Add support for arm64e (Fixes #4490)

python_bindings/src/halide/halide_/PyEnums.cpp

@@ -189,6 +189,7 @@ void define_enums(py::module &m) {
         .value("ARMv87a", Target::Feature::ARMv87a)
         .value("ARMv88a", Target::Feature::ARMv88a)
         .value("ARMv89a", Target::Feature::ARMv89a)
+        .value("ARM64e", Target::Feature::ARM64e)
         .value("SanitizerCoverage", Target::Feature::SanitizerCoverage)
         .value("ProfileByTimer", Target::Feature::ProfileByTimer)
         .value("SPIRV", Target::Feature::SPIRV)

src/CodeGen_ARM.cpp

@@ -45,6 +45,11 @@ namespace {
 //
 // v8r has no relation to anything.
 Target complete_arm_target(Target t) {
+    // If arm64e is set, assume at least arm8.3a
+    if (t.has_feature(Target::ARM64e)) {
+        t.set_feature(Target::ARMv83a);
+    }
+
     constexpr int num_arm_v8_features = 10;
     static const Target::Feature arm_v8_features[num_arm_v8_features] = {
         Target::ARMv89a,
@@ -2473,7 +2478,11 @@ string CodeGen_ARM::mcpu_target() const {
         }
     } else {
         if (target.os == Target::IOS) {
-            return "apple-a7";
+            if (target.has_feature(Target::ARM64e)) {
+                return "apple-a12";
+            } else {
+                return "apple-a7";
+            }
         } else if (target.os == Target::OSX) {
             return "apple-m1";
         } else if (target.has_feature(Target::SVE2)) {

src/LLVM_Runtime_Linker.cpp

@@ -504,7 +504,11 @@ llvm::Triple get_triple_for_target(const Target &target) {
         } else {
             user_assert(target.bits == 64) << "Target bits must be 32 or 64\n";
 #ifdef WITH_AARCH64
-            triple.setArch(llvm::Triple::aarch64);
+            if (target.has_feature(Target::ARM64e)) {
+                triple.setArch(llvm::Triple::aarch64, llvm::Triple::AArch64SubArch_arm64e);
+            } else {
+                triple.setArch(llvm::Triple::aarch64);
+            }
 #else
             user_error << "AArch64 llvm target not enabled in this build of Halide\n";
 #endif

src/Target.cpp

@@ -698,6 +698,7 @@ const std::map<std::string, Target::Feature> feature_name_map = {
     {"armv87a", Target::ARMv87a},
     {"armv88a", Target::ARMv88a},
     {"armv89a", Target::ARMv89a},
+    {"arm64e", Target::ARM64e},
     {"sanitizer_coverage", Target::SanitizerCoverage},
     {"profile_by_timer", Target::ProfileByTimer},
     {"spirv", Target::SPIRV},
@@ -1504,7 +1505,7 @@ bool Target::get_runtime_compatible_target(const Target &other, Target &result)
     // (c) must match across both targets; it is an error if one target has the feature and the other doesn't
 
     // clang-format off
-    const std::array<Feature, 33> union_features = {{
+    const std::array<Feature, 34> union_features = {{
         // These are true union features.
         CUDA,
         D3D12Compute,
@@ -1545,6 +1546,8 @@ bool Target::get_runtime_compatible_target(const Target &other, Target &result)
         ARMv87a,
         ARMv88a,
         ARMv89a,
+
+        ARM64e,
     }};
     // clang-format on
 

src/Target.h

@@ -164,6 +164,7 @@ struct Target {
         ARMv87a = halide_target_feature_armv87a,
         ARMv88a = halide_target_feature_armv88a,
         ARMv89a = halide_target_feature_armv89a,
+        ARM64e = halide_target_feature_arm64e,
         SanitizerCoverage = halide_target_feature_sanitizer_coverage,
         ProfileByTimer = halide_target_feature_profile_by_timer,
         SPIRV = halide_target_feature_spirv,

src/runtime/HalideRuntime.h

@@ -1444,6 +1444,7 @@ typedef enum halide_target_feature_t {
     halide_target_feature_armv87a,                ///< Enable ARMv8.7a instructions
     halide_target_feature_armv88a,                ///< Enable ARMv8.8a instructions
     halide_target_feature_armv89a,                ///< Enable ARMv8.9a instructions
+    halide_target_feature_arm64e,                 ///< Enable ARM64e (requires ARMv8.3a)
     halide_target_feature_sanitizer_coverage,     ///< Enable hooks for SanitizerCoverage support.
     halide_target_feature_profile_by_timer,       ///< Alternative to halide_target_feature_profile using timer interrupt for systems without threads or applicartions that need to avoid them.
     halide_target_feature_spirv,                  ///< Enable SPIR-V code generation support.

test/correctness/cross_compilation.cpp

@@ -31,6 +31,7 @@ int main(int argc, char **argv) {
         "arm-64-ios-armv87a",
         "arm-64-ios-armv88a",
         "arm-64-ios-armv89a",
+        "arm-64-ios-arm64e",
         "arm-64-linux",
         "arm-64-noos-semihosting",
         "arm-64-windows",

test/correctness/simd_op_check.h

@@ -130,6 +130,7 @@ class SimdOpCheckTest {
                  Target::ARMv87a,
                  Target::ARMv88a,
                  Target::ARMv89a,
+                 Target::ARM64e,
                  Target::AVX,
                  Target::AVX2,
                  Target::AVX512,