Actually setup jit targets when compiling packageimages instead of targeting only one (#54471)

gbaraldi · web-flow · commit ad407a6d2198 · 2024-07-11T10:06:25.000+02:00
diff --git a/src/codegen.cpp b/src/codegen.cpp
@@ -7438,8 +7438,11 @@ static Function* gen_cfun_wrapper(
     ctx.builder.ClearInsertionPoint();
 
     if (aliasname) {
-        GlobalAlias::create(cw->getValueType(), cw->getType()->getAddressSpace(),
+        auto alias = GlobalAlias::create(cw->getValueType(), cw->getType()->getAddressSpace(),
                             GlobalValue::ExternalLinkage, aliasname, cw, M);
+        if(ctx.emission_context.TargetTriple.isOSBinFormatCOFF()) {
+            alias->setDLLStorageClass(GlobalValue::DLLStorageClassTypes::DLLExportStorageClass);
+        }
     }
 
     if (nest) {
diff --git a/src/llvm-multiversioning.cpp b/src/llvm-multiversioning.cpp
@@ -677,6 +677,7 @@ void CloneCtx::rewrite_alias(GlobalAlias *alias, Function *F)
     trampoline->removeFnAttr("julia.mv.reloc");
     trampoline->removeFnAttr("julia.mv.clones");
     trampoline->addFnAttr("julia.mv.alias");
+    trampoline->setDLLStorageClass(alias->getDLLStorageClass());
     alias->eraseFromParent();
 
     uint32_t id;
diff --git a/src/processor_arm.cpp b/src/processor_arm.cpp
@@ -1890,12 +1890,56 @@ const std::pair<std::string,std::string> &jl_get_llvm_disasm_target(void)
     return res;
 }
 
+#ifndef __clang_gcanalyzer__
 llvm::SmallVector<jl_target_spec_t, 0> jl_get_llvm_clone_targets(void)
 {
-    if (jit_targets.empty())
-        jl_error("JIT targets not initialized");
+
+    auto &cmdline = get_cmdline_targets();
+    check_cmdline(cmdline, true);
+    llvm::SmallVector<TargetData<feature_sz>, 0> image_targets;
+    for (auto &arg: cmdline) {
+        auto data = arg_target_data(arg, image_targets.empty());
+        image_targets.push_back(std::move(data));
+    }
+    auto ntargets = image_targets.size();
+    if (image_targets.empty())
+        jl_error("No targets specified");
     llvm::SmallVector<jl_target_spec_t, 0> res;
-    for (auto &target: jit_targets) {
+    // Now decide the clone condition.
+    for (size_t i = 1; i < ntargets; i++) {
+        auto &t = image_targets[i];
+        if (t.en.flags & JL_TARGET_CLONE_ALL)
+            continue;
+        auto &features0 = image_targets[t.base].en.features;
+        // Always clone when code checks CPU features
+        t.en.flags |= JL_TARGET_CLONE_CPU;
+        static constexpr uint32_t clone_fp16[] = {Feature::fp16fml,Feature::fullfp16};
+        for (auto fe: clone_fp16) {
+            if (!test_nbit(features0, fe) && test_nbit(t.en.features, fe)) {
+                t.en.flags |= JL_TARGET_CLONE_FLOAT16;
+                break;
+            }
+        }
+        // The most useful one in general...
+        t.en.flags |= JL_TARGET_CLONE_LOOP;
+#ifdef _CPU_ARM_
+        static constexpr uint32_t clone_math[] = {Feature::vfp3, Feature::vfp4, Feature::neon};
+        for (auto fe: clone_math) {
+            if (!test_nbit(features0, fe) && test_nbit(t.en.features, fe)) {
+                t.en.flags |= JL_TARGET_CLONE_MATH;
+                break;
+            }
+        }
+        static constexpr uint32_t clone_simd[] = {Feature::neon};
+        for (auto fe: clone_simd) {
+            if (!test_nbit(features0, fe) && test_nbit(t.en.features, fe)) {
+                t.en.flags |= JL_TARGET_CLONE_SIMD;
+                break;
+            }
+        }
+#endif
+    }
+    for (auto &target: image_targets) {
         auto features_en = target.en.features;
         auto features_dis = target.dis.features;
         for (auto &fename: feature_names) {
@@ -1916,6 +1960,8 @@ llvm::SmallVector<jl_target_spec_t, 0> jl_get_llvm_clone_targets(void)
     return res;
 }
 
