[v1.05][ARM] Preeliminary support for SVE detection (#259)

Makefile

@@ -38,10 +38,16 @@ ifneq ($(OS),Windows_NT)
 		CFLAGS += -DARCH_PPC -std=gnu99 -fstack-protector-all -Wno-language-extension-token
 	else ifeq ($(arch), $(filter $(arch), arm aarch64_be aarch64 arm64 armv8b armv8l armv7l armv6l))
 		SRC_DIR=src/arm/
-		SOURCE += $(COMMON_SRC) $(SRC_DIR)midr.c $(SRC_DIR)uarch.c $(SRC_COMMON)soc.c $(SRC_DIR)soc.c $(SRC_COMMON)pci.c $(SRC_DIR)udev.c
+		SOURCE += $(COMMON_SRC) $(SRC_DIR)midr.c $(SRC_DIR)uarch.c $(SRC_COMMON)soc.c $(SRC_DIR)soc.c $(SRC_COMMON)pci.c $(SRC_DIR)udev.c sve.o
 		HEADERS += $(COMMON_HDR) $(SRC_DIR)midr.h $(SRC_DIR)uarch.h  $(SRC_COMMON)soc.h $(SRC_DIR)soc.h $(SRC_COMMON)pci.h $(SRC_DIR)udev.c $(SRC_DIR)socs.h
 		CFLAGS += -DARCH_ARM -Wno-unused-parameter -std=c99 -fstack-protector-all
 
+		# Check if the compiler supports -march=armv8-a+sve. We will use it (if supported) to compile SVE detection code later
+		is_sve_flag_supported := $(shell $(CC) -march=armv8-a+sve -c $(SRC_DIR)sve.c -o sve_test.o 2> /dev/null && echo 'yes'; rm -f sve_test.o)
+		ifeq ($(is_sve_flag_supported), yes)
+			SVE_FLAGS += -march=armv8-a+sve
+		endif
+
 		ifeq ($(os), Darwin)
 			SOURCE += $(SRC_COMMON)sysctl.c
 			HEADERS += $(SRC_COMMON)sysctl.h
@@ -91,6 +97,9 @@ freq_avx.o: Makefile $(SRC_DIR)freq/freq_avx.c $(SRC_DIR)freq/freq_avx.h $(SRC_D
 freq_avx512.o: Makefile $(SRC_DIR)freq/freq_avx512.c $(SRC_DIR)freq/freq_avx512.h $(SRC_DIR)freq/freq.h
 	$(CC) $(CFLAGS) $(SANITY_FLAGS) -c -mavx512f -pthread $(SRC_DIR)freq/freq_avx512.c -o $@
 
+sve.o: Makefile $(SRC_DIR)sve.c $(SRC_DIR)sve.h
+	$(CC) $(CFLAGS) $(SANITY_FLAGS) $(SVE_FLAGS) -c $(SRC_DIR)sve.c -o $@
+
 $(OUTPUT): Makefile $(SOURCE) $(HEADERS)
 ifeq ($(GIT_VERSION),"")
 	$(CC) $(CFLAGS) $(SANITY_FLAGS) $(SOURCE) -o $(OUTPUT)

src/arm/midr.c

@@ -19,6 +19,7 @@
 #include "udev.h"
 #include "midr.h"
 #include "uarch.h"
+#include "sve.h"
 
