Merge remote-tracking branch 'origin/main' into abadams/zen4_natural_vector_size

abadams · abadams · commit f395da4f26da · 2025-02-13T10:08:16.000-08:00
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -196,7 +196,7 @@ option(THREADS_PREFER_PTHREAD_FLAG "When enabled, prefer to use the -pthread fla
 find_package(Threads REQUIRED)
 
 ## LLVM
-find_package(Halide_LLVM 18...20 REQUIRED
+find_package(Halide_LLVM 18...99 REQUIRED  # Use 99 to fake a minimum-only constraint
              COMPONENTS WebAssembly X86
              OPTIONAL_COMPONENTS AArch64 ARM Hexagon NVPTX PowerPC RISCV)
 
diff --git a/src/CodeGen_LLVM.cpp b/src/CodeGen_LLVM.cpp
@@ -748,7 +748,7 @@ Value *CodeGen_LLVM::register_destructor(llvm::Function *destructor_fn, Value *o
     IRBuilderBase::InsertPoint here = builder->saveIP();
     BasicBlock *dtors = get_destructor_block();
 
-    builder->SetInsertPoint(dtors->getFirstNonPHI());
+    builder->SetInsertPoint(dtors->getFirstNonPHIIt());
 
     PHINode *error_code = dyn_cast<PHINode>(dtors->begin());
     internal_assert(error_code) << "The destructor block is supposed to start with a phi node\n";
diff --git a/src/CodeGen_PTX_Dev.cpp b/src/CodeGen_PTX_Dev.cpp
@@ -156,6 +156,8 @@ void CodeGen_PTX_Dev::add_kernel(Stmt stmt,
         }
     }
 
+    function->setCallingConv(llvm::CallingConv::PTX_Kernel);
+
     // Make the initial basic block
     entry_block = BasicBlock::Create(*context, "entry", function);
     builder->SetInsertPoint(entry_block);
diff --git a/src/IROperator.cpp b/src/IROperator.cpp
@@ -496,7 +496,6 @@ Expr lossless_cast(Type t,
                 Expr a = lossless_cast(t, op->a, scope, cache);
                 Expr b = lossless_cast(t, op->b, scope, cache);
                 if (a.defined() && b.defined()) {
-                    debug(0) << a << " " << b << "\n";
                     return Min::make(a, b);
                 }
             } else if (const Max *op = e.as<Max>()) {
diff --git a/src/runtime/aarch64_cpu_features.cpp b/src/runtime/aarch64_cpu_features.cpp
@@ -21,24 +21,24 @@ extern "C" unsigned long getauxval(unsigned long type);
 
 namespace {
 
-void set_platform_features(CpuFeatures &features) {
+void set_platform_features(CpuFeatures *features) {
     unsigned long hwcaps = getauxval(AT_HWCAP);
     unsigned long hwcaps2 = getauxval(AT_HWCAP2);
 
     if (hwcaps & HWCAP_ASIMDDP) {
-        features.set_available(halide_target_feature_arm_dot_prod);
+        halide_set_available_cpu_feature(features, halide_target_feature_arm_dot_prod);
     }
 
     if (hwcaps & HWCAP_ASIMDHP) {
-        features.set_available(halide_target_feature_arm_fp16);
+        halide_set_available_cpu_feature(features, halide_target_feature_arm_fp16);
     }
 
     if (hwcaps & HWCAP_SVE) {
-        features.set_available(halide_target_feature_sve);
+        halide_set_available_cpu_feature(features, halide_target_feature_sve);
     }
 
     if (hwcaps2 & HWCAP2_SVE2) {
-        features.set_available(halide_target_feature_sve2);
+        halide_set_available_cpu_feature(features, halide_target_feature_sve2);
     }
 }
 
@@ -56,13 +56,13 @@ bool sysctl_is_set(const char *name) {
     return sysctlbyname(name, &enabled, &enabled_len, nullptr, 0) == 0 && enabled;
 }
 
