From 6ae014d6ed7d641b7211be6d3f63e5107ddbdc11 Mon Sep 17 00:00:00 2001
From: Andrea Mazzoleni <amadvance@gmail.com>
Date: Mon, 5 Oct 2026 13:46:22 +0200
Subject: [PATCH] Align integer gen and rec loops to 64 bytes

---
 raid/NOTES      |  6 ++++++
 raid/int.c      | 19 +++++++++++++++++++
 raid/internal.h | 15 +++++++++++++++
 3 files changed, 40 insertions(+)

diff --git a/raid/NOTES b/raid/NOTES
index 20028207..b095fc2a 100644
--- a/raid/NOTES
+++ b/raid/NOTES
@@ -1055,6 +1055,12 @@ The implementation deliberately targets at most six parity levels.
 Short list of ideas already tried or rejected
 ---------------------------------------------
 
+* Loop alignment to 64 bytes:
+  improves scalar integer routines (raid/int.c) by +4% to +15% by avoiding fetch
+  and decode stalls on compact loops. Tried on vector SIMD routines (SSE, AVX,
+  NEON), with no improvement because wide unrolled loops are memory-bound and
+  cached in the uop cache. Kept only for integer routines.
+
 * AVX512 128-byte generation processing:
   tried, can be slower.
 
diff --git a/raid/int.c b/raid/int.c
index 0aa5148e..9587c4d4 100644
--- a/raid/int.c
+++ b/raid/int.c
@@ -9,6 +9,7 @@
  *
  * Uses 16-byte chunks across four 32-bit words.
  */
+__aligned_loops(64)
 void raid_gen1_int32(int nd, size_t size, void **vv, int streaming)
 {
 	uint8_t **v = (uint8_t **)vv;
@@ -49,6 +50,7 @@ void raid_gen1_int32(int nd, size_t size, void **vv, int streaming)
  *
  * Uses 32-byte chunks across four 64-bit words.
  */
+__aligned_loops(64)
 void raid_gen1_int64(int nd, size_t size, void **vv, int streaming)
 {
 	uint8_t **v = (uint8_t **)vv;
@@ -193,6 +195,7 @@ static __always_inline void raid_gen2_int64_gen(int nd, size_t size, void **vv,
  *
  * Uses Horner's method with 8-byte chunks across two 32-bit words.
  */
+__aligned_loops(64)
 void raid_genz_int32_raid(int nd, size_t size, void **vv, int streaming)
 {
 	uint8_t **v = (uint8_t **)vv;
@@ -250,6 +253,7 @@ void raid_genz_int32_raid(int nd, size_t size, void **vv, int streaming)
  *
  * Uses Horner's method with 16-byte chunks across two 64-bit words.
  */
+__aligned_loops(64)
 void raid_genz_int64_raid(int nd, size_t size, void **vv, int streaming)
 {
 	uint8_t **v = (uint8_t **)vv;
@@ -307,6 +311,7 @@ void raid_genz_int64_raid(int nd, size_t size, void **vv, int streaming)
  *
  * Precomputes per-disk multiplication tables and processes 2-byte chunks.
  */
+__aligned_loops(64)
 void raid_gen2_int8(int nd, size_t size, void **vv, int streaming)
 {
 	uint8_t **v = (uint8_t **)vv;
@@ -361,6 +366,7 @@ void raid_gen2_int8(int nd, size_t size, void **vv, int streaming)
  *
  * Precomputes per-disk multiplication tables and processes 2-byte chunks.
  */
+__aligned_loops(64)
 void raid_gen3_int8(int nd, size_t size, void **vv, int streaming)
 {
 	uint8_t **v = (uint8_t **)vv;
@@ -425,6 +431,7 @@ void raid_gen3_int8(int nd, size_t size, void **vv, int streaming)
  *
  * Precomputes per-disk multiplication tables and processes 2-byte chunks.
  */
+__aligned_loops(64)
 void raid_gen4_int8(int nd, size_t size, void **vv, int streaming)
 {
 	uint8_t **v = (uint8_t **)vv;
@@ -498,6 +505,7 @@ void raid_gen4_int8(int nd, size_t size, void **vv, int streaming)
  *
  * Precomputes per-disk multiplication tables and processes 2-byte chunks.
  */
+__aligned_loops(64)
 void raid_gen5_int8(int nd, size_t size, void **vv, int streaming)
 {
 	uint8_t **v = (uint8_t **)vv;
@@ -580,6 +588,7 @@ void raid_gen5_int8(int nd, size_t size, void **vv, int streaming)
  *
  * Precomputes per-disk multiplication tables and processes 2-byte chunks.
  */
+__aligned_loops(64)
 void raid_gen6_int8(int nd, size_t size, void **vv, int streaming)
 {
 	uint8_t **v = (uint8_t **)vv;
@@ -1008,30 +1017,35 @@ static __always_inline void raid_recX_int8(int nr, int has_p, int *id, int *ip,
 	}
 }
 
