ld: pru: Merge the bss input sections into data

The popular method to load PRU firmware is through the remoteproc Linux
kernel driver.  In order to save a few bytes from the firmware, the PRU
CRT0 is spared from calling memset for the bss segment [1].  Instead the
host loader is supposed to zero out the bss segment.  This is important
for PRU, which typically has only 8KB for instruction memory.

The legacy non-mainline PRU host driver relied on the default
behaviour of the kernel core remoteproc [2].  That default is to zero
out the loadable memory regions not backed by file storage (i.e. the
bss sections).  This worked for the libgloss' CRT0.

But the PRU loader merged in mainline Linux explicitly changes the
default behaviour [3].  It no longer is zeroing out memory regions.
Hence the bss sections are not initialized - neither by CRT0, nor by the
host loader.

This patch fixes the issue by aligning the GNU LD default linker script
with the mainline Linux kernel expectation.  Since the mainline kernel
driver is submitted by the PRU manufacturer itself (Text Instruments),
we can consider that as defining the ABI.

This change has been tested on Beaglebone AI-64 [4].  Static counter
variables in the firmware are now always starting from zero, as
expected.  There was only one new toolchain test failure in orphan3.d,
due to reordering of the output sections.  I believe this is a harmless
issue.  I could not rewrite the PASS criteria to ignore the output
section ordering, so I have disabled that test case for PRU.

[1] https://sourceware.org/git/?p=newlib-cygwin.git;a=blob;f=libgloss/pru/crt0.S;h=b3f0d53a93acc372f461007553e7688ca77753c9;hb=HEAD#l40
[2] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/remoteproc/remoteproc_elf_loader.c?h=v6.1#n228
[3] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/remoteproc/pru_rproc.c?h=v6.1#n641
[4] https://beagleboard.org/ai-64

ld/ChangeLog:

	* scripttempl/pru.sc (.data): Merge .bss input sections into the
	.data output section.
	* testsuite/ld-elf/orphan3.d: Disable for PRU.

Signed-off-by: Dimitar Dimitrov <dimitar@dinux.eu>

2 files changed, 15 insertions, 11 deletions

diff --git a/ld/scripttempl/pru.sc b/ld/scripttempl/pru.sc
index 56d07be..7cb8069 100644
--- a/ld/scripttempl/pru.sc
+++ b/ld/scripttempl/pru.sc
@@ -149,17 +149,13 @@ SECTIONS
     ${RELOCATING+*(.gnu.linkonce.r*)}
     ${RELOCATING+. = ALIGN(4);}
     ${RELOCATING+ PROVIDE (_data_end = .) ; }
-  } ${RELOCATING+ > dmem }
 
-  /* Linux remoteproc loader requires the resource_table section
-     start address to be aligned to 8 bytes.  */
-  .resource_table ${RELOCATING-0} ${RELOCATING+ ALIGN(8)} :
-  {
-    KEEP (*(.resource_table))
-  } ${RELOCATING+ > dmem}
-
-  .bss ${RELOCATING-0} :
-  {
+    ${RELOCATING+/* Merge the bss input sections into the output
+      data section.  The Linux kernel's remoteproc PRU ELF loader
+      will not memzero the bss section.  The CRT0 will not either, in order
+      to reduce the final firmware's instruction memory size.  Hence
+      present bss sections as regular data sections, at the negligible
+      expense of increasing the ELF file size.  */}
     ${RELOCATING+ PROVIDE (_bss_start = .) ; }
     *(.bss)
     ${RELOCATING+ *(.bss.*)}
@@ -167,6 +163,14 @@ SECTIONS
     ${RELOCATING+*(.gnu.linkonce.b*)}
     ${RELOCATING+*(COMMON)}
     ${RELOCATING+ PROVIDE (_bss_end = .) ; }
+
+  } ${RELOCATING+ > dmem}
+
+  /* Linux remoteproc loader requires the resource_table section
+     start address to be aligned to 8 bytes.  */
+  .resource_table ${RELOCATING-0} ${RELOCATING+ ALIGN (8)} :
+  {
+    KEEP (*(.resource_table))
   } ${RELOCATING+ > dmem}
 
   /* Global data not cleared after reset.  */
diff --git a/ld/testsuite/ld-elf/orphan3.d b/ld/testsuite/ld-elf/orphan3.d
index af6ce25..ec10164 100644
--- a/ld/testsuite/ld-elf/orphan3.d
+++ b/ld/testsuite/ld-elf/orphan3.d
@@ -7,7 +7,7 @@
 #ld:
 #readelf: -S --wide
 #xfail: [uses_genelf]
-#xfail: xstormy16-*-*
+#xfail: xstormy16-*-* pru-*-*
 
 #...
   \[[ 0-9]+\] \.foo +PROGBITS +[0-9a-f]+ +[0-9a-f]+ +0+20 +0+ +A +0 +0 +[0-9]+