     // LLVM 21 calls getFixedValue() on scalable TypeSize objects in the
     // AArch64 backend, triggering an assertion. Fixed in LLVM 22 by:
     // https://github.com/llvm/llvm-project/commit/d1500d12be60 (PR #169764)
-    if (Internal::get_llvm_version() >= 210 &&
-        Internal::get_llvm_version() < 220 &&
+    if (Internal::get_llvm_version() < 220 &&
         target.has_feature(Target::SVE2)) {
         printf("[SKIP] LLVM 21 has known getFixedValue() assertion failures on SVE scalable types.\n");
         return 0;

test/performance/nested_vectorization_gemm.cpp

@@ -11,14 +11,6 @@ int main(int argc, char **argv) {
         return 0;
     }
 
-    // LLVM 20 is missing llvm.vector.interleave4/deinterleave4 for SVE.
-    if (Internal::get_llvm_version() < 210 &&
-        target.has_feature(Target::SVE2)) {
-        printf("[SKIP] LLVM %d is missing SVE vector interleave/deinterleave intrinsics.\n",
-               Internal::get_llvm_version());
-        return 0;
-    }
-
     // 8-bit mat-mul into 32-bit accumulator
     {