More robust ISA detection on ARM/Linux

7 files changed, 72 insertions, 40 deletions

diff --git a/CMakeLists.txt b/CMakeLists.txt
index f750a4c..ed829f6 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -84,12 +84,12 @@ ELSEIF(CMAKE_SYSTEM_PROCESSOR MATCHES "^(armv5te|armv7|armv7f|armv7s|armv7k|armv
       src/arm/linux/chipset.c
       src/arm/linux/midr.c)
     IF(CMAKE_SYSTEM_PROCESSOR MATCHES "^(armv5te|armv7l|armv7-a)$")
-      LIST(APPEND CPUINFO_SRCS src/arm/linux/arm32-isa.c)
+      LIST(APPEND CPUINFO_SRCS src/arm/linux/aarch32-isa.c)
       IF(CMAKE_SYSTEM_NAME STREQUAL "Android" AND ANDROID_ABI STREQUAL "armeabi")
-        SET_SOURCE_FILES_PROPERTIES(src/arm/linux/arm32-isa.c PROPERTIES COMPILE_FLAGS -marm)
+        SET_SOURCE_FILES_PROPERTIES(src/arm/linux/aarch32-isa.c PROPERTIES COMPILE_FLAGS -marm)
       ENDIF()
     ELSEIF(CMAKE_SYSTEM_PROCESSOR STREQUAL "aarch64")
-      LIST(APPEND CPUINFO_SRCS src/arm/linux/arm64-isa.c)
+      LIST(APPEND CPUINFO_SRCS src/arm/linux/aarch64-isa.c)
     ENDIF()
   ELSEIF(CMAKE_SYSTEM_NAME STREQUAL "Darwin")
     LIST(APPEND CPUINFO_SRCS
diff --git a/configure.py b/configure.py
index 271c454..0d8777a 100755
--- a/configure.py
+++ b/configure.py
@@ -46,9 +46,9 @@ def main(args):
                     "arm/linux/chipset.c",
                 ]
                 if build.target.is_arm:
-                    sources.append("arm/linux/arm32-isa.c")
+                    sources.append("arm/linux/aarch32-isa.c")
                 elif build.target.is_arm64:
-                    sources.append("arm/linux/arm64-isa.c")
+                    sources.append("arm/linux/aarch64-isa.c")
 
         if build.target.is_macos:
             sources += ["mach/topology.c"]
diff --git a/jni/Android.mk b/jni/Android.mk
index e8c5621..f63d29c 100644
--- a/jni/Android.mk
+++ b/jni/Android.mk
@@ -22,13 +22,13 @@ LOCAL_SRC_FILES += \
 	$(LOCAL_PATH)/src/arm/linux/midr.c \
 	$(LOCAL_PATH)/src/arm/android/properties.c
 ifeq ($(TARGET_ARCH_ABI),armeabi)
-LOCAL_SRC_FILES += $(LOCAL_PATH)/src/arm/linux/arm32-isa.c.arm
+LOCAL_SRC_FILES += $(LOCAL_PATH)/src/arm/linux/aarch32-isa.c.arm
 endif # armeabi
 ifeq ($(TARGET_ARCH_ABI),armeabi-v7a)
-LOCAL_SRC_FILES += $(LOCAL_PATH)/src/arm/linux/arm32-isa.c
+LOCAL_SRC_FILES += $(LOCAL_PATH)/src/arm/linux/aarch32-isa.c
 endif # armeabi-v7a
 ifeq ($(TARGET_ARCH_ABI),arm64-v8a)
-LOCAL_SRC_FILES += $(LOCAL_PATH)/src/arm/linux/arm64-isa.c
+LOCAL_SRC_FILES += $(LOCAL_PATH)/src/arm/linux/aarch64-isa.c
 endif # arm64-v8a
 endif # armeabi, armeabi-v7a, or arm64-v8a
 ifeq ($(TARGET_ARCH_ABI),$(filter $(TARGET_ARCH_ABI),x86 x86_64))
@@ -74,13 +74,13 @@ LOCAL_SRC_FILES += \
 	$(LOCAL_PATH)/src/arm/linux/midr.c \
 	$(LOCAL_PATH)/src/arm/android/properties.c
 ifeq ($(TARGET_ARCH_ABI),armeabi)
-LOCAL_SRC_FILES += $(LOCAL_PATH)/src/arm/linux/arm32-isa.c.arm
+LOCAL_SRC_FILES += $(LOCAL_PATH)/src/arm/linux/aarch32-isa.c.arm
 endif # armeabi
 ifeq ($(TARGET_ARCH_ABI),armeabi-v7a)
-LOCAL_SRC_FILES += $(LOCAL_PATH)/src/arm/linux/arm32-isa.c
+LOCAL_SRC_FILES += $(LOCAL_PATH)/src/arm/linux/aarch32-isa.c
 endif # armeabi-v7a
 ifeq ($(TARGET_ARCH_ABI),arm64-v8a)
-LOCAL_SRC_FILES += $(LOCAL_PATH)/src/arm/linux/arm64-isa.c
+LOCAL_SRC_FILES += $(LOCAL_PATH)/src/arm/linux/aarch64-isa.c
 endif # arm64-v8a
 endif # armeabi, armeabi-v7a, or arm64-v8a
 ifeq ($(TARGET_ARCH_ABI),$(filter $(TARGET_ARCH_ABI),x86 x86_64))
diff --git a/src/arm/linux/arm32-isa.c b/src/arm/linux/aarch32-isa.c
index 092d50f..2d36182 100644
--- a/src/arm/linux/arm32-isa.c
+++ b/src/arm/linux/aarch32-isa.c
@@ -25,11 +25,13 @@
 