-void set_platform_features(CpuFeatures &features) {
+void set_platform_features(CpuFeatures *features) {
     if (sysctl_is_set("hw.optional.arm.FEAT_DotProd")) {
-        features.set_available(halide_target_feature_arm_dot_prod);
+        halide_set_available_cpu_feature(features, halide_target_feature_arm_dot_prod);
     }
 
     if (sysctl_is_set("hw.optional.arm.FEAT_FP16")) {
-        features.set_available(halide_target_feature_arm_fp16);
+        halide_set_available_cpu_feature(features, halide_target_feature_arm_fp16);
     }
 }
 
@@ -84,20 +84,20 @@ extern "C" BOOL IsProcessorFeaturePresent(DWORD feature);
 
 namespace {
 
-void set_platform_features(CpuFeatures &features) {
+void set_platform_features(CpuFeatures *features) {
     // This is the strategy used by Google's cpuinfo library for
     // detecting fp16 arithmetic support on Windows.
     if (!IsProcessorFeaturePresent(PF_FLOATING_POINT_EMULATED) &&
         IsProcessorFeaturePresent(PF_ARM_FMAC_INSTRUCTIONS_AVAILABLE)) {
-        features.set_available(halide_target_feature_arm_fp16);
+        halide_set_available_cpu_feature(features, halide_target_feature_arm_fp16);
     }
 
     if (IsProcessorFeaturePresent(PF_ARM_V82_DP_INSTRUCTIONS_AVAILABLE)) {
-        features.set_available(halide_target_feature_arm_dot_prod);
+        halide_set_available_cpu_feature(features, halide_target_feature_arm_dot_prod);
     }
 
     if (IsProcessorFeaturePresent(PF_ARM_SVE_INSTRUCTIONS_AVAILABLE)) {
-        features.set_available(halide_target_feature_sve);
+        halide_set_available_cpu_feature(features, halide_target_feature_sve);
     }
 }
 
@@ -107,28 +107,27 @@ void set_platform_features(CpuFeatures &features) {
 
 namespace {
 
-void set_platform_features(CpuFeatures &) {
+void set_platform_features(CpuFeatures *) {
 }
 
 }  // namespace
 
 #endif
 
-WEAK CpuFeatures halide_get_cpu_features() {
-    CpuFeatures features;
-    features.set_known(halide_target_feature_arm_dot_prod);
-    features.set_known(halide_target_feature_arm_fp16);
-    features.set_known(halide_target_feature_armv7s);
-    features.set_known(halide_target_feature_no_neon);
-    features.set_known(halide_target_feature_sve);
-    features.set_known(halide_target_feature_sve2);
+extern "C" WEAK int halide_get_cpu_features(CpuFeatures *features) {
+    halide_set_known_cpu_feature(features, halide_target_feature_arm_dot_prod);
+    halide_set_known_cpu_feature(features, halide_target_feature_arm_fp16);
+    halide_set_known_cpu_feature(features, halide_target_feature_armv7s);
+    halide_set_known_cpu_feature(features, halide_target_feature_no_neon);
+    halide_set_known_cpu_feature(features, halide_target_feature_sve);
+    halide_set_known_cpu_feature(features, halide_target_feature_sve2);
 
     // All ARM architectures support "No Neon".
-    features.set_available(halide_target_feature_no_neon);
+    halide_set_available_cpu_feature(features, halide_target_feature_no_neon);
 
     set_platform_features(features);
 
-    return features;
+    return halide_error_code_success;
 }
 
 }  // namespace Internal
diff --git a/src/runtime/arm_cpu_features.cpp b/src/runtime/arm_cpu_features.cpp
@@ -18,15 +18,15 @@ extern "C" unsigned long getauxval(unsigned long type);
 
 namespace {
 
-void set_platform_features(CpuFeatures &features) {
+void set_platform_features(CpuFeatures *features) {
     unsigned long hwcaps = getauxval(AT_HWCAP);
 
     if (hwcaps & HWCAP_ASIMDDP) {
-        features.set_available(halide_target_feature_arm_dot_prod);
+        halide_set_available_cpu_feature(features, halide_target_feature_arm_dot_prod);
     }
 
     if (hwcaps & HWCAP_ASIMDHP) {
-        features.set_available(halide_target_feature_arm_fp16);
+        halide_set_available_cpu_feature(features, halide_target_feature_arm_fp16);
     }
 }
 
