Skip to contentSkip to conversation

xor: enable auto-vectorization in Clang

PATCH In review Ard Biesheuvel 2 participants 4 messages Feb 5, 2022 – Feb 11, 2022

Subsystems (10)
Arm (32-bit) x86 PowerPC Crypto API +6 more subsystems
lore
Share & tools
lore
#0 lore raw

v4 cover xor: enable auto-vectorization in Clang

Update the xor_blocks() prototypes so that the compiler understands that the inputs always refer to distinct regions of memory. This is implied by the existing implementations, as they use different granularities for the load/xor/store loops.

With that, we can fix the ARM/Clang version, which refuses to SIMD vectorize otherwise, and throws a spurious warning related to the GCC version being incompatible.

Changes since v3:
- revert broken PPC argument rename - doing it fully results in too much pointless churn, and the 'inner' altivec routines are not strictly part of the xor_blocks API anyway

Changes since v2:
- fix arm64 build after rebase
- name PPC argument names consistently
- add Nick's acks and link tags

Changes since v1:
- fix PPC build
- add Nathan's Tested-by

Cc Nick Desaulniers <[email protected]>
Cc Nathan Chancellor <[email protected]>

Ard Biesheuvel (2):
lib/xor: make xor prototypes more friendly to compiler vectorization
crypto: arm/xor - make vectorized C code Clang-friendly

 arch/alpha/include/asm/xor.h           | 53 ++++++++----
 arch/arm/include/asm/xor.h             | 42 ++++++----
 arch/arm/lib/xor-neon.c                | 12 +--
 arch/arm64/include/asm/xor.h           | 21 +++--
 arch/arm64/lib/xor-neon.c              | 46 +++++++----
 arch/ia64/include/asm/xor.h            | 21 +++--
 arch/powerpc/include/asm/xor_altivec.h | 25 +++---
 arch/powerpc/lib/xor_vmx.c             | 28 ++++---
 arch/powerpc/lib/xor_vmx.h             | 27 ++++---
 arch/powerpc/lib/xor_vmx_glue.c        | 32 ++++----
 arch/s390/lib/xor.c                    | 21 +++--
 arch/sparc/include/asm/xor_32.h        | 21 +++--
 arch/sparc/include/asm/xor_64.h        | 42 ++++++----
 arch/x86/include/asm/xor.h             | 42 ++++++----
 arch/x86/include/asm/xor_32.h          | 42 ++++++----
 arch/x86/include/asm/xor_avx.h         | 21 +++--
 include/asm-generic/xor.h              | 84 +++++++++++++-------
 include/linux/raid/xor.h               | 21 +++--
 18 files changed, 384 insertions(+), 217 deletions(-)
Signature · 1 line
2.30.2
#1 lore raw

PATCH v4 1/2 lib/xor: make xor prototypes more friendly to compiler vectorization

Modern compilers are perfectly capable of extracting parallelism from the XOR routines, provided that the prototypes reflect the nature of the input accurately, in particular, the fact that the input vectors are expected not to overlap. This is not documented explicitly, but is implied by the interchangeability of the various C routines, some of which use temporary variables while others don't: this means that these routines only behave identically for non-overlapping inputs.

So let's decorate these input vectors with the __restrict modifier, which informs the compiler that there is no overlap. While at it, make the input-only vectors pointer-to-const as well.

---
 arch/alpha/include/asm/xor.h           | 53 ++++++++----
 arch/arm/include/asm/xor.h             | 42 ++++++----
 arch/arm64/include/asm/xor.h           | 21 +++--
 arch/arm64/lib/xor-neon.c              | 46 +++++++----
 arch/ia64/include/asm/xor.h            | 21 +++--
 arch/powerpc/include/asm/xor_altivec.h | 25 +++---
 arch/powerpc/lib/xor_vmx.c             | 28 ++++---
 arch/powerpc/lib/xor_vmx.h             | 27 ++++---
 arch/powerpc/lib/xor_vmx_glue.c        | 32 ++++----
 arch/s390/lib/xor.c                    | 21 +++--
 arch/sparc/include/asm/xor_32.h        | 21 +++--
 arch/sparc/include/asm/xor_64.h        | 42 ++++++----
 arch/x86/include/asm/xor.h             | 42 ++++++----
 arch/x86/include/asm/xor_32.h          | 42 ++++++----
 arch/x86/include/asm/xor_avx.h         | 21 +++--
 include/asm-generic/xor.h              | 84 +++++++++++++-------
 include/linux/raid/xor.h               | 21 +++--
 17 files changed, 381 insertions(+), 208 deletions(-)
arch/alpha/include/asm/xor.h +36 -17
index 5aeb4fb3cb7c..e0de0c233ab9 100644
@@ -5,24 +5,43 @@
  * Optimized RAID-5 checksumming functions for alpha EV5 and EV6
  */
 
