The MIPS %hi and %lo relocation operators act as a pair on REL targets;
you need to see the partnering %lo in order to calculate the addend for
a %hi and detect carries properly. The assembler therefore complains if
%hi is used without a corresponding %lo.
The MIPS backend tries to remove unmatched %his in mips_reorg.
The problem is that it can wrongly keep a %hi(foo) in cases like:
asm ("" : "=m" (foo));
where a LO_SUM MEM appears as an operand at the rtl level but where
%lo(foo) doesn't appear in the assembly code.
It's very hard without fully parsing an asm to know whether it will need
the LO_SUM or not. E.g. the operand could appear in the asm string
but could be protected by an .if or only be used in a comment. But it
should be safe to assume that the LO_SUM in operand N isn't needed if
%N or %<letter>N doesn't appear in the string, since AFAIK there's
no other way for the asm to know which register contains the high part.
Specifically, things like:
asm ("sw $0,foo" : "=m" (foo));
work fine either way, since the %hi is implicit in the SW macro.
I split out the routine to detect which operands are used since it
seemed like it might be more generally useful.
Tested on mips64-linux-gnu, where it fixes the gcc.dg/guality/pr36728-2.c
failures for -mabi=32. OK for the target-independent bits?
Thanks,
Richard
gcc/
PR target/52125
* rtl.h (get_referenced_operands): Declare.
* recog.c (get_referenced_operands): New function.
* config/mips/mips.c (mips_reorg_process_insns): Check which asm
operands have been referenced when recording LO_SUM references.
gcc/testsuite/
PR target/52125
* gcc.dg/pr48774.c: Remove skip for mips_rel.
* gcc.target/mips/pr52125.c: New test.