[PATCH 5/5] -march=native: MOVBE support

From: Alexey Dobriyan <adobriyan@gmail.com>
To: linux-kernel@vger.kernel.org
Cc: x86@kernel.org, tglx@linutronix.de, mingo@redhat.com,
	hpa@zytor.com, Alexey Dobriyan <adobriyan@gmail.com>
Subject: [PATCH 5/5] -march=native: MOVBE support
Date: Fri,  8 Dec 2017 01:41:54 +0300	[thread overview]
Message-ID: <20171207224154.4687-5-adobriyan@gmail.com> (raw)
In-Reply-To: <20171207224154.4687-1-adobriyan@gmail.com>

Use MOVBE if it is available.

This doesn't save code size as MOVBE seems to be as long as MOV+BSWAP,
It is not clear if it saves uop, maybe it will in the future.

Do it because it is easy, I guess.
---
 arch/x86/crypto/des3_ede-asm_64.S | 28 ++++++++++++++++++++++++++++
 arch/x86/net/bpf_jit.S            | 12 ++++++++++++
 scripts/kconfig/cpuid.c           |  4 ++++
 scripts/march-native.sh           |  3 ++-
 4 files changed, 46 insertions(+), 1 deletion(-)

diff --git a/arch/x86/crypto/des3_ede-asm_64.S b/arch/x86/crypto/des3_ede-asm_64.S
index 8e49ce117494..007319ea1f62 100644
--- a/arch/x86/crypto/des3_ede-asm_64.S
+++ b/arch/x86/crypto/des3_ede-asm_64.S
@@ -159,6 +159,15 @@
 
 #define dummy2(a, b) /*_*/
 
+#ifdef CONFIG_MARCH_NATIVE_MOVBE
+#define read_block(io, left, right) \
+	movbe	 (io), left##d; \
+	movbe	4(io), right##d;
+
+#define write_block(io, left, right) \
+	movbe	left##d,   (io); \
+	movbe	right##d, 4(io);
+#else
 #define read_block(io, left, right) \
 	movl    (io), left##d; \
 	movl   4(io), right##d; \
@@ -170,6 +179,7 @@
 	bswapl right##d; \
 	movl   left##d,   (io); \
 	movl   right##d, 4(io);
+#endif
 
 ENTRY(des3_ede_x86_64_crypt_blk)
 	/* input:
@@ -443,6 +453,14 @@ ENTRY(des3_ede_x86_64_crypt_blk_3way)
 	pushq %rsi /* dst */
 
 	/* load input */
+#ifdef CONFIG_MARCH_NATIVE_MOVBE
+	movbe 0 * 4(%rdx), RL0d;
+	movbe 1 * 4(%rdx), RR0d;
+	movbe 2 * 4(%rdx), RL1d;
+	movbe 3 * 4(%rdx), RR1d;
+	movbe 4 * 4(%rdx), RL2d;
+	movbe 5 * 4(%rdx), RR2d;
+#else
 	movl 0 * 4(%rdx), RL0d;
 	movl 1 * 4(%rdx), RR0d;
 	movl 2 * 4(%rdx), RL1d;
@@ -456,6 +474,7 @@ ENTRY(des3_ede_x86_64_crypt_blk_3way)
 	bswapl RR1d;
 	bswapl RL2d;
 	bswapl RR2d;
+#endif
 
 	initial_permutation3(RL, RR);
 
@@ -516,6 +535,14 @@ ENTRY(des3_ede_x86_64_crypt_blk_3way)
 
 	final_permutation3(RR, RL);
 
+#ifdef CONFIG_MARCH_NATIVE_MOVBE
+	movbe RR0d, 0 * 4(%rsi);
+	movbe RL0d, 1 * 4(%rsi);
+	movbe RR1d, 2 * 4(%rsi);
+	movbe RL1d, 3 * 4(%rsi);
+	movbe RR2d, 4 * 4(%rsi);
+	movbe RL2d, 5 * 4(%rsi);
+#else
 	bswapl RR0d;
 	bswapl RL0d;
 	bswapl RR1d;
@@ -530,6 +557,7 @@ ENTRY(des3_ede_x86_64_crypt_blk_3way)
 	movl RL1d, 3 * 4(%rsi);
 	movl RR2d, 4 * 4(%rsi);
 	movl RL2d, 5 * 4(%rsi);
+#endif
 