 bool cores_are_equal(int c1pos, int c2pos, uint32_t* midr_array, int32_t* freq_array) {
   return midr_array[c1pos] == midr_array[c2pos] && freq_array[c1pos] == freq_array[c2pos];
@@ -168,6 +169,15 @@ struct features* get_features_info(void) {
     feat->SHA1 = hwcaps & HWCAP_SHA1;
     feat->SHA2 = hwcaps & HWCAP_SHA2;
     feat->NEON = hwcaps & HWCAP_ASIMD;
+    feat->SVE = hwcaps & HWCAP_SVE;
+
+    hwcaps = getauxval(AT_HWCAP2);
+    if (errno == ENOENT) {
+      printWarn("Unable to retrieve AT_HWCAP2 using getauxval");
+    }
+    else {
+      feat->SVE2 = hwcaps & HWCAP2_SVE2;
+    }
   }
 #else
   else {
@@ -183,6 +193,8 @@ struct features* get_features_info(void) {
     feat->CRC32 = hwcaps & HWCAP2_CRC32;
     feat->SHA1 = hwcaps & HWCAP2_SHA1;
     feat->SHA2 = hwcaps & HWCAP2_SHA2;
+    feat->SVE = false;
+    feat->SVE2 = false;
   }
 #endif // ifdef __aarch64__
 #elif defined __APPLE__ || __MACH__
@@ -192,8 +204,14 @@ struct features* get_features_info(void) {
   feat->SHA1 = true;
   feat->SHA2 = true;
   feat->NEON = true;
+  feat->SVE = false;
+  feat->SVE2 = false;
 #endif  // ifdef __linux__
 
+  if (feat->SVE || feat->SVE2) {
+    feat->cntb = sve_cntb();
+  }
+
   return feat;
 }
 
@@ -430,7 +448,7 @@ char* get_str_topology(struct cpuInfo* cpu, struct topology* topo, bool dual_soc
 
 char* get_str_features(struct cpuInfo* cpu) {
   struct features* feat = cpu->feat;
-  uint32_t max_len = strlen("NEON,SHA1,SHA2,AES,CRC32,") + 1;
+  uint32_t max_len = strlen("NEON,SHA1,SHA2,AES,CRC32,SVE,SVE2") + 1;
   uint32_t len = 0;
   char* string = ecalloc(max_len, sizeof(char));
 
@@ -438,6 +456,14 @@ char* get_str_features(struct cpuInfo* cpu) {
     strcat(string, "NEON,");
     len += 5;
   }
+  if(feat->SVE) {
+    strcat(string, "SVE,");
+    len += 4;
+  }
+  if(feat->SVE2) {
+    strcat(string, "SVE2,");
+    len += 5;
+  }
   if(feat->SHA1) {
     strcat(string, "SHA1,");
     len += 5;

src/arm/sve.c

@@ -0,0 +1,15 @@
+#include <stdint.h>
+#include "../common/global.h"
+
+// https://learn.arm.com/learning-paths/servers-and-cloud-computing/sve/sve_basics/#:~:text=Using%20a%20text%20editor%20of%20your%20choice%2C%20copy,svcntb%28%29%29%3B%20%7D%20This%20program%20prints%20the%20vector%20length
+uint64_t sve_cntb(void) {
+  #ifdef __ARM_FEATURE_SVE
+    uint64_t x0 = 0;
+    __asm volatile("cntb %0"
+         : "=r"(x0));
+    return x0;
+  #else
+    printWarn("sve_cntb: Hardware supports SVE, but it was not enabled by the compiler");
+    return 0;
+  #endif
+}

src/arm/sve.h

@@ -0,0 +1,6 @@
+#ifndef __SVE_DETECTION__
+#define __SVE_DETECTION__
+
+uint64_t sve_cntb(void);
+
+#endif

src/arm/uarch.c

@@ -294,7 +294,10 @@ int get_vpus_width(struct cpuInfo* cpu) {
     case UARCH_NEOVERSE_V1:
       return 256;
     default:
-      if(cpu->feat->NEON) {
+      if (cpu->feat->SVE && cpu->feat->cntb > 0) {
+        return cpu->feat->cntb * 8;
+      }
+      else if (cpu->feat->NEON) {
         if(is_ARMv8_or_newer(cpu)) {
           return 128;
         }

src/common/cpu.h

@@ -124,6 +124,9 @@ struct features {
   bool SHA1;
   bool SHA2;
   bool CRC32;
+  bool SVE;
+  bool SVE2;
+  uint32_t cntb;
 #endif  
 };