@@ -68,17 +68,17 @@ bool is_armv7s() {
     return type == CPU_TYPE_ARM && subtype == CPU_SUBTYPE_ARM_V7S;
 }
 
-void set_platform_features(CpuFeatures &features) {
+void set_platform_features(CpuFeatures *features) {
     if (is_armv7s()) {
-        features.set_available(halide_target_feature_armv7s);
+        halide_set_available_cpu_feature(features, halide_target_feature_armv7s);
     }
 
     if (sysctl_is_set("hw.optional.arm.FEAT_DotProd")) {
-        features.set_available(halide_target_feature_arm_dot_prod);
+        halide_set_available_cpu_feature(features, halide_target_feature_arm_dot_prod);
     }
 
     if (sysctl_is_set("hw.optional.arm.FEAT_FP16")) {
-        features.set_available(halide_target_feature_arm_fp16);
+        halide_set_available_cpu_feature(features, halide_target_feature_arm_fp16);
     }
 }
 
@@ -88,28 +88,27 @@ void set_platform_features(CpuFeatures &features) {
 
 namespace {
 
-void set_platform_features(CpuFeatures &) {
+void set_platform_features(CpuFeatures *) {
 }
 
 }  // namespace
 
 #endif
 
-WEAK CpuFeatures halide_get_cpu_features() {
-    CpuFeatures features;
-    features.set_known(halide_target_feature_arm_dot_prod);
-    features.set_known(halide_target_feature_arm_fp16);
-    features.set_known(halide_target_feature_armv7s);
-    features.set_known(halide_target_feature_no_neon);
-    features.set_known(halide_target_feature_sve);
-    features.set_known(halide_target_feature_sve2);
+extern "C" WEAK int halide_get_cpu_features(CpuFeatures *features) {
+    halide_set_known_cpu_feature(features, halide_target_feature_arm_dot_prod);
+    halide_set_known_cpu_feature(features, halide_target_feature_arm_fp16);
+    halide_set_known_cpu_feature(features, halide_target_feature_armv7s);
+    halide_set_known_cpu_feature(features, halide_target_feature_no_neon);
+    halide_set_known_cpu_feature(features, halide_target_feature_sve);
+    halide_set_known_cpu_feature(features, halide_target_feature_sve2);
 
     // All ARM architectures support "No Neon".
-    features.set_available(halide_target_feature_no_neon);
+    halide_set_available_cpu_feature(features, halide_target_feature_no_neon);
 
     set_platform_features(features);
 
-    return features;
+    return halide_error_code_success;
 }
 
 }  // namespace Internal
diff --git a/src/runtime/can_use_target.cpp b/src/runtime/can_use_target.cpp
@@ -40,21 +40,23 @@ WEAK int halide_default_can_use_target_features(int count, const uint64_t *featu
 
         static_assert(sizeof(halide_cpu_features_storage) == sizeof(CpuFeatures), "CpuFeatures Mismatch");
         if (!halide_cpu_features_initialized) {
-            CpuFeatures tmp = halide_get_cpu_features();
+            CpuFeatures tmp;
+            int error = halide_get_cpu_features(&tmp);
+            halide_abort_if_false(nullptr, error == halide_error_code_success);
             memcpy(&halide_cpu_features_storage, &tmp, sizeof(tmp));
             halide_cpu_features_initialized = true;
         }
     }
 