-extern void xor_alpha_2(unsigned long, unsigned long *, unsigned long *);
-extern void xor_alpha_3(unsigned long, unsigned long *, unsigned long *,
-		        unsigned long *);
-extern void xor_alpha_4(unsigned long, unsigned long *, unsigned long *,
-		        unsigned long *, unsigned long *);
-extern void xor_alpha_5(unsigned long, unsigned long *, unsigned long *,
-		        unsigned long *, unsigned long *, unsigned long *);
+extern void
+xor_alpha_2(unsigned long bytes, unsigned long * __restrict p1,
+	    const unsigned long * __restrict p2);
+extern void
+xor_alpha_3(unsigned long bytes, unsigned long * __restrict p1,
+	    const unsigned long * __restrict p2,
+	    const unsigned long * __restrict p3);
+extern void
+xor_alpha_4(unsigned long bytes, unsigned long * __restrict p1,
+	    const unsigned long * __restrict p2,
+	    const unsigned long * __restrict p3,
+	    const unsigned long * __restrict p4);
+extern void
+xor_alpha_5(unsigned long bytes, unsigned long * __restrict p1,
+	    const unsigned long * __restrict p2,
+	    const unsigned long * __restrict p3,
+	    const unsigned long * __restrict p4,
+	    const unsigned long * __restrict p5);
 
-extern void xor_alpha_prefetch_2(unsigned long, unsigned long *,
-				 unsigned long *);
-extern void xor_alpha_prefetch_3(unsigned long, unsigned long *,
-				 unsigned long *, unsigned long *);
-extern void xor_alpha_prefetch_4(unsigned long, unsigned long *,
-				 unsigned long *, unsigned long *,
-				 unsigned long *);
-extern void xor_alpha_prefetch_5(unsigned long, unsigned long *,
-				 unsigned long *, unsigned long *,
-				 unsigned long *, unsigned long *);
+extern void
+xor_alpha_prefetch_2(unsigned long bytes, unsigned long * __restrict p1,
+		     const unsigned long * __restrict p2);
+extern void
+xor_alpha_prefetch_3(unsigned long bytes, unsigned long * __restrict p1,
+		     const unsigned long * __restrict p2,
+		     const unsigned long * __restrict p3);
+extern void
+xor_alpha_prefetch_4(unsigned long bytes, unsigned long * __restrict p1,
+		     const unsigned long * __restrict p2,
+		     const unsigned long * __restrict p3,
+		     const unsigned long * __restrict p4);
+extern void
+xor_alpha_prefetch_5(unsigned long bytes, unsigned long * __restrict p1,
+		     const unsigned long * __restrict p2,
+		     const unsigned long * __restrict p3,
+		     const unsigned long * __restrict p4,
+		     const unsigned long * __restrict p5);
 
 asm("								\n\
 	.text							\n\
arch/arm/include/asm/xor.h +28 -14

90 lines not on this page · raw message

arch/arm64/include/asm/xor.h +14 -7
index 947f6a4f1aa0..befcd8a7abc9 100644
@@ -16,7 +16,8 @@
 extern struct xor_block_template const xor_block_inner_neon;
 
 static void
-xor_neon_2(unsigned long bytes, unsigned long *p1, unsigned long *p2)
+xor_neon_2(unsigned long bytes, unsigned long * __restrict p1,
+	   const unsigned long * __restrict p2)
 {
 	kernel_neon_begin();
 	xor_block_inner_neon.do_2(bytes, p1, p2);
@@ -24,8 +25,9 @@ xor_neon_2(unsigned long bytes, unsigned long *p1, unsigned long *p2)
 }
 
