Use own clear cache function for aarch64

.github/workflows/c-cpp.yml

@@ -2,7 +2,6 @@ name: C/C++ CI
 
 on:
   push:
-    branches: [ master ]
   pull_request:
 
 jobs:

src/jit_compiler_a64.cpp

@@ -98,9 +98,7 @@ JitCompilerA64::JitCompilerA64()
 	memset(reg_changed_offset, 0, sizeof(reg_changed_offset));
 	memcpy(code, (void*) randomx_program_aarch64, CodeSize);
 
-#ifdef __GNUC__
-	__builtin___clear_cache(reinterpret_cast<char*>(code), reinterpret_cast<char*>(code + CodeSize));
-#endif
+	randomx_clear_cache(code, code + CodeSize);
 }
 
 JitCompilerA64::~JitCompilerA64()
@@ -169,9 +167,7 @@ void JitCompilerA64::generateProgram(Program& program, ProgramConfiguration& con
 	codePos = ((uint8_t*)randomx_program_aarch64_update_spMix1) - ((uint8_t*)randomx_program_aarch64);
 	emit32(ARMV8A::EOR | 10 | (IntRegMap[config.readReg0] << 5) | (IntRegMap[config.readReg1] << 16), code, codePos);
 
-#ifdef __GNUC__
-	__builtin___clear_cache(reinterpret_cast<char*>(code + MainLoopBegin), reinterpret_cast<char*>(code + codePos));
-#endif
+	randomx_clear_cache(code + MainLoopBegin, code + codePos);
 }
 
 void JitCompilerA64::generateProgramLight(Program& program, ProgramConfiguration& config, uint32_t datasetOffset)
@@ -226,9 +222,7 @@ void JitCompilerA64::generateProgramLight(Program& program, ProgramConfiguration
 	emit32(ARMV8A::ADD_IMM_LO | 2 | (2 << 5) | (imm_lo << 10), code, codePos);
 	emit32(ARMV8A::ADD_IMM_HI | 2 | (2 << 5) | (imm_hi << 10), code, codePos);
 
-#ifdef __GNUC__
-	__builtin___clear_cache(reinterpret_cast<char*>(code + MainLoopBegin), reinterpret_cast<char*>(code + codePos));
-#endif
+	randomx_clear_cache(code + MainLoopBegin, code + codePos);
 }
 
 template<size_t N>
@@ -344,9 +338,7 @@ void JitCompilerA64::generateSuperscalarHash(SuperscalarProgram(&programs)[N], s
 	memcpy(code + codePos, p1, p2 - p1);
 	codePos += p2 - p1;
 
-#ifdef __GNUC__
-	__builtin___clear_cache(reinterpret_cast<char*>(code + CodeSize), reinterpret_cast<char*>(code + codePos));
-#endif
+	randomx_clear_cache(code + CodeSize, code + codePos);
 }
 
 template void JitCompilerA64::generateSuperscalarHash(SuperscalarProgram(&programs)[RANDOMX_CACHE_ACCESSES], std::vector<uint64_t> &reciprocalCache);

src/jit_compiler_a64_static.S

@@ -33,6 +33,7 @@
 
 	.arch armv8-a
 	.text
+	.global DECL(randomx_clear_cache)
 	.global DECL(randomx_program_aarch64)
 	.global DECL(randomx_program_aarch64_main_loop)
 	.global DECL(randomx_program_aarch64_vm_instructions)
@@ -52,6 +53,57 @@
 	.global DECL(randomx_calc_dataset_item_aarch64_store_result)
 	.global DECL(randomx_calc_dataset_item_aarch64_end)
 
+	.balign 4
+DECL(randomx_clear_cache):
+	# x0 = begin
+	# x1 = end
+
+	# Range check
+	cmp	x0, x1
+	bhs	randomx_clear_cache_exit
+
+	# Read "Cache Type Register, EL0"
+	# https://developer.arm.com/documentation/ddi0488/h/system-control/aarch64-register-descriptions/cache-type-register--el0
+	mrs	x4, ctr_el0
+
+	# [19:16] DminLine
+	# Log2 of the number of words in the smallest cache line of all the data and unified caches that the processor controls
+	lsr	x2, x4, 16
+	and	x2, x2, 15
+	mov	x3, 4
+	lsl	x3, x3, x2
+
+	# Invalidate all data cache lines between x0 and x1
+	mov	x2, x0
+randomx_dcache_invalidate_loop:
+	#dc	cvau, x2
+	add	x2, x2, x3
+	cmp	x2, x1
+	blo	randomx_clear_cache_loop
+
+	# Data Synchronization Barrier
+	dsb	ish
+
+	# [3:0] IminLine
+	# Log2 of the number of words in the smallest cache line of all the Instruction Caches that the processor controls
+	and	x2, x4, 15
+	mov	x3, 4
+	lsl	x3, x3, x2
+
+	# Invalidate all instruction cache lines between x0 and x1
+	mov	x2, x0
+randomx_clear_cache_loop:
+	#ic	ivau, x2
+	add	x2, x2, x3
+	cmp	x2, x1
+	blo	randomx_clear_cache_loop
+
+	# Instruction Synchronization Barrier
+	isb	sy
+
+randomx_clear_cache_exit:
+	ret
+
 #include "configuration.h"
 
 # Register allocation
@@ -106,7 +158,6 @@
 # v30 -> E 'or' mask  = 0x3*00000000******3*00000000******
 # v31 -> scale mask   = 0x81f000000000000081f0000000000000
 
-	.balign 4
 DECL(randomx_program_aarch64):
 	# Save callee-saved registers
 	sub	sp, sp, 192

src/jit_compiler_a64_static.hpp

@@ -30,6 +30,7 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #pragma once
 
 extern "C" {
+	void randomx_clear_cache(void* begin, void* end);
 	void randomx_program_aarch64(void* reg, void* mem, void* scratchpad, uint64_t iterations);
 	void randomx_program_aarch64_main_loop();
 	void randomx_program_aarch64_vm_instructions();