-    if (count != CpuFeatures::kWordCount) {
+    if (count != cpu_feature_mask_size) {
         // This should not happen unless our runtime is out of sync with the rest of libHalide.
 #ifdef DEBUG_RUNTIME
-        debug(nullptr) << "count " << count << " CpuFeatures::kWordCount " << CpuFeatures::kWordCount << "\n";
+        debug(nullptr) << "count " << count << " cpu_feature_mask_size " << cpu_feature_mask_size << "\n";
 #endif
         halide_error(nullptr, "Internal error: wrong structure size passed to halide_can_use_target_features()\n");
     }
     const CpuFeatures *cpu_features = reinterpret_cast<const CpuFeatures *>(&halide_cpu_features_storage[0]);
-    for (int i = 0; i < CpuFeatures::kWordCount; ++i) {
+    for (int i = 0; i < cpu_feature_mask_size; ++i) {
         uint64_t m;
         if ((m = (features[i] & cpu_features->known[i])) != 0) {
             if ((m & cpu_features->available[i]) != m) {
diff --git a/src/runtime/cpu_features.h b/src/runtime/cpu_features.h
@@ -8,42 +8,37 @@ namespace Halide {
 namespace Runtime {
 namespace Internal {
 
-// Return two masks:
+// Size of CPU feature mask large enough to cover all Halide target features
+static constexpr int cpu_feature_mask_size = (halide_target_feature_end + 63) / (sizeof(uint64_t) * 8);
+
+// Contains two masks:
 // One with all the CPU-specific features that might possible be available on this architecture ('known'),
 // and one with the subset that are actually present ('available').
 struct CpuFeatures {
-    static const int kWordCount = (halide_target_feature_end + 63) / (sizeof(uint64_t) * 8);
-
-    ALWAYS_INLINE void set_known(int i) {
-        known[i >> 6] |= ((uint64_t)1) << (i & 63);
-    }
-
-    ALWAYS_INLINE void set_available(int i) {
-        available[i >> 6] |= ((uint64_t)1) << (i & 63);
-    }
-
-    ALWAYS_INLINE bool test_known(int i) const {
-        return (known[i >> 6] & ((uint64_t)1) << (i & 63)) != 0;
-    }
-
-    ALWAYS_INLINE bool test_available(int i) const {
-        return (available[i >> 6] & ((uint64_t)1) << (i & 63)) != 0;
-    }
-
-    ALWAYS_INLINE
-    CpuFeatures() {
-        for (int i = 0; i < kWordCount; ++i) {
-            known[i] = 0;
-            available[i] = 0;
-        }
-    }
-
-    uint64_t known[kWordCount];      // mask of the CPU features we know how to detect
-    uint64_t available[kWordCount];  // mask of the CPU features that are available
-                                     // (always a subset of 'known')
+    uint64_t known[cpu_feature_mask_size] = {0};      // mask of the CPU features we know how to detect
+    uint64_t available[cpu_feature_mask_size] = {0};  // mask of the CPU features that are available
+                                                      // (always a subset of 'known')
 };
 
-extern WEAK CpuFeatures halide_get_cpu_features();
+ALWAYS_INLINE void halide_set_known_cpu_feature(CpuFeatures *features, int i) {
+    features->known[i >> 6] |= ((uint64_t)1) << (i & 63);
+}
+
+ALWAYS_INLINE void halide_set_available_cpu_feature(CpuFeatures *features, int i) {
+    features->available[i >> 6] |= ((uint64_t)1) << (i & 63);
+}
+
+ALWAYS_INLINE bool halide_test_known_cpu_feature(CpuFeatures *features, int i) {
+    return (features->known[i >> 6] & ((uint64_t)1) << (i & 63)) != 0;
+}
+
+ALWAYS_INLINE bool halide_test_available_cpu_feature(CpuFeatures *features, int i) {
+    return (features->available[i >> 6] & ((uint64_t)1) << (i & 63)) != 0;
+}
+
+// NOTE: This method is not part of the public API, but we push it into extern "C" to
+//       avoid name mangling mismatches between platforms. See: https://github.com/halide/Halide/issues/8565
+extern "C" WEAK int halide_get_cpu_features(CpuFeatures *features);
 
 }  // namespace Internal
 }  // namespace Runtime
diff --git a/src/runtime/hexagon_cpu_features.cpp b/src/runtime/hexagon_cpu_features.cpp
@@ -5,9 +5,9 @@ namespace Halide {
 namespace Runtime {
 namespace Internal {
 