+__aligned_loops(64)
 void raid_gen2_int32_raid(int nd, size_t size, void **vv, int streaming)
 {
 	(void)streaming;
 	raid_gen2_int32_gen(nd, size, vv, 2);
 }
 
+__aligned_loops(64)
 void raid_gen2_int32_aes(int nd, size_t size, void **vv, int streaming)
 {
 	(void)streaming;
 	raid_gen2_int32_gen(nd, size, vv, 3);
 }
 
+__aligned_loops(64)
 void raid_gen2_int64_raid(int nd, size_t size, void **vv, int streaming)
 {
 	(void)streaming;
 	raid_gen2_int64_gen(nd, size, vv, 2);
 }
 
+__aligned_loops(64)
 void raid_gen2_int64_aes(int nd, size_t size, void **vv, int streaming)
 {
 	(void)streaming;
 	raid_gen2_int64_gen(nd, size, vv, 3);
 }
 
+__aligned_loops(64)
 void raid_rec1_int8(int nr, int *id, int *ip, int nd, size_t size, void **vv)
 {
 	BUG_ON(nr != 1);
@@ -1042,6 +1056,7 @@ void raid_rec1_int8(int nr, int *id, int *ip, int nd, size_t size, void **vv)
 		raid_recX_int8(1, 0, id, ip, nd, size, vv);
 }
 
+__aligned_loops(64)
 void raid_rec2_int8(int nr, int *id, int *ip, int nd, size_t size, void **vv)
 {
 	BUG_ON(nr != 2);
@@ -1052,6 +1067,7 @@ void raid_rec2_int8(int nr, int *id, int *ip, int nd, size_t size, void **vv)
 		raid_recX_int8(2, 0, id, ip, nd, size, vv);
 }
 
+__aligned_loops(64)
 void raid_rec3_int8(int nr, int *id, int *ip, int nd, size_t size, void **vv)
 {
 	BUG_ON(nr != 3);
@@ -1061,6 +1077,7 @@ void raid_rec3_int8(int nr, int *id, int *ip, int nd, size_t size, void **vv)
 		raid_recX_int8(3, 0, id, ip, nd, size, vv);
 }
 
+__aligned_loops(64)
 void raid_rec4_int8(int nr, int *id, int *ip, int nd, size_t size, void **vv)
 {
 	BUG_ON(nr != 4);
@@ -1070,6 +1087,7 @@ void raid_rec4_int8(int nr, int *id, int *ip, int nd, size_t size, void **vv)
 		raid_recX_int8(4, 0, id, ip, nd, size, vv);
 }
 
+__aligned_loops(64)
 void raid_rec5_int8(int nr, int *id, int *ip, int nd, size_t size, void **vv)
 {
 	BUG_ON(nr != 5);
@@ -1079,6 +1097,7 @@ void raid_rec5_int8(int nr, int *id, int *ip, int nd, size_t size, void **vv)
 		raid_recX_int8(5, 0, id, ip, nd, size, vv);
 }
 
+__aligned_loops(64)
 void raid_rec6_int8(int nr, int *id, int *ip, int nd, size_t size, void **vv)
 {
 	BUG_ON(nr != 6);
diff --git a/raid/internal.h b/raid/internal.h
index 4fbe9b89..0314c95a 100644
--- a/raid/internal.h
+++ b/raid/internal.h
@@ -61,7 +61,22 @@
  * Forced alignment.
  */
 #ifndef __aligned
+#if defined(__GNUC__) || defined(__clang__)
 #define __aligned(a) __attribute__((aligned(a)))
+#else
+#define __aligned(a)
+#endif
+#endif
+
+/*
+ * Forced loops alignment.
+ */
+#ifndef __aligned_loops
+#if defined(__GNUC__) && !defined(__clang__)
+#define __aligned_loops(n) __attribute__((optimize("align-loops=" #n)))
+#else
+#define __aligned_loops(n)
+#endif
 #endif
 
 /*