crc32: Add crc32 implementation using __builtin_aarch64_crc32b

ARMv8.0 has optional crc32 instruction for crc32 calculation. The
instruction is mandatory since ARMv8.1. The crc32 calculation is
faster using the dedicated instruction, e.g. 1.4 GHz iMX8MN gives:

  => time crc32 0x50000000 0x2000000
  time: 0.126 seconds # crc32 instruction
  time: 0.213 seconds # software crc32

Add implementation using the compiler builtin wrapper for the crc32
instruction and enable it by default, since we don't support any
platforms which do not implement this instruction.

Signed-off-by: Marek Vasut <[email protected]>
Cc: Simon Glass <[email protected]>
[trini: Make crc32_table guarded by CONFIG_ARM64_CRC32]
Signed-off-by: Tom Rini <[email protected]>

diff --git a/arch/arm/Kconfig b/arch/arm/Kconfig
index f0fd57f8d6d675471c3b50c2f40a634f7029879e..95102d386b4ca04e4cd8a11ba8d268bd939e3b8e 100644 (file)
--- a/arch/arm/Kconfig
+++ b/arch/arm/Kconfig
@@ -9,6 +9,16 @@ config ARM64
        select PHYS_64BIT
        select SYS_CACHE_SHIFT_6
 
+config ARM64_CRC32
+       bool "Enable support for CRC32 instruction"
+       depends on ARM64
+       default y
+       help
+         ARMv8 implements dedicated crc32 instruction for crc32 calculation.
+         This is faster than software crc32 calculation. This instruction may
+         not be present on all ARMv8.0, but is always present on ARMv8.1 and
+         newer.
+
 config POSITION_INDEPENDENT
        bool "Generate position-independent pre-relocation code"
        depends on ARM64 || CPU_V7A

diff --git a/arch/arm/Makefile b/arch/arm/Makefile
index c68e598a675b96647abd36f0a2a07bbb17868df6..ce977bf6323f1b8fe4f993101a725118029ebe8c 100644 (file)
--- a/arch/arm/Makefile
+++ b/arch/arm/Makefile
@@ -18,7 +18,11 @@ arch-$(CONFIG_CPU_V7A)               =$(call cc-option, -march=armv7-a, \
                                 $(call cc-option, -march=armv7))
 arch-$(CONFIG_CPU_V7M)         =-march=armv7-m
 arch-$(CONFIG_CPU_V7R)         =-march=armv7-r
+ifeq ($(CONFIG_ARM64_CRC32),y)
+arch-$(CONFIG_ARM64)           =-march=armv8-a+crc
+else
 arch-$(CONFIG_ARM64)           =-march=armv8-a
+endif
 
 # On Tegra systems we must build SPL for the armv4 core on the device
 # but otherwise we can use the value in CONFIG_SYS_ARM_ARCH

diff --git a/lib/crc32.c b/lib/crc32.c
index f2acc107fe4c102b63c612724545355308d606fd..5a3127e03adad2cb77449696452d11d283df5563 100644 (file)
--- a/lib/crc32.c
+++ b/lib/crc32.c
@@ -84,7 +84,7 @@ static void __efi_runtime make_crc_table(void)
   }
   crc_table_empty = 0;
 }
-#else
+#elif !defined(CONFIG_ARM64_CRC32)
 /* ========================================================================
  * Table of CRC-32's of all single-byte values (made by make_crc_table)
  */
@@ -184,6 +184,12 @@ const uint32_t * ZEXPORT get_crc_table()
  */
 uint32_t __efi_runtime crc32_no_comp(uint32_t crc, const Bytef *buf, uInt len)
 {
+#ifdef CONFIG_ARM64_CRC32
+    crc = cpu_to_le32(crc);
+    while (len--)
+        crc = __builtin_aarch64_crc32b(crc, *buf++);
+    return le32_to_cpu(crc);
+#else
     const uint32_t *tab = crc_table;
     const uint32_t *b =(const uint32_t *)buf;
     size_t rem_len;
@@ -221,6 +227,7 @@ uint32_t __efi_runtime crc32_no_comp(uint32_t crc, const Bytef *buf, uInt len)
     }
 
     return le32_to_cpu(crc);
+#endif
 }
 #undef DO_CRC