On 30/09/15 11:42, Claudiu Zissulescu wrote:
This patch adds basic support for Synopsys' ARCv2 CPUs.
There is an awful lot of places that have TARGET_ARC700 || TARGET_V2 etc.
Maybe it's time for some new feature-oriented macros? Like:
TARGET_ARC600_FAMILY ARC600 and ARC601, for ARC600 pipeline idiosyncrasies
(currently tested by ruling out everything else... OK when that was just the
ARC700, and ARCtangent-A5 shared lots of the same issues, but
backwards now)
TARGET_LP_WR_INTERLOCK loop count register can be read in very next
instruction after it has been written to by an ordinary instruction
(Which unfortunately comes at a penalty for every LP use).
TARGET_MPY MPY support
TARGET_SLOW_CARRY high latency for carry input to arithmetic
comments to the patch in detail:
arc.c
For arc_secondary_reload:
The new code is puzzling at first glance (address reloading is generally
the job of find_reloads_toplev and its direct and indirect callees).
Sifting through the port I can see why it may be needed, but this is far
from obvious, so it requires a comment. Something like:
/* For ARC base register + offset addressing, the validity of the address
is mode-dependent for most of the offset range, as the offset can be
scaled by the access size.
We don't expose these as mode-dependent addresses in the
mode_dependent_address_p target hook, because that would disable
lots of optimizations, and most uses of these addresses are for 32 or
64 bit accesses anyways, which are fine.
However, that leaves some addresses for 8 / 16 bit values not properly
reloaded by the generic code, which is why we have to schedule
secondary
reloads for these. */
Also, the test is unnecessarily strict for HImode. Instead of
(INTVAL (XEXP (addr, 1)) < -256
|| INTVAL (XEXP (addr, 1)) > 255)
you can use:
!RTX_OK_FOR_OFFSET_P (mode, XEXP (addr, 1))
which takes potential scaling for HImode into account.
If you want to allow LIMMs for loads, this is also the place to
avoid unnecessary reloads in the first place - rather than making them
no-ops as the comment in arc_secondary_reload_conv suggests.
In arc_secondary_reload, you would just have to demur scheduling a reloads
for a large offset if in_p is set (that wouldn't allow them in the general
case, though - just the mode-dependent stuff).
To make this work, there would have to be a pattern-alternative to
match, though. Currently, we don't allow large offsets for ARC memory
operands. You could create a special constraint (not officially a memory
constraint) that allows MEMs with large offsets, and add it to the load
alternatives of the move patterns.
Another approach you could experiment with (don't know if it'll work)
would be to allow the full 32 bit offset range in arc_legitimate_address_p,
but claim that that which can not be stored with a 32 bit store is mode
dependent. The SH is sort of a precedent for this as it had pre-decrement /
post-increment addresses which are valid only for store / load,
respectively.
But it also has to use predicates to separate load and store operands.
You'd probably will also need to make arc's move_dest_operand predicate
reject larger offsets in MEMs.
We already reject scaled indices there.
arc_secondary_reload_conv:
I don't see why you test the address range again - isn't this function
only called when you have already scheduled a reload for the express
purpose of reloading the address?
If you do want to re-do the test (for sanity checking?), again,
!RTX_OK_FOR_OFFSET_P (GET_MODE (mem), XEXP (addr, 1))
will consider the scaling for HImode.
in arc_loop_hazard
typo (moved, pre-existing, but still...): instert
arc.md
"*commutative_binary_mult_comparison_result_used":
You will also have to fix mult_operator in predicates.md to make this work
for TARGET_V2 .
"mulhisi3_imm" / "umulhisi3_imm"
I suspect that the hardware does not support multiplying with arbitrary
32 bit constants. If so, the predicates and constraints for these
alternatives are wrong. Without these properly in place, cse / combine etc.
might generate constants the hardware process as the compiler thinks it
does.
"simple_return"
Here you say a return from interrupt is nocond - but you don't change
the "p_return_i" pattern, so presumably the machine-independent
if-conversion
can still generate a conditional return from interrupt.
predicates.md:
proper_comparison_operator
When are we passed a comparison where the first operand is VOIDmode?
I thought we generally canonicalize so that this does not happen?