 static void
-xor_neon_3(unsigned long bytes, unsigned long *p1, unsigned long *p2,
-		unsigned long *p3)
+xor_neon_3(unsigned long bytes, unsigned long * __restrict p1,
+	   const unsigned long * __restrict p2,
+	   const unsigned long * __restrict p3)
 {
 	kernel_neon_begin();
 	xor_block_inner_neon.do_3(bytes, p1, p2, p3);
@@ -33,8 +35,10 @@ xor_neon_3(unsigned long bytes, unsigned long *p1, unsigned long *p2,
 }
 
 static void
-xor_neon_4(unsigned long bytes, unsigned long *p1, unsigned long *p2,
-		unsigned long *p3, unsigned long *p4)
+xor_neon_4(unsigned long bytes, unsigned long * __restrict p1,
+	   const unsigned long * __restrict p2,
+	   const unsigned long * __restrict p3,
+	   const unsigned long * __restrict p4)
 {
 	kernel_neon_begin();
 	xor_block_inner_neon.do_4(bytes, p1, p2, p3, p4);
@@ -42,8 +46,11 @@ xor_neon_4(unsigned long bytes, unsigned long *p1, unsigned long *p2,
 }
 
 static void
-xor_neon_5(unsigned long bytes, unsigned long *p1, unsigned long *p2,
-		unsigned long *p3, unsigned long *p4, unsigned long *p5)
+xor_neon_5(unsigned long bytes, unsigned long * __restrict p1,
+	   const unsigned long * __restrict p2,
+	   const unsigned long * __restrict p3,
+	   const unsigned long * __restrict p4,
+	   const unsigned long * __restrict p5)
 {
 	kernel_neon_begin();
 	xor_block_inner_neon.do_5(bytes, p1, p2, p3, p4, p5);
arch/arm64/lib/xor-neon.c +29 -17

88 lines not on this page · raw message

arch/ia64/include/asm/xor.h +14 -7
index 673051bf9d7d..6785f70d3208 100644
@@ -4,13 +4,20 @@
  */
 
 
-extern void xor_ia64_2(unsigned long, unsigned long *, unsigned long *);
-extern void xor_ia64_3(unsigned long, unsigned long *, unsigned long *,
-		       unsigned long *);
-extern void xor_ia64_4(unsigned long, unsigned long *, unsigned long *,
-		       unsigned long *, unsigned long *);
-extern void xor_ia64_5(unsigned long, unsigned long *, unsigned long *,
-		       unsigned long *, unsigned long *, unsigned long *);
+extern void xor_ia64_2(unsigned long bytes, unsigned long * __restrict p1,
+		       const unsigned long * __restrict p2);
+extern void xor_ia64_3(unsigned long bytes, unsigned long * __restrict p1,
+		       const unsigned long * __restrict p2,
+		       const unsigned long * __restrict p3);
+extern void xor_ia64_4(unsigned long bytes, unsigned long * __restrict p1,
+		       const unsigned long * __restrict p2,
+		       const unsigned long * __restrict p3,
+		       const unsigned long * __restrict p4);
+extern void xor_ia64_5(unsigned long bytes, unsigned long * __restrict p1,
+		       const unsigned long * __restrict p2,
+		       const unsigned long * __restrict p3,
+		       const unsigned long * __restrict p4,
+		       const unsigned long * __restrict p5);
 
 static struct xor_block_template xor_block_ia64 = {
 	.name =	"ia64",
arch/powerpc/include/asm/xor_altivec.h +14 -11
index 6ca923510b59..294620a25f80 100644
@@ -3,17 +3,20 @@
 #define _ASM_POWERPC_XOR_ALTIVEC_H
 
