diff --git a/sys/arch/sparc64/stand/ofwboot/Makefile b/sys/arch/sparc64/stand/ofwboot/Makefile
index c256b4d3a769..a98aaef70c6f 100644
--- a/sys/arch/sparc64/stand/ofwboot/Makefile
+++ b/sys/arch/sparc64/stand/ofwboot/Makefile
@@ -1,4 +1,4 @@
-#	$NetBSD: Makefile,v 1.13 2000/01/23 17:04:11 mycroft Exp $
+#	$NetBSD: Makefile,v 1.14 2000/03/06 01:29:04 eeh Exp $
 
 CURDIR=	${.CURDIR}
 S=	${CURDIR}/../../../..
@@ -14,9 +14,9 @@ SRCS=		srt0.s Locore.c boot.c ofdev.c alloc.c net.c netif_of.c
 .PATH:		${S}/arch/sparc64/sparc64
 
 .if CROSS
-CEXTRAFLAGS=	-nostdinc
+CEXTRAFLAGS?=	-nostdinc
 .else
-CEXTRAFLAGS=	-D_LP64
+CEXTRAFLAGS?=	-D_LP64
 .endif
 CFLAGS+=	${COPTS} -DNETIF_DEBUG -D__ELF__ ${CEXTRAFLAGS} -D_STANDALONE
 
@@ -91,7 +91,7 @@ ${PROG}: ${OBJS} ${LIBSA} ${LIBZ} ${LIBKERN}
 .if CROSS
 AFLAGS=		-x assembler-with-cpp -Wa,-Av9a -traditional-cpp -D_LOCORE -D__ELF__
 .else
-AFLAGS=		-x assembler-with-cpp -traditional-cpp -D_LOCORE -D__ELF__ -D_LP64 -Wa,-Av9 #-Wa,-Av8plusa
+AFLAGS=		-x assembler-with-cpp -traditional-cpp -D_LOCORE -D__ELF__ -Wa,-Av9a ${CEXTRAFLAGS} #-Wa,-Av8plusa
 .endif
 NORMAL_S=	${CC} ${AFLAGS} ${CPPFLAGS} -c $<
 srt0.o: srt0.s
diff --git a/sys/arch/sparc64/stand/ofwboot/boot.c b/sys/arch/sparc64/stand/ofwboot/boot.c
index ca93aeb30f04..8b66b7c9ae3d 100644
--- a/sys/arch/sparc64/stand/ofwboot/boot.c
+++ b/sys/arch/sparc64/stand/ofwboot/boot.c
@@ -1,4 +1,4 @@
-/*	$NetBSD: boot.c,v 1.10 1999/10/25 14:04:38 kleink Exp $	*/
+/*	$NetBSD: boot.c,v 1.11 2000/03/06 01:29:04 eeh Exp $	*/
 #define DEBUG
 /*
  * Copyright (c) 1997, 1999 Eduardo E. Horvath.  All rights reserved.
@@ -61,6 +61,8 @@
 #include <sparc64/stand/ofwboot/ofdev.h>
 #include <sparc64/stand/ofwboot/openfirm.h>
 
+#define	MEG	(1024*1024)
+
 /*
  * Boot device is derived from ROM provided information, or if there is none,
  * this list is used in sequence, to find a kernel.
@@ -403,6 +405,7 @@ elf32_exec(fd, elf, entryp, ssymp, esymp)
 	void *addr;
 	size_t size;
 	int i, first = 1;
+	long align;
 	int n;
 
 	/*
@@ -429,7 +432,15 @@ elf32_exec(fd, elf, entryp, ssymp, esymp)
 		printf("%s%lu@0x%lx", first ? "" : "+", phdr.p_filesz,
 		    (u_long)phdr.p_vaddr);
 		(void)lseek(fd, phdr.p_offset, SEEK_SET);
-/* NB need to do 4MB allocs here */
+
+		/* 
+		 * If the segment's VA is aligned on a 4MB boundary, align its
+		 * request 4MB aligned physical memory.  Otherwise use default
+		 * alignment.
+		 */
+		align = phdr.p_align;
+		if ((phdr.p_vaddr & (4*MEG-1)) == 0)
+			align = 4*MEG;
 		if (OF_claim((void *)phdr.p_vaddr, phdr.p_memsz, phdr.p_align) ==
 		    (void *)-1)
 			panic("cannot claim memory");
diff --git a/sys/arch/sparc64/stand/ofwboot/elfXX_exec.c b/sys/arch/sparc64/stand/ofwboot/elfXX_exec.c
index 1068db056fef..7e3bc1590c1b 100644
--- a/sys/arch/sparc64/stand/ofwboot/elfXX_exec.c
+++ b/sys/arch/sparc64/stand/ofwboot/elfXX_exec.c
@@ -1,6 +1,7 @@
-/*	$NetBSD: elfXX_exec.c,v 1.4 1999/10/25 14:04:38 kleink Exp $	*/
+/*	$NetBSD: elfXX_exec.c,v 1.5 2000/03/06 01:29:04 eeh Exp $	*/
 
