diff options
author | Jakub Jelinek <jakub@redhat.com> | 2022-04-26 10:11:58 +0200 |
---|---|---|
committer | Jakub Jelinek <jakub@redhat.com> | 2022-04-26 10:11:58 +0200 |
commit | 7d31c678d68d7b6820a958584619ca763b0eb9c5 (patch) | |
tree | 3f8d3e92e1860d492cbaffdef9beadd65404187e /gcc/ifcvt.cc | |
parent | d4836ac9acd0c991a4fe1dec9438773a2c4eb5ac (diff) | |
download | gcc-7d31c678d68d7b6820a958584619ca763b0eb9c5.zip gcc-7d31c678d68d7b6820a958584619ca763b0eb9c5.tar.gz gcc-7d31c678d68d7b6820a958584619ca763b0eb9c5.tar.bz2 |
ifcvt: Improve noce_try_store_flag_mask [PR105314]
The following testcase regressed on riscv due to the splitting of critical
edges in the sink pass, similarly to x86_64 compared to GCC 11 we now swap
the edges, whether true or false edge goes to an empty forwarded bb.
From GIMPLE POV, those 2 forms are equivalent, but as can be seen here, for
some ifcvt opts it matters one way or another.
On this testcase, noce_try_store_flag_mask used to trigger and transformed
if (pseudo2) pseudo1 = 0;
into
pseudo1 &= -(pseudo2 == 0);
But with the swapped edges ifcvt actually sees
if (!pseudo2) pseudo3 = pseudo1; else pseudo3 = 0;
and noce_try_store_flag_mask punts. IMHO there is no reason why it
should punt those, it is equivalent to
pseudo3 = pseudo1 & -(pseudo2 == 0);
and especially if the target has 3 operand AND, it shouldn't be any more
costly (and even with 2 operand AND, it might very well happen that RA
can make it happen without any extra moves).
Initially I've just removed the rtx_equal_p calls from the conditions
and didn't add anything there, but that broke aarch64 bootstrap and
regressed some testcases on x86_64, where if_info->a or if_info->b could be
some larger expression that we can't force into a register.
Furthermore, the case where both if_info->a and if_info->b are constants is
better handled by other ifcvt optimizations like noce_try_store_flag
or noce_try_inverse_constants or noce