[Cbe-oss-dev] [PATCH] libspe2: Tweak extended __ea support

[Alan Modra]

This patch delays libspe relocation processing until after the image
has been copied to ld_buffer.  The benefit in doing so is that only
relocations applied to the ._ea section need to be applied directly to
the image.  Others that apply to loaded spu segments can be applied
to ld_buffer, which potentially saves memory since their ppu image
pages are not dirtied.

I've also fixed a bug in casting that could affect a 32-bit libspe
handling an -mea64 spu object.  (Elf64_Xword)(long)start might result
in the high 32 bits being set to 1's due to sign extension.  I don't
know whether this is a problem in practice, but it seems better to be
safe by keeping the high 32 bits zero in this case.

Signed-off-by: Alan Modra <amodra at bigpond.net.au>

-- 
Alan Modra
Australia Development Lab, IBM

diff -urp libspe2-2.2.80.old/spebase/elf_loader.c libspe2-2.2.80/spebase/elf_loader.c
--- libspe2-2.2.80.old/spebase/elf_loader.c	2008-04-12 17:03:43.000000000 +0930
+++ libspe2-2.2.80/spebase/elf_loader.c	2008-05-12 07:53:27.000000000 +0930
@@ -201,17 +201,26 @@ copy_to_ld_buffer(spe_program_handle_t *
    without a symbol, which are to locations within ._ea.  */
 
 static void
-apply_relocations(spe_program_handle_t *handle, Elf32_Shdr *rh, Elf32_Shdr *sh)
+apply_relocations (spe_program_handle_t *handle,
+		   void *buf,
+		   Elf32_Shdr *rh,
+		   Elf32_Shdr *sh,
+		   Elf32_Shdr *eah)
 {
 #define R_SPU_PPU32 15
 #define R_SPU_PPU64 16
 	void *start = handle->elf_image;
 	Elf32_Rela *r, *r_end;
-	/* Relocations in an executable specify r_offset as a virtual
-	   address, but we are applying the reloc in the image before
-	   the section has been copied to its destination sh_addr.
-	   Adjust so as to poke relative to the image base.  */
-	void *reloc_base = start + sh->sh_offset - sh->sh_addr;
+	void *reloc_base = buf;
+
+	if (sh == eah)
+	{
+		/* Relocations in an executable specify r_offset as a
+		   virtual address, but if we are applying relocs for
+		   ._ea then the location we want is in the image.
+		   Adjust so as to poke relative to the image base.  */
+		reloc_base = start + sh->sh_offset - sh->sh_addr;
+	}
 
 	r = start + rh->sh_offset;
 	r_end = (void *)r + rh->sh_size;
@@ -219,15 +228,17 @@ apply_relocations(spe_program_handle_t *
 	for (; r < r_end; ++r)
 	{
 		if (r->r_info == ELF32_R_INFO(0,R_SPU_PPU32)) {
+			/* v is in ._ea */
 			Elf32_Word *loc = reloc_base + r->r_offset;
-			Elf32_Word v = (Elf32_Word)(long)start + r->r_addend;
+			Elf32_Word v = (unsigned long)start + r->r_addend;
 			/* Don't dirty pages unnecessarily.  */
 			if (*loc != v)
 				*loc = v;
 			DEBUG_PRINTF("PPU32(%p) = %#x\n", loc, v);
 		} else if (r->r_info == ELF32_R_INFO(0,R_SPU_PPU64)) {
+			/* v is in ._ea */
 			Elf64_Xword *loc = reloc_base + r->r_offset;
-			Elf64_Xword v = (Elf64_Xword)(long)start + r->r_addend;
+			Elf64_Xword v = (unsigned long)start + r->r_addend;
 			if (*loc != v)
 				*loc = v;
 			DEBUG_PRINTF("PPU64(%p) = %#llx\n", loc, v);
@@ -244,6 +255,7 @@ _base_spe_load_spe_elf (spe_program_hand
 
 	Elf32_Shdr *shdr;
 	Elf32_Shdr *sh;
+	Elf32_Shdr *eah;
 
 	Elf32_Off  toe_addr = 0;
 	long	toe_size = 0;
@@ -272,17 +284,18 @@ _base_spe_load_spe_elf (spe_program_hand
 
 	/* traverse the sections to locate the toe segment */
 	/* by specification, the toe sections are grouped together in a segment */
+	eah = 0;
 	for (sh = shdr; sh < &shdr[ehdr->e_shnum]; ++sh)
 	{
 		DEBUG_PRINTF("section name: %s ( start: 0x%04x, size: 0x%04x)\n", str_table+sh->sh_name, sh->sh_offset, sh->sh_size );
-		if (sh->sh_type == SHT_RELA)
-			apply_relocations(handle, sh, &shdr[sh->sh_info]);
 		if (strcmp(".toe", str_table+sh->sh_name) == 0) {
 			DEBUG_PRINTF("section offset: %d\n", sh->sh_offset);
 			toe_size += sh->sh_size;
 			if ((toe_addr == 0) || (toe_addr > sh->sh_addr))
 				toe_addr = sh->sh_addr;
 		}
+		else if (strcmp("._ea", str_table+sh->sh_name) == 0)
+			eah = sh;
 		/* Disabled : Actually not needed, only good for testing
 		if (strcmp(".bss", str_table+sh->sh_name) == 0) {
 			DEBUG_PRINTF("zeroing .bss section:\n");
@@ -328,6 +341,12 @@ _base_spe_load_spe_elf (spe_program_hand
 		  return -errno;
 	  }
 
+	if (eah)
+		for (sh = shdr; sh < &shdr[ehdr->e_shnum]; ++sh)
+			if (sh->sh_type == SHT_RELA)
+				apply_relocations (handle, ld_buffer,
+						   sh, &shdr[sh->sh_info], eah);
+
 	/* Remember where the code wants to be started */
 	ld_info->entry = ehdr->e_entry;
 	DEBUG_PRINTF ("entry = 0x%x\n", ehdr->e_entry);