 /*
+ * Copyright (c) 1998-2000 Eduardo Horvath.  All rights reserved.
  * Copyright (c) 1997 Jason R. Thorpe.  All rights reserved.
  * Copyright (C) 1995, 1996 Wolfgang Solfrank.
  * Copyright (C) 1995, 1996 TooLs GmbH.
@@ -44,6 +45,7 @@
 
 #define CONCAT(x,y)	__CONCAT(x,y)
 #define CAT3(s,m,e)	CONCAT(s,CONCAT(m,e))
+#define	MEG	(1024*1024)
 
 #if 0
 int	CAT3(elf,ELFSIZE,_exec) __P((int, CAT3(Elf,ELFSIZE,_Ehdr) *, u_int64_t *, void **, void **));
@@ -61,6 +63,7 @@ CAT3(elf, ELFSIZE, _exec)(fd, elf, entryp, ssymp, esymp)
 	CAT3(Elf,ELFSIZE,_Off) off;
 	void *addr;
 	size_t size;
+	u_int align;
 	int i, first = 1;
 	int n;
 
@@ -72,18 +75,14 @@ CAT3(elf, ELFSIZE, _exec)(fd, elf, entryp, ssymp, esymp)
 	printf("elf%d_exec: ", ELFSIZE);
 #endif
 	printf("Booting %s\n", opened_name);
-printf("reading %ld program headers\n", (long)elf->e_phnum);
 
 	for (i = 0; i < elf->e_phnum; i++) {
 		CAT3(Elf,ELFSIZE,_Phdr) phdr;
-printf("reading phdr %d at %lx\n", i, (long)(elf->e_phoff + sizeof(phdr) * i));
 		size = lseek(fd, (size_t)(elf->e_phoff + sizeof(phdr) * i), SEEK_SET);
-printf("lseek sez: %lx %s\n", (long)size, (size<0)?strerror(errno):"");
 		if (read(fd, (void *)&phdr, sizeof(phdr)) != sizeof(phdr)) {
 			printf("read phdr: %s\n", strerror(errno));
 			return (1);
 		}
-printf("reading phdr worked, type %lx flags %lx\n", (long)phdr.p_type, (long)phdr.p_flags);
 		if (phdr.p_type != PT_LOAD ||
 		    (phdr.p_flags & (PF_W|PF_X)) == 0)
 			continue;
@@ -92,8 +91,15 @@ printf("reading phdr worked, type %lx flags %lx\n", (long)phdr.p_type, (long)phd
 		printf("%s%lu@0x%lx", first ? "" : "+", (u_long)phdr.p_filesz,
 		    (u_long)phdr.p_vaddr);
 		(void)lseek(fd, (size_t)phdr.p_offset, SEEK_SET);
-/* NB need to do 4MB allocs here */
-		if (OF_claim((void *)(long)phdr.p_vaddr, phdr.p_memsz, phdr.p_align) ==
+		/* 
+		 * If the segment's VA is aligned on a 4MB boundary, align its
+		 * request 4MB aligned physical memory.  Otherwise use default
+		 * alignment.
+		 */
+		align = phdr.p_align;
+		if ((phdr.p_vaddr & (4*MEG-1)) == 0)
+			align = 4*MEG;
+		if (OF_claim((void *)(long)phdr.p_vaddr, phdr.p_memsz, align) ==
 		    (void *)-1)
 			panic("cannot claim memory");
 		if (read(fd, (void *)(long)phdr.p_vaddr, phdr.p_filesz) !=