 
 void cpuinfo_arm_linux_decode_isa_from_proc_cpuinfo(
-	const struct cpuinfo_arm_linux_processor processor[restrict static 1],
+	uint32_t features,
+	uint32_t features2,
+	uint32_t midr,
+	uint32_t architecture_version,
+	uint32_t architecture_flags,
 	struct cpuinfo_arm_isa isa[restrict static 1])
 {
-	const uint32_t midr = processor->midr;
-	uint32_t architecture_version = processor->architecture_version;
 	if (architecture_version >= 8) {
 		/*
 		 * ARMv7 code running on ARMv8: IDIV, VFP, NEON are always supported,
@@ -60,8 +62,13 @@ void cpuinfo_arm_linux_decode_isa_from_proc_cpuinfo(
 			architecture_version = 6;
 		}
 
-		const uint32_t features = processor->features;
-		const uint32_t architecture_flags = processor->architecture_flags;
+		if (architecture_version < 7) {
+			const uint32_t armv7_features_mask = CPUINFO_ARM_LINUX_FEATURE_VFPV3 | CPUINFO_ARM_LINUX_FEATURE_VFPV3D16 | CPUINFO_ARM_LINUX_FEATURE_VFPD32 |
+				CPUINFO_ARM_LINUX_FEATURE_VFPV4 | CPUINFO_ARM_LINUX_FEATURE_NEON | CPUINFO_ARM_LINUX_FEATURE_IDIVT | CPUINFO_ARM_LINUX_FEATURE_IDIVA;
+			if (features & armv7_features_mask) {
+				architecture_version = 7;
+			}
+		}
 		if ((architecture_version >= 6) || (features & CPUINFO_ARM_LINUX_FEATURE_EDSP) || (architecture_flags & CPUINFO_ARM_LINUX_ARCH_E)) {
 			isa->armv5e = true;
 		}
@@ -175,7 +182,6 @@ void cpuinfo_arm_linux_decode_isa_from_proc_cpuinfo(
 		}
 	}
 
-	const uint32_t features2 = processor->features2;
 	if (features2 & CPUINFO_ARM_LINUX_FEATURE2_AES) {
 		isa->aes = true;
 	}
diff --git a/src/arm/linux/arm64-isa.c b/src/arm/linux/aarch64-isa.c
index ae92d48..0b67a12 100644
--- a/src/arm/linux/arm64-isa.c
+++ b/src/arm/linux/aarch64-isa.c
@@ -5,10 +5,9 @@
 
 
 void cpuinfo_arm64_linux_decode_isa_from_proc_cpuinfo(
-	const struct cpuinfo_arm_linux_processor processor[restrict static 1],
+	uint32_t features,
 	struct cpuinfo_arm_isa isa[restrict static 1])
 {
-	const uint32_t features = processor->features;
 	if (features & CPUINFO_ARM_LINUX_FEATURE_AES) {
 		isa->aes = true;
 	}
diff --git a/src/arm/linux/api.h b/src/arm/linux/api.h
index 1c31ba1..6764032 100644
--- a/src/arm/linux/api.h
+++ b/src/arm/linux/api.h
@@ -247,11 +247,15 @@ bool cpuinfo_arm_linux_parse_proc_cpuinfo(
 