 	popq %r15;
 	popq %r14;
diff --git a/arch/x86/net/bpf_jit.S b/arch/x86/net/bpf_jit.S
index b33093f84528..17fe33750298 100644
--- a/arch/x86/net/bpf_jit.S
+++ b/arch/x86/net/bpf_jit.S
@@ -34,8 +34,12 @@ FUNC(sk_load_word_positive_offset)
 	sub	%esi,%eax		# hlen - offset
 	cmp	$3,%eax
 	jle	bpf_slow_path_word
+#ifdef CONFIG_MARCH_NATIVE_MOVBE
+	movbe	(SKBDATA,%rsi),%eax
+#else
 	mov     (SKBDATA,%rsi),%eax
 	bswap   %eax  			/* ntohl() */
+#endif
 	ret
 
 FUNC(sk_load_half)
@@ -80,8 +84,12 @@ FUNC(sk_load_byte_positive_offset)
 bpf_slow_path_word:
 	bpf_slow_path_common(4)
 	js	bpf_error
+#ifdef CONFIG_MARCH_NATIVE_MOVBE
+	movbe	32(%rbp),%eax
+#else
 	mov	32(%rbp),%eax
 	bswap	%eax
+#endif
 	ret
 
 bpf_slow_path_half:
@@ -118,8 +126,12 @@ bpf_slow_path_word_neg:
 
 FUNC(sk_load_word_negative_offset)
 	sk_negative_common(4)
+#ifdef CONFIG_MARCH_NATIVE_MOVBE
+	movbe	(%rax), %eax
+#else
 	mov	(%rax), %eax
 	bswap	%eax
+#endif
 	ret
 
 bpf_slow_path_half_neg:
diff --git a/scripts/kconfig/cpuid.c b/scripts/kconfig/cpuid.c
index ecb285183581..2c23c8699ae6 100644
--- a/scripts/kconfig/cpuid.c
+++ b/scripts/kconfig/cpuid.c
@@ -42,6 +42,7 @@ static inline void cpuid2(uint32_t eax0, uint32_t ecx0, uint32_t *eax, uint32_t
 	);
 }
 
+static bool movbe	= false;
 static bool popcnt	= false;
 static bool rep_movsb	= false;
 static bool rep_stosb	= false;
@@ -56,6 +57,8 @@ static void intel(void)
 		cpuid(1, &eax, &ecx, &edx, &ebx);
 //		printf("%08x %08x %08x %08x\n", eax, ecx, edx, ebx);
 
+		if (ecx & (1 << 22))
+			movbe = true;
 		if (ecx & (1 << 23))
 			popcnt = true;
 	}
@@ -86,6 +89,7 @@ int main(int argc, char *argv[])
 		intel();
 
 #define _(x)	if (streq(opt, #x)) return x ? EXIT_SUCCESS : EXIT_FAILURE
+	_(movbe);
 	_(popcnt);
 	_(rep_movsb);
 	_(rep_stosb);
diff --git a/scripts/march-native.sh b/scripts/march-native.sh
index d3adf0edb2be..93f6a9bd4a6c 100755
--- a/scripts/march-native.sh
+++ b/scripts/march-native.sh
@@ -30,6 +30,7 @@ option() {
 }
 
 if test -x "$CPUID"; then
+	"$CPUID" movbe		&& option "CONFIG_MARCH_NATIVE_MOVBE"
 	"$CPUID" popcnt		&& option "CONFIG_MARCH_NATIVE_POPCNT"
 	"$CPUID" rep_movsb	&& option "CONFIG_MARCH_NATIVE_REP_MOVSB"
 	"$CPUID" rep_stosb	&& option "CONFIG_MARCH_NATIVE_REP_STOSB"
@@ -76,7 +77,7 @@ for i in $COLLECT_GCC_OPTIONS; do
 		-mhle)		option "CONFIG_MARCH_NATIVE_HLE"	;;
 		-mlzcnt)	option "CONFIG_MARCH_NATIVE_LZCNT"	;;
 		-mmmx)		option "CONFIG_MARCH_NATIVE_MMX"	;;
-		-mmovbe)	option "CONFIG_MARCH_NATIVE_MOVBE"	;;
+		-mmovbe);;
 		-mpclmul)	option "CONFIG_MARCH_NATIVE_PCLMUL"	;;
 		-mpopcnt);;
 		-mprfchw)	option "CONFIG_MARCH_NATIVE_PREFETCHW"	;;
-- 
2.13.6