+#endif
+
 extern "C" int jl_test_cpu_feature(jl_cpu_feature_t feature)
 {
     if (feature >= 32 * feature_sz)
diff --git a/src/processor_fallback.cpp b/src/processor_fallback.cpp
@@ -144,13 +144,27 @@ const std::pair<std::string,std::string> &jl_get_llvm_disasm_target(void)
                 jl_get_cpu_features_llvm(), {{}, 0}, {{}, 0}, 0});
     return res;
 }
-
+#ifndef __clang_gcanalyzer__
 llvm::SmallVector<jl_target_spec_t, 0> jl_get_llvm_clone_targets(void)
 {
-    if (jit_targets.empty())
-        jl_error("JIT targets not initialized");
+
+    auto &cmdline = get_cmdline_targets();
+    check_cmdline(cmdline, true);
+    llvm::SmallVector<TargetData<1>, 0> image_targets;
+    for (auto &arg: cmdline) {
+        auto data = arg_target_data(arg, image_targets.empty());
+        image_targets.push_back(std::move(data));
+    }
+    auto ntargets = image_targets.size();
+    // Now decide the clone condition.
+    for (size_t i = 1; i < ntargets; i++) {
+        auto &t = image_targets[i];
+        t.en.flags |= JL_TARGET_CLONE_ALL;
+    }
+    if (image_targets.empty())
+        jl_error("No image targets found");
     llvm::SmallVector<jl_target_spec_t, 0> res;
-    for (auto &target: jit_targets) {
+    for (auto &target: image_targets) {
         jl_target_spec_t ele;
         std::tie(ele.cpu_name, ele.cpu_features) = get_llvm_target_str(target);
         ele.data = serialize_target_data(target.name, target.en.features,
@@ -161,6 +175,7 @@ llvm::SmallVector<jl_target_spec_t, 0> jl_get_llvm_clone_targets(void)
     }
     return res;
 }
+#endif
 
 JL_DLLEXPORT jl_value_t *jl_cpu_has_fma(int bits)
 {
diff --git a/src/processor_x86.cpp b/src/processor_x86.cpp
@@ -910,6 +910,8 @@ static uint32_t pkgimg_init_cb(const void *id, jl_value_t **rejection_reason)
     return match.best_idx;
 }
 