 #if CPUINFO_ARCH_ARM
 	void cpuinfo_arm_linux_decode_isa_from_proc_cpuinfo(
-		const struct cpuinfo_arm_linux_processor processors[restrict static 1],
+		uint32_t features,
+		uint32_t features2,
+		uint32_t midr,
+		uint32_t architecture_version,
+		uint32_t architecture_flags,
 		struct cpuinfo_arm_isa isa[restrict static 1]);
 #elif CPUINFO_ARCH_ARM64
 	void cpuinfo_arm64_linux_decode_isa_from_proc_cpuinfo(
-		const struct cpuinfo_arm_linux_processor processors[restrict static 1],
+		uint32_t features,
 		struct cpuinfo_arm_isa isa[restrict static 1]);
 #endif
 
diff --git a/src/arm/linux/init.c b/src/arm/linux/init.c
index 6a7ab5d..d7ad54a 100644
--- a/src/arm/linux/init.c
+++ b/src/arm/linux/init.c
@@ -169,7 +169,9 @@ void cpuinfo_arm_linux_init(void) {
 	}
 
 	uint32_t usable_processors = 0;
-	uint32_t known_processors = 0;
+	#if CPUINFO_ARCH_ARM
+	uint32_t last_midr = 0, last_architecture_version = 0, last_architecture_flags = 0;
+	#endif
 	for (uint32_t i = 0; i < arm_linux_processors_count; i++) {
 		arm_linux_processors[i].system_processor_id = i;
 		if (bitmask_all(arm_linux_processors[i].flags, CPUINFO_LINUX_MASK_USABLE)) {
@@ -183,25 +185,15 @@ void cpuinfo_arm_linux_init(void) {
 				cpuinfo_log_info("processor %"PRIu32" is not listed in /proc/cpuinfo", i);
 			}
 
-			if (bitmask_all(arm_linux_processors[i].flags, CPUINFO_ARM_LINUX_VALID_INFO)) {
-				if (known_processors == 0) {
-					/*
-					 * This is the first processor for which we have complete information.
-					 * Use it to detect instruction set architecture.
-					 * If there are several cores with different microarchitecture, we expect
-					 * Linux kernel to report the same ISA extensions for each of them.
-					 */
-					#if CPUINFO_ARCH_ARM
-						cpuinfo_arm_linux_decode_isa_from_proc_cpuinfo(
-							&arm_linux_processors[i], &cpuinfo_isa);
-					#elif CPUINFO_ARCH_ARM64
-						cpuinfo_arm64_linux_decode_isa_from_proc_cpuinfo(
-							&arm_linux_processors[i], &cpuinfo_isa);
-					#endif
+			#if CPUINFO_ARCH_ARM
+				if (bitmask_all(arm_linux_processors[i].flags, CPUINFO_ARM_LINUX_VALID_MIDR)) {
+					last_midr = arm_linux_processors[i].midr;
 				}
-
-				known_processors += 1;
-			}
+				if (bitmask_all(arm_linux_processors[i].flags, CPUINFO_ARM_LINUX_VALID_ARCHITECTURE)) {
+					last_architecture_version = arm_linux_processors[i].architecture_version;
+					last_architecture_flags   = arm_linux_processors[i].architecture_flags;
+				}
+			#endif
 		} else {
 			/* Processor reported in /proc/cpuinfo, but not in possible and/or present lists: log and ignore */
 			if (!(arm_linux_processors[i].flags & CPUINFO_ARM_LINUX_VALID_PROCESSOR)) {
@@ -210,6 +202,37 @@ void cpuinfo_arm_linux_init(void) {
 		}
 	}
 
+	/* Detect ISA. If different processors report different ISA features, take the intersection. */
+	uint32_t isa_features = 0, processors_with_features = 0;
+	#if CPUINFO_ARCH_ARM
+		uint32_t isa_features2 = 0;
+	#endif
+	for (uint32_t i = 0; i < arm_linux_processors_count; i++) {
+		if (bitmask_all(arm_linux_processors[i].flags, CPUINFO_LINUX_MASK_USABLE | CPUINFO_ARM_LINUX_VALID_FEATURES)) {
+			if (processors_with_features == 0) {
+				isa_features = arm_linux_processors[i].features;
+				#if CPUINFO_ARCH_ARM
+					isa_features2 = arm_linux_processors[i].features2;
+				#endif
+			} else {
+				isa_features &= arm_linux_processors[i].features;
+				#if CPUINFO_ARCH_ARM
+					isa_features2 &= arm_linux_processors[i].features2;
+				#endif
+			}
+			processors_with_features += 1;
+		}
+	}
+	#if CPUINFO_ARCH_ARM
+		cpuinfo_arm_linux_decode_isa_from_proc_cpuinfo(
+			isa_features, isa_features2,
+			last_midr, last_architecture_version, last_architecture_flags,
+			&cpuinfo_isa);
+	#elif CPUINFO_ARCH_ARM64
+		cpuinfo_arm64_linux_decode_isa_from_proc_cpuinfo(
+			isa_features, &cpuinfo_isa);
+	#endif
+
 	/* Detect min/max frequency and package ID */
 	for (uint32_t i = 0; i < arm_linux_processors_count; i++) {
 		if (bitmask_all(arm_linux_processors[i].flags, CPUINFO_LINUX_MASK_USABLE)) {