-WEAK CpuFeatures halide_get_cpu_features() {
+extern "C" WEAK int halide_get_cpu_features(CpuFeatures *features) {
     // Hexagon has no CPU-specific Features.
-    return CpuFeatures();
+    return halide_error_code_success;
 }
 
 }  // namespace Internal
diff --git a/src/runtime/powerpc_cpu_features.cpp b/src/runtime/powerpc_cpu_features.cpp
@@ -8,27 +8,29 @@
 
 #define PPC_FEATURE2_ARCH_2_07 0x80000000
 
-extern "C" unsigned long int getauxval(unsigned long int);
+extern "C" {
+
+unsigned long int getauxval(unsigned long int);
+}
 
 namespace Halide {
 namespace Runtime {
 namespace Internal {
 
-WEAK CpuFeatures halide_get_cpu_features() {
-    CpuFeatures features;
-    features.set_known(halide_target_feature_vsx);
-    features.set_known(halide_target_feature_power_arch_2_07);
+extern "C" WEAK int halide_get_cpu_features(CpuFeatures *features) {
+    halide_set_known_cpu_feature(features, halide_target_feature_vsx);
+    halide_set_known_cpu_feature(features, halide_target_feature_power_arch_2_07);
 
     const unsigned long hwcap = getauxval(AT_HWCAP);
     const unsigned long hwcap2 = getauxval(AT_HWCAP2);
 
     if (hwcap & PPC_FEATURE_HAS_VSX) {
-        features.set_available(halide_target_feature_vsx);
+        halide_set_available_cpu_feature(features, halide_target_feature_vsx);
     }
     if (hwcap2 & PPC_FEATURE2_ARCH_2_07) {
-        features.set_available(halide_target_feature_power_arch_2_07);
+        halide_set_available_cpu_feature(features, halide_target_feature_power_arch_2_07);
     }
-    return features;
+    return halide_error_code_success;
 }
 
 }  // namespace Internal
diff --git a/src/runtime/riscv_cpu_features.cpp b/src/runtime/riscv_cpu_features.cpp
@@ -5,9 +5,9 @@ namespace Halide {
 namespace Runtime {
 namespace Internal {
 
-WEAK CpuFeatures halide_get_cpu_features() {
+extern "C" WEAK int halide_get_cpu_features(Halide::Runtime::Internal::CpuFeatures *features) {
     // For now, no version specific features, though RISCV promises to have many.
-    return CpuFeatures();
+    return halide_error_code_success;
 }
 
 }  // namespace Internal
diff --git a/src/runtime/wasm_cpu_features.cpp b/src/runtime/wasm_cpu_features.cpp
diff --git a/src/runtime/x86_cpu_features.cpp b/src/runtime/x86_cpu_features.cpp
diff --git a/vcpkg.json b/vcpkg.json

Original file line number	Diff line number	Diff line change
`@@ -156,6 +156,8 @@ void CodeGen_PTX_Dev::add_kernel(Stmt stmt,`
`156`	`156`	`}`
`157`	`157`	`}`
`158`	`158`
	`159`	`+ function->setCallingConv(llvm::CallingConv::PTX_Kernel);`
	`160`	`+`
`159`	`161`	`// Make the initial basic block`
`160`	`162`	`entry_block = BasicBlock::Create(*context, "entry", function);`
`161`	`163`	`builder->SetInsertPoint(entry_block);`
Original file line number	Diff line number	Diff line change
`@@ -496,7 +496,6 @@ Expr lossless_cast(Type t,`
`496`	`496`	`Expr a = lossless_cast(t, op->a, scope, cache);`
`497`	`497`	`Expr b = lossless_cast(t, op->b, scope, cache);`
`498`	`498`	`if (a.defined() && b.defined()) {`
`499`		`- debug(0) << a << " " << b << "\n";`
`500`	`499`	`return Min::make(a, b);`
`501`	`500`	`}`
`502`	`501`	`} else if (const Max *op = e.as<Max>()) {`