 #ifdef CONFIG_ALTIVEC
-
-void xor_altivec_2(unsigned long bytes, unsigned long *v1_in,
-		   unsigned long *v2_in);
-void xor_altivec_3(unsigned long bytes, unsigned long *v1_in,
-		   unsigned long *v2_in, unsigned long *v3_in);
-void xor_altivec_4(unsigned long bytes, unsigned long *v1_in,
-		   unsigned long *v2_in, unsigned long *v3_in,
-		   unsigned long *v4_in);
-void xor_altivec_5(unsigned long bytes, unsigned long *v1_in,
-		   unsigned long *v2_in, unsigned long *v3_in,
-		   unsigned long *v4_in, unsigned long *v5_in);
+void xor_altivec_2(unsigned long bytes, unsigned long * __restrict p1,
+		   const unsigned long * __restrict p2);
+void xor_altivec_3(unsigned long bytes, unsigned long * __restrict p1,
+		   const unsigned long * __restrict p2,
+		   const unsigned long * __restrict p3);
+void xor_altivec_4(unsigned long bytes, unsigned long * __restrict p1,
+		   const unsigned long * __restrict p2,
+		   const unsigned long * __restrict p3,
+		   const unsigned long * __restrict p4);
+void xor_altivec_5(unsigned long bytes, unsigned long * __restrict p1,
+		   const unsigned long * __restrict p2,
+		   const unsigned long * __restrict p3,
+		   const unsigned long * __restrict p4,
+		   const unsigned long * __restrict p5);
 
 #endif
 #endif /* _ASM_POWERPC_XOR_ALTIVEC_H */
arch/powerpc/lib/xor_vmx.c +18 -10
index 54e61979e80e..aab49d056d18 100644
@@ -49,8 +49,9 @@ typedef vector signed char unative_t;
 		V1##_3 = vec_xor(V1##_3, V2##_3);	\
 	} while (0)
 
-void __xor_altivec_2(unsigned long bytes, unsigned long *v1_in,
-		     unsigned long *v2_in)
+void __xor_altivec_2(unsigned long bytes,
+		     unsigned long * __restrict v1_in,
+		     const unsigned long * __restrict v2_in)
 {
 	DEFINE(v1);
 	DEFINE(v2);
@@ -67,8 +68,10 @@ void __xor_altivec_2(unsigned long bytes, unsigned long *v1_in,
 	} while (--lines > 0);
 }
 
-void __xor_altivec_3(unsigned long bytes, unsigned long *v1_in,
-		     unsigned long *v2_in, unsigned long *v3_in)
+void __xor_altivec_3(unsigned long bytes,
+		     unsigned long * __restrict v1_in,
+		     const unsigned long * __restrict v2_in,
+		     const unsigned long * __restrict v3_in)
 {
 	DEFINE(v1);
 	DEFINE(v2);
@@ -89,9 +92,11 @@ void __xor_altivec_3(unsigned long bytes, unsigned long *v1_in,
 	} while (--lines > 0);
 }
 
-void __xor_altivec_4(unsigned long bytes, unsigned long *v1_in,
-		     unsigned long *v2_in, unsigned long *v3_in,
-		     unsigned long *v4_in)
+void __xor_altivec_4(unsigned long bytes,
+		     unsigned long * __restrict v1_in,
+		     const unsigned long * __restrict v2_in,
+		     const unsigned long * __restrict v3_in,
+		     const unsigned long * __restrict v4_in)
 {
 	DEFINE(v1);
 	DEFINE(v2);
@@ -116,9 +121,12 @@ void __xor_altivec_4(unsigned long bytes, unsigned long *v1_in,
 	} while (--lines > 0);
 }
 
-void __xor_altivec_5(unsigned long bytes, unsigned long *v1_in,
-		     unsigned long *v2_in, unsigned long *v3_in,
-		     unsigned long *v4_in, unsigned long *v5_in)
+void __xor_altivec_5(unsigned long bytes,
+		     unsigned long * __restrict v1_in,
+		     const unsigned long * __restrict v2_in,
+		     const unsigned long * __restrict v3_in,
+		     const unsigned long * __restrict v4_in,
+		     const unsigned long * __restrict v5_in)
 {
 	DEFINE(v1);
 	DEFINE(v2);
arch/powerpc/lib/xor_vmx.h +14 -13
index 5c2b0839b179..573c41d90dac 100644
@@ -6,16 +6,17 @@
  * outside of the enable/disable altivec block.
  */
 