+//This function serves as a fallback during bootstrapping, at that point we don't have a sysimage with native code
+// so we won't call sysimg_init_cb, else this function shouldn't do anything.
 static void ensure_jit_target(bool imaging)
 {
     auto &cmdline = get_cmdline_targets();
@@ -1102,13 +1104,82 @@ const std::pair<std::string,std::string> &jl_get_llvm_disasm_target(void)
             {feature_masks, 0}, {{}, 0}, 0});
     return res;
 }
-
+//This function parses the -C command line to figure out which targets to multiversion to.
+#ifndef __clang_gcanalyzer__
 llvm::SmallVector<jl_target_spec_t, 0> jl_get_llvm_clone_targets(void)
 {
-    if (jit_targets.empty())
-        jl_error("JIT targets not initialized");
+    auto &cmdline = get_cmdline_targets();
+    check_cmdline(cmdline, true);
+    llvm::SmallVector<TargetData<feature_sz>, 0> image_targets;
+    for (auto &arg: cmdline) {
+        auto data = arg_target_data(arg, image_targets.empty());
+        image_targets.push_back(std::move(data));
+    }
+
+    auto ntargets = image_targets.size();
+    // Now decide the clone condition.
+    for (size_t i = 1; i < ntargets; i++) {
+        auto &t = image_targets[i];
+        if (t.en.flags & JL_TARGET_CLONE_ALL)
+            continue;
+        // Always clone when code checks CPU features
+        t.en.flags |= JL_TARGET_CLONE_CPU;
+        // The most useful one in general...
+        t.en.flags |= JL_TARGET_CLONE_LOOP;
+        auto &features0 = image_targets[t.base].en.features;
+        // Special case for KNL/KNM since they're so different
+        if (!(t.dis.flags & JL_TARGET_CLONE_ALL)) {
+            if ((t.name == "knl" || t.name == "knm") &&
+                image_targets[t.base].name != "knl" && image_targets[t.base].name != "knm") {
+                t.en.flags |= JL_TARGET_CLONE_ALL;
+                break;
+            }
+        }
+        static constexpr uint32_t clone_math[] = {Feature::fma, Feature::fma4};
+        static constexpr uint32_t clone_simd[] = {Feature::sse3, Feature::ssse3,
+                                                  Feature::sse41, Feature::sse42,
+                                                  Feature::avx, Feature::avx2,
+                                                  Feature::vaes, Feature::vpclmulqdq,
+                                                  Feature::sse4a, Feature::avx512f,
+                                                  Feature::avx512dq, Feature::avx512ifma,
+                                                  Feature::avx512pf, Feature::avx512er,
+                                                  Feature::avx512cd, Feature::avx512bw,
+                                                  Feature::avx512vl, Feature::avx512vbmi,
+                                                  Feature::avx512vpopcntdq, Feature::avxvnni,
+                                                  Feature::avx512vbmi2, Feature::avx512vnni,
+                                                  Feature::avx512bitalg, Feature::avx512bf16,
+                                                  Feature::avx512vp2intersect, Feature::avx512fp16};
+        for (auto fe: clone_math) {
+            if (!test_nbit(features0, fe) && test_nbit(t.en.features, fe)) {
+                t.en.flags |= JL_TARGET_CLONE_MATH;
+                break;
+            }
+        }
+        for (auto fe: clone_simd) {
+            if (!test_nbit(features0, fe) && test_nbit(t.en.features, fe)) {
+                t.en.flags |= JL_TARGET_CLONE_SIMD;
+                break;
+            }
+        }
+        static constexpr uint32_t clone_fp16[] = {Feature::avx512fp16};
+        for (auto fe: clone_fp16) {
+            if (!test_nbit(features0, fe) && test_nbit(t.en.features, fe)) {
+                t.en.flags |= JL_TARGET_CLONE_FLOAT16;
+                break;
+            }
+        }
+        static constexpr uint32_t clone_bf16[] = {Feature::avx512bf16};
+        for (auto fe: clone_bf16) {
+            if (!test_nbit(features0, fe) && test_nbit(t.en.features, fe)) {
+                t.en.flags |= JL_TARGET_CLONE_BFLOAT16;
+                break;
+            }
+        }
+    }
+    if (image_targets.empty())
+        jl_error("No targets specified");
     llvm::SmallVector<jl_target_spec_t, 0> res;
-    for (auto &target: jit_targets) {
+    for (auto &target: image_targets) {
         auto features_en = target.en.features;
         auto features_dis = target.dis.features;
         for (auto &fename: feature_names) {
@@ -1128,6 +1199,7 @@ llvm::SmallVector<jl_target_spec_t, 0> jl_get_llvm_clone_targets(void)
     }
     return res;
 }
+#endif
 
 extern "C" int jl_test_cpu_feature(jl_cpu_feature_t feature)
 {
diff --git a/test/precompile.jl b/test/precompile.jl
@@ -1963,6 +1963,19 @@ precompile_test_harness("Test flags") do load_path
     @test !Base.isprecompiled(id, ;flags=current_flags)
 end
 
+if Base.get_bool_env("CI", false) && (Sys.ARCH === :x86_64 || Sys.ARCH === :aarch64)
+    @testset "Multiversioning" begin # This test isn't the most robust because it relies on being in CI,
+        pkg = Base.identify_package("Test")  # but we need better target reflection to make a better one.
+        cachefiles = Base.find_all_in_cache_path(pkg)
+        pkgpath = Base.locate_package(pkg)
+        idx = findfirst(cachefiles) do cf
+            Base.stale_cachefile(pkgpath, cf) !== true
+        end
+        targets = Base.parse_image_targets(Base.parse_cache_header(cachefiles[idx])[7])
+        @test length(targets) > 1
+    end
+end
+
 precompile_test_harness("No backedge precompile") do load_path
     # Test that the system doesn't accidentally forget to revalidate a method without backedges
     write(joinpath(load_path, "NoBackEdges.jl"),
