prefetch

Signed-off-by: Oliver Upton <oliver.upton@linux.dev>

10 files changed, 42 insertions, 0 deletions

diff --git a/Makefile b/Makefile
index a424394..c4b94f2 100644
--- a/Makefile
+++ b/Makefile
@@ -7,6 +7,7 @@ OBJS	+= memcpy_ldp_sttr.o
 OBJS	+= memcpy_ldp_stp.o
 OBJS	+= memcpy_ldr_stp.o
 OBJS	+= memcpy_ldtr_stp.o
+OBJS	+= memcpy_prfm_2k_ldp_str.o
 
 memcpy: $(OBJS)
 	$(CC) $(CFLAGS) $(OBJS) -o $@
diff --git a/copy_template.h b/copy_template.h
index 3ed0124..3c15f25 100644
--- a/copy_template.h
+++ b/copy_template.h
@@ -18,6 +18,7 @@ D_h	.req	x10
 	sub	count, count, #64
 
 1:
+	prfm1	src
 	stp1	A_l, A_h, dst, #16
 	ldp1	A_l, A_h, src, #16
 	stp1	B_l, B_h, dst, #16
diff --git a/main.c b/main.c
index 3542501..6f28f68 100644
--- a/main.c
+++ b/main.c
@@ -20,6 +20,7 @@ enum mode {
 	LDR_STP,
 	LDTR_STP,
 	LDP_STP,
+	PRFM_2K_LDP_STR,
 
 	NR_MODES
 };
@@ -40,6 +41,7 @@ static void pr_modes(void)
 	PR_MODE(LDR_STP);
 	PR_MODE(LDTR_STP);
 	PR_MODE(LDP_STP);
+	PR_MODE(PRFM_2K_LDP_STR);
 }
 
 static void pr_help(const char *progname)
@@ -92,6 +94,9 @@ static void do_memcpy(void *dst, const void *src, size_t chunk_size)
 	case LDP_STP:
 		memcpy_ldp_stp(dst, src, chunk_size);
 		break;
+	case PRFM_2K_LDP_STR:
+		memcpy_prfm_2k_ldp_str(dst, src, chunk_size);
+		break;
 	default:
 		assert(0);
 	}
diff --git a/memcpy.h b/memcpy.h
index e0fce8c..947177c 100644
--- a/memcpy.h
+++ b/memcpy.h
@@ -8,5 +8,6 @@ void memcpy_ldp_sttr(void *dst, const void *src, size_t count);
 void memcpy_ldr_stp(void *dst, const void *src, size_t count);
 void memcpy_ldtr_stp(void *dst, const void *src, size_t count);
 void memcpy_ldp_stp(void *dst, const void *src, size_t count);
+void memcpy_prfm_2k_ldp_str(void *dst, const void *src, size_t count);
 
 #endif  /* __MEMCPY_H__ */
diff --git a/memcpy_ldp_stp.S b/memcpy_ldp_stp.S
index a9851f5..87b676f 100644
--- a/memcpy_ldp_stp.S
+++ b/memcpy_ldp_stp.S
@@ -1,3 +1,6 @@
+.macro	prfm1, Xn
+.endm
+
 .macro	ldp1, Xt1, Xt2, Xn, imm
 	ldp	\Xt1, \Xt2, [\Xn], \imm
 .endm
diff --git a/memcpy_ldp_str.S b/memcpy_ldp_str.S
index 0b51165..a61eddb 100644
--- a/memcpy_ldp_str.S
+++ b/memcpy_ldp_str.S
@@ -1,3 +1,6 @@
+.macro	prfm1, Xn
+.endm
+
 .macro	ldp1, Xt1, Xt2, Xn, imm
 	ldp	\Xt1, \Xt2, [\Xn], \imm
 .endm
diff --git a/memcpy_ldp_sttr.S b/memcpy_ldp_sttr.S
index b517d7a..e346104 100644
--- a/memcpy_ldp_sttr.S
+++ b/memcpy_ldp_sttr.S
@@ -1,3 +1,6 @@
+.macro	prfm1, Xn
+.endm
+
 .macro	ldp1, Xt1, Xt2, Xn, imm
 	ldp	\Xt1, \Xt2, [\Xn], \imm
 .endm
diff --git a/memcpy_ldr_stp.S b/memcpy_ldr_stp.S
index 78aed07..2c0081b 100644
--- a/memcpy_ldr_stp.S
+++ b/memcpy_ldr_stp.S
@@ -1,3 +1,6 @@
+.macro	prfm1, Xn
+.endm
+
 .macro	ldp1, Xt1, Xt2, Xn, imm
 	ldr	\Xt1, [\Xn]
 	ldr	\Xt2, [\Xn, #8]
diff --git a/memcpy_ldtr_stp.S b/memcpy_ldtr_stp.S
index e7ea408..df8eb05 100644
--- a/memcpy_ldtr_stp.S
+++ b/memcpy_ldtr_stp.S
@@ -1,3 +1,6 @@
+.macro	prfm1, Xn
+.endm
+
 .macro	ldp1, Xt1, Xt2, Xn, imm
 	ldtr	\Xt1, [\Xn]
 	ldtr	\Xt2, [\Xn, #8]
diff --git a/memcpy_prfm_2k_ldp_str.S b/memcpy_prfm_2k_ldp_str.S
new file mode 100644
index 0000000..6d4a52c
--- /dev/null
+++ b/memcpy_prfm_2k_ldp_str.S
@@ -0,0 +1,19 @@
+.macro	prfm1, Xn
+	prfm	pldl2keep, [\Xn, #0x800]
+	prfm	pldl2keep, [\Xn, #0x840]
+.endm
+
+.macro	ldp1, Xt1, Xt2, Xn, imm
+	ldp	\Xt1, \Xt2, [\Xn], \imm
+.endm
+
+.macro	stp1, Xt1, Xt2, Xn, imm
+	str	\Xt1, [\Xn]
+	str	\Xt2, [\Xn, #8]
+	add	\Xn, \Xn, \imm
+.endm
+
+.globl memcpy_prfm_2k_ldp_str
+memcpy_prfm_2k_ldp_str:
+	#include "copy_template.h"
+	ret