-void __xor_altivec_2(unsigned long bytes, unsigned long *v1_in,
-			     unsigned long *v2_in);
-
-void __xor_altivec_3(unsigned long bytes, unsigned long *v1_in,
-			     unsigned long *v2_in, unsigned long *v3_in);
-
-void __xor_altivec_4(unsigned long bytes, unsigned long *v1_in,
-			     unsigned long *v2_in, unsigned long *v3_in,
-			     unsigned long *v4_in);
-
-void __xor_altivec_5(unsigned long bytes, unsigned long *v1_in,
-			     unsigned long *v2_in, unsigned long *v3_in,
-			     unsigned long *v4_in, unsigned long *v5_in);
+void __xor_altivec_2(unsigned long bytes, unsigned long * __restrict p1,
+		     const unsigned long * __restrict p2);
+void __xor_altivec_3(unsigned long bytes, unsigned long * __restrict p1,
+		     const unsigned long * __restrict p2,
+		     const unsigned long * __restrict p3);
+void __xor_altivec_4(unsigned long bytes, unsigned long * __restrict p1,
+		     const unsigned long * __restrict p2,
+		     const unsigned long * __restrict p3,
+		     const unsigned long * __restrict p4);
+void __xor_altivec_5(unsigned long bytes, unsigned long * __restrict p1,
+		     const unsigned long * __restrict p2,
+		     const unsigned long * __restrict p3,
+		     const unsigned long * __restrict p4,
+		     const unsigned long * __restrict p5);
arch/powerpc/lib/xor_vmx_glue.c +18 -14
index 80dba916c367..35d917ece4d1 100644
@@ -12,47 +12,51 @@
 #include <asm/xor_altivec.h>
 #include "xor_vmx.h"
 
-void xor_altivec_2(unsigned long bytes, unsigned long *v1_in,
-		   unsigned long *v2_in)
+void xor_altivec_2(unsigned long bytes, unsigned long * __restrict p1,
+		   const unsigned long * __restrict p2)
 {
 	preempt_disable();
 	enable_kernel_altivec();
-	__xor_altivec_2(bytes, v1_in, v2_in);
+	__xor_altivec_2(bytes, p1, p2);
 	disable_kernel_altivec();
 	preempt_enable();
 }
 EXPORT_SYMBOL(xor_altivec_2);
 
-void xor_altivec_3(unsigned long bytes,  unsigned long *v1_in,
-		   unsigned long *v2_in, unsigned long *v3_in)
+void xor_altivec_3(unsigned long bytes, unsigned long * __restrict p1,
+		   const unsigned long * __restrict p2,
+		   const unsigned long * __restrict p3)
 {
 	preempt_disable();
 	enable_kernel_altivec();
-	__xor_altivec_3(bytes, v1_in, v2_in, v3_in);
+	__xor_altivec_3(bytes, p1, p2, p3);
 	disable_kernel_altivec();
 	preempt_enable();
 }
 EXPORT_SYMBOL(xor_altivec_3);
 
-void xor_altivec_4(unsigned long bytes,  unsigned long *v1_in,
-		   unsigned long *v2_in, unsigned long *v3_in,
-		   unsigned long *v4_in)
+void xor_altivec_4(unsigned long bytes, unsigned long * __restrict p1,
+		   const unsigned long * __restrict p2,
+		   const unsigned long * __restrict p3,
+		   const unsigned long * __restrict p4)
 {
 	preempt_disable();
 	enable_kernel_altivec();
-	__xor_altivec_4(bytes, v1_in, v2_in, v3_in, v4_in);
+	__xor_altivec_4(bytes, p1, p2, p3, p4);
 	disable_kernel_altivec();
 	preempt_enable();
 }
 EXPORT_SYMBOL(xor_altivec_4);
 
-void xor_altivec_5(unsigned long bytes,  unsigned long *v1_in,
-		   unsigned long *v2_in, unsigned long *v3_in,
-		   unsigned long *v4_in, unsigned long *v5_in)
+void xor_altivec_5(unsigned long bytes, unsigned long * __restrict p1,
+		   const unsigned long * __restrict p2,
+		   const unsigned long * __restrict p3,
+		   const unsigned long * __restrict p4,
+		   const unsigned long * __restrict p5)
 {
 	preempt_disable();
 	enable_kernel_altivec();
-	__xor_altivec_5(bytes, v1_in, v2_in, v3_in, v4_in, v5_in);
+	__xor_altivec_5(bytes, p1, p2, p3, p4, p5);
 	disable_kernel_altivec();
 	preempt_enable();
 }
arch/s390/lib/xor.c +14 -7
index a963c3d8ad0d..fb924a8041dc 100644
@@ -11,7 +11,8 @@
 #include <linux/raid/xor.h>
 #include <asm/xor.h>
 
