As diagnosed by the submitter, LRA can generate unaligned accesses when
SLOW_UNALIGNED_ACCESS is 1, for example when STRICT_ALIGNMENT is 1, because
simplify_operand_subreg can wrongly reject aligned accesses when reloading
SUBREGs of MEMs in favor of unaligned ones. The fix is to add the missing
guard on the alignment before invoking SLOW_UNALIGNED_ACCESS (as in every
other use of SLOW_UNALIGNED_ACCESS in the compiler).
Tested on arm-eabi, approved by Vladimir, applied on the mainline.
2016-11-17 Pip Cet <pip...@gmail.com>
Eric Botcazou <ebotca...@adacore.com>
PR rtl-optimization/78355
* doc/tm.texi.in (SLOW_UNALIGNED_ACCESS): Document that the macro
only needs to deal with unaligned accesses.
* doc/tm.texi: Regenerate.
* lra-constraints.c (simplify_operand_subreg): Only invoke
SLOW_UNALIGNED_ACCESS on innermode if the MEM is not aligned enough.
--
Eric Botcazou
Index: doc/tm.texi.in
===================================================================
--- doc/tm.texi.in (revision 242377)
+++ doc/tm.texi.in (working copy)
@@ -4654,7 +4654,8 @@ other fields in the same word of the str
Define this macro to be the value 1 if memory accesses described by the
@var{mode} and @var{alignment} parameters have a cost many times greater
than aligned accesses, for example if they are emulated in a trap
-handler.
+handler. This macro is invoked only for unaligned accesses, i.e. when
+@code{@var{alignment} < GET_MODE_ALIGNMENT (@var{mode})}.
When this macro is nonzero, the compiler will act as if
@code{STRICT_ALIGNMENT} were nonzero when generating code for block
Index: lra-constraints.c
===================================================================
--- lra-constraints.c (revision 242377)
+++ lra-constraints.c (working copy)
@@ -1486,9 +1486,10 @@ simplify_operand_subreg (int nop, machin
equivalences in function lra_constraints) and because for spilled
pseudos we allocate stack memory enough for the biggest
corresponding paradoxical subreg. */
- if (!SLOW_UNALIGNED_ACCESS (mode, MEM_ALIGN (reg))
- || SLOW_UNALIGNED_ACCESS (innermode, MEM_ALIGN (reg))
- || MEM_ALIGN (reg) >= GET_MODE_ALIGNMENT (mode))
+ if (!(MEM_ALIGN (reg) < GET_MODE_ALIGNMENT (mode)
+ && SLOW_UNALIGNED_ACCESS (mode, MEM_ALIGN (reg)))
+ || (MEM_ALIGN (reg) < GET_MODE_ALIGNMENT (innermode)
+ && SLOW_UNALIGNED_ACCESS (innermode, MEM_ALIGN (reg))))
return true;
/* INNERMODE is fast, MODE slow. Reload the mem in INNERMODE. */