OpenE2K
/
binutils-gdb
12
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Fix PR 24571 - Relaxation does not shorten jmp or call to target at pc-relative range boundary 

The range check done to transform an absolute call/jump to a pc-relative one is
off-by-one, and that causes this shortening optimization to be missed if the
branch target is right at the range boundary.

In the non-shrinkable case, the range is what is mentioned in the ISA - -4094
bytes in the backward direction, and 4096 bytes in the positive direction.

In the shrinkable case, the forward jump range increases by two bytes (deleted
because of the shortening from call/jmp to rcall/rjmp), and therefore, the
range is -4094 in the reverse, and 4098 in the positive direction.

Fix the ranges for !shrinkable and shrinkable cases, and add a test caes to
ensure jumps to max forward and backward ranges get relaxed to rjmp.
This commit is contained in:
Senthil Kumar Selvaraj 2019-05-16 12:12:33 +05:30
parent 338ba75534
commit 7622049e0b
5 changed files with 55 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
bfd/ChangeLog
View File

@ -1,3 +1,9 @@
2019-05-21 Senthil Kumar Selvaraj <senthilkumar.selvaraj@microchip.com>
PR ld/24571
* bfd/elf32-avr.c (elf32_avr_relax_section): Adjust range check
when computing distance_short_enough.
2019-05-21 Senthil Kumar Selvaraj <senthilkumar.selvaraj@microchip.com>
PR ld/24564

18
bfd/elf32-avr.c
View File

@ -2643,16 +2643,28 @@ elf32_avr_relax_section (bfd *abfd,
/* Compute the distance from this insn to the branch target. */
gap = value - dot;
/* The ISA manual states that addressable range is PC - 2k + 1 to
PC + 2k. In bytes, that would be -4094 <= PC <= 4096. The range
is shifted one word to the right, because pc-relative instructions
implicitly add one word i.e. rjmp 0 jumps to next insn, not the
current one.
Therefore, for the !shrinkable case, the range is as above.
If shrinkable, then the current code only deletes bytes 3 and
4 of the absolute call/jmp, so the forward jump range increases
by 2 bytes, but the backward (negative) jump range remains
the same. */
/* Check if the gap falls in the range that can be accommodated
in 13bits signed (It is 12bits when encoded, as we deal with
word addressing). */
if (!shrinkable && ((int) gap >= -4096 && (int) gap <= 4095))
if (!shrinkable && ((int) gap >= -4094 && (int) gap <= 4096))
distance_short_enough = 1;
/* If shrinkable, then we can check for a range of distance which
is two bytes farther on both the directions because the call
is two bytes farther on the positive direction because the call
or jump target will be closer by two bytes after the
relaxation. */
else if (shrinkable && ((int) gap >= -4094 && (int) gap <= 4097))
else if (shrinkable && ((int) gap >= -4094 && (int) gap <= 4098))
distance_short_enough = 1;
/* Here we handle the wrap-around case. E.g. for a 16k device

6
ld/ChangeLog
View File

@ -1,3 +1,9 @@
2019-05-21 Senthil Kumar Selvaraj <senthilkumar.selvaraj@microchip.com>
PR ld/24571
* ld/testsuite/ld-avr/relax-insn-at-range-boundary.d: New test.
* ld/testsuite/ld-avr/relax-insn-at-range-boundary.s: New test.
2019-05-21 Senthil Kumar Selvaraj <senthilkumar.selvaraj@microchip.com>
PR ld/24564

17
ld/testsuite/ld-avr/relax-insn-at-range-boundary.d Normal file
View File

@ -0,0 +1,17 @@
#name: AVR relaxation, jump to symbol at ends of pc-relative range boundary
#as: -mlink-relax -mavr51
#ld: --relax -mavr51
#source: relax-insn-at-range-boundary.s
#objdump: -d
#target: avr-*-*
#...
00000000.*
...
ffc: 00 00 nop
ffe: 00 c8 rjmp .-4096 ; 0x0 .*
1000: ff c7 rjmp .+4094 ; 0x2000 <forward_target>
...
00002000 <forward_target>:
#...

11
ld/testsuite/ld-avr/relax-insn-at-range-boundary.s Normal file
View File

@ -0,0 +1,11 @@
.section ".text", "ax",@progbits
.global main
main:
backward_target: ; Exactly -4094 bytes from jmp
.ds.b 4094, 0
jmp backward_target
jmp forward_target
.ds.b 4094, 0
forward_target: ; Exactly 4098 bytes before relax, 4096 bytes after relax
nop