-static void xor_xc_2(unsigned long bytes, unsigned long *p1, unsigned long *p2)
+static void xor_xc_2(unsigned long bytes, unsigned long * __restrict p1,
+		     const unsigned long * __restrict p2)
 {
 	asm volatile(
 		"	larl	1,2f\n"
@@ -32,8 +33,9 @@ static void xor_xc_2(unsigned long bytes, unsigned long *p1, unsigned long *p2)
 		: "0", "1", "cc", "memory");
 }
 
-static void xor_xc_3(unsigned long bytes, unsigned long *p1, unsigned long *p2,
-		     unsigned long *p3)
+static void xor_xc_3(unsigned long bytes, unsigned long * __restrict p1,
+		     const unsigned long * __restrict p2,
+		     const unsigned long * __restrict p3)
 {
 	asm volatile(
 		"	larl	1,2f\n"
@@ -58,8 +60,10 @@ static void xor_xc_3(unsigned long bytes, unsigned long *p1, unsigned long *p2,
 		: : "0", "1", "cc", "memory");
 }
 
-static void xor_xc_4(unsigned long bytes, unsigned long *p1, unsigned long *p2,
-		     unsigned long *p3, unsigned long *p4)
+static void xor_xc_4(unsigned long bytes, unsigned long * __restrict p1,
+		     const unsigned long * __restrict p2,
+		     const unsigned long * __restrict p3,
+		     const unsigned long * __restrict p4)
 {
 	asm volatile(
 		"	larl	1,2f\n"
@@ -88,8 +92,11 @@ static void xor_xc_4(unsigned long bytes, unsigned long *p1, unsigned long *p2,
 		: : "0", "1", "cc", "memory");
 }
 
-static void xor_xc_5(unsigned long bytes, unsigned long *p1, unsigned long *p2,
-		     unsigned long *p3, unsigned long *p4, unsigned long *p5)
+static void xor_xc_5(unsigned long bytes, unsigned long * __restrict p1,
+		     const unsigned long * __restrict p2,
+		     const unsigned long * __restrict p3,
+		     const unsigned long * __restrict p4,
+		     const unsigned long * __restrict p5)
 {
 	asm volatile(
 		"	larl	1,2f\n"
arch/sparc/include/asm/xor_32.h +14 -7
index 3e5af37e4b9c..0351813cf3af 100644
@@ -13,7 +13,8 @@
  */
 
 static void
-sparc_2(unsigned long bytes, unsigned long *p1, unsigned long *p2)
+sparc_2(unsigned long bytes, unsigned long * __restrict p1,
+	const unsigned long * __restrict p2)
 {
 	int lines = bytes / (sizeof (long)) / 8;
 
@@ -50,8 +51,9 @@ sparc_2(unsigned long bytes, unsigned long *p1, unsigned long *p2)
 }
 
 static void
-sparc_3(unsigned long bytes, unsigned long *p1, unsigned long *p2,
-	unsigned long *p3)
+sparc_3(unsigned long bytes, unsigned long * __restrict p1,
+	const unsigned long * __restrict p2,
+	const unsigned long * __restrict p3)
 {
 	int lines = bytes / (sizeof (long)) / 8;
 
@@ -101,8 +103,10 @@ sparc_3(unsigned long bytes, unsigned long *p1, unsigned long *p2,
 }
 
 static void
-sparc_4(unsigned long bytes, unsigned long *p1, unsigned long *p2,
-	unsigned long *p3, unsigned long *p4)
+sparc_4(unsigned long bytes, unsigned long * __restrict p1,
+	const unsigned long * __restrict p2,
+	const unsigned long * __restrict p3,
+	const unsigned long * __restrict p4)
 {
 	int lines = bytes / (sizeof (long)) / 8;
 
@@ -165,8 +169,11 @@ sparc_4(unsigned long bytes, unsigned long *p1, unsigned long *p2,
 }
 
 static void
-sparc_5(unsigned long bytes, unsigned long *p1, unsigned long *p2,
-	unsigned long *p3, unsigned long *p4, unsigned long *p5)
+sparc_5(unsigned long bytes, unsigned long * __restrict p1,
+	const unsigned long * __restrict p2,
+	const unsigned long * __restrict p3,
+	const unsigned long * __restrict p4,
+	const unsigned long * __restrict p5)
 {
 	int lines = bytes / (sizeof (long)) / 8;
 
arch/sparc/include/asm/xor_64.h +28 -14

54 lines not on this page · raw message

arch/x86/include/asm/xor.h +28 -14

90 lines not on this page · raw message

arch/x86/include/asm/xor_32.h +28 -14

90 lines not on this page · raw message

arch/x86/include/asm/xor_avx.h +14 -7
index 0c4e5b5e3852..7f81dd5897f4 100644
@@ -26,7 +26,8 @@
 		BLOCK4(8) \
 		BLOCK4(12)
 
-static void xor_avx_2(unsigned long bytes, unsigned long *p0, unsigned long *p1)
+static void xor_avx_2(unsigned long bytes, unsigned long * __restrict p0,
+		      const unsigned long * __restrict p1)
 {
 	unsigned long lines = bytes >> 9;
 
@@ -52,8 +53,9 @@ do { \
 	kernel_fpu_end();
 }
 
-static void xor_avx_3(unsigned long bytes, unsigned long *p0, unsigned long *p1,
-	unsigned long *p2)
+static void xor_avx_3(unsigned long bytes, unsigned long * __restrict p0,
+		      const unsigned long * __restrict p1,
+		      const unsigned long * __restrict p2)
 {
 	unsigned long lines = bytes >> 9;
 
@@ -82,8 +84,10 @@ do { \
 	kernel_fpu_end();
 }
 
-static void xor_avx_4(unsigned long bytes, unsigned long *p0, unsigned long *p1,
-	unsigned long *p2, unsigned long *p3)
+static void xor_avx_4(unsigned long bytes, unsigned long * __restrict p0,
+		      const unsigned long * __restrict p1,
+		      const unsigned long * __restrict p2,
+		      const unsigned long * __restrict p3)
 {
 	unsigned long lines = bytes >> 9;
 
@@ -115,8 +119,11 @@ do { \
 	kernel_fpu_end();
 }
 
-static void xor_avx_5(unsigned long bytes, unsigned long *p0, unsigned long *p1,
-	unsigned long *p2, unsigned long *p3, unsigned long *p4)
+static void xor_avx_5(unsigned long bytes, unsigned long * __restrict p0,
+	     const unsigned long * __restrict p1,
+	     const unsigned long * __restrict p2,
+	     const unsigned long * __restrict p3,
+	     const unsigned long * __restrict p4)
 {
 	unsigned long lines = bytes >> 9;
 
include/asm-generic/xor.h +56 -28

180 lines not on this page · raw message

include/linux/raid/xor.h +14 -7
index 2a9fee8ddae3..51b811b62322 100644
@@ -11,13 +11,20 @@ struct xor_block_template {
         struct xor_block_template *next;
         const char *name;
         int speed;
-	void (*do_2)(unsigned long, unsigned long *, unsigned long *);
-	void (*do_3)(unsigned long, unsigned long *, unsigned long *,
-		     unsigned long *);
-	void (*do_4)(unsigned long, unsigned long *, unsigned long *,
-		     unsigned long *, unsigned long *);
-	void (*do_5)(unsigned long, unsigned long *, unsigned long *,
-		     unsigned long *, unsigned long *, unsigned long *);
+	void (*do_2)(unsigned long, unsigned long * __restrict,
+		     const unsigned long * __restrict);
+	void (*do_3)(unsigned long, unsigned long * __restrict,
+		     const unsigned long * __restrict,
+		     const unsigned long * __restrict);
+	void (*do_4)(unsigned long, unsigned long * __restrict,
+		     const unsigned long * __restrict,
+		     const unsigned long * __restrict,
+		     const unsigned long * __restrict);
+	void (*do_5)(unsigned long, unsigned long * __restrict,
+		     const unsigned long * __restrict,
+		     const unsigned long * __restrict,
+		     const unsigned long * __restrict,
+		     const unsigned long * __restrict);
 };
 
 #endif
Signature · 1 line
2.30.2
#2 lore raw

PATCH v4 2/2 crypto: arm/xor - make vectorized C code Clang-friendly

In reply to Ard Biesheuvel

The ARM version of the accelerated XOR routines are simply the 8-way C routines passed through the auto-vectorizer with SIMD codegen enabled. This used to require GCC version 4.6 at least, but given that 5.1 is now the baseline, this check is no longer necessary, and actually misidentifies Clang as GCC < 4.6 as Clang defines the GCC major/minor as well, but makes no attempt at doing this in a way that conveys feature parity with a certain version of GCC (which would not be a great idea in the first place).

So let's drop the version check, and make the auto-vectorize pragma (which is based on a GCC-specific command line option) GCC-only. Since Clang performs SIMD auto-vectorization by default at -O2, no pragma is necessary here.

---
 arch/arm/lib/xor-neon.c | 12 +++---------
 1 file changed, 3 insertions(+), 9 deletions(-)
arch/arm/lib/xor-neon.c +3 -9
index b99dd8e1c93f..522510baed49 100644
@@ -17,17 +17,11 @@ MODULE_LICENSE("GPL");
 /*
  * Pull in the reference implementations while instructing GCC (through
  * -ftree-vectorize) to attempt to exploit implicit parallelism and emit
- * NEON instructions.
+ * NEON instructions. Clang does this by default at O2 so no pragma is
+ * needed.
  */
-#if __GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 6)
+#ifdef CONFIG_CC_IS_GCC
 #pragma GCC optimize "tree-vectorize"
-#else
-/*
- * While older versions of GCC do not generate incorrect code, they fail to
- * recognize the parallel nature of these functions, and emit plain ARM code,
- * which is known to be slower than the optimized ARM code in asm-arm/xor.h.
- */
-#warning This code requires at least version 4.6 of GCC
 #endif
 
 #pragma GCC diagnostic ignored "-Wunused-variable"
Signature · 1 line
2.30.2
maintainer
#3 lore raw
In reply to Ard Biesheuvel
Quoted 41 lines · Ard Biesheuvel

Update the xor_blocks() prototypes so that the compiler understands that the inputs always refer to distinct regions of memory. This is implied by the existing implementations, as they use different granularities for the load/xor/store loops.

With that, we can fix the ARM/Clang version, which refuses to SIMD vectorize otherwise, and throws a spurious warning related to the GCC version being incompatible.

Changes since v3:
- revert broken PPC argument rename - doing it fully results in too much pointless churn, and the 'inner' altivec routines are not strictly part of the xor_blocks API anyway

Changes since v2:
- fix arm64 build after rebase
- name PPC argument names consistently
- add Nick's acks and link tags

Changes since v1:
- fix PPC build
- add Nathan's Tested-by

Cc Nick Desaulniers <[email protected]>
Cc Nathan Chancellor <[email protected]>

Ard Biesheuvel (2):
lib/xor: make xor prototypes more friendly to compiler vectorization
crypto: arm/xor - make vectorized C code Clang-friendly

 arch/alpha/include/asm/xor.h           | 53 ++++++++----
 arch/arm/include/asm/xor.h             | 42 ++++++----
 arch/arm/lib/xor-neon.c                | 12 +--
 arch/arm64/include/asm/xor.h           | 21 +++--
 arch/arm64/lib/xor-neon.c              | 46 +++++++----
 arch/ia64/include/asm/xor.h            | 21 +++--
 arch/powerpc/include/asm/xor_altivec.h | 25 +++---
 arch/powerpc/lib/xor_vmx.c             | 28 ++++---
 arch/powerpc/lib/xor_vmx.h             | 27 ++++---
 arch/powerpc/lib/xor_vmx_glue.c        | 32 ++++----
 arch/s390/lib/xor.c                    | 21 +++--
Quoted 8 lines
 arch/sparc/include/asm/xor_32.h        | 21 +++--
 arch/sparc/include/asm/xor_64.h        | 42 ++++++----
 arch/x86/include/asm/xor.h             | 42 ++++++----
 arch/x86/include/asm/xor_32.h          | 42 ++++++----
 arch/x86/include/asm/xor_avx.h         | 21 +++--
 include/asm-generic/xor.h              | 84 +++++++++++++-------
 include/linux/raid/xor.h               | 21 +++--
 18 files changed, 384 insertions(+), 217 deletions(-)

All applied. Thanks.

Signature · 3 lines
Email: Herbert Xu <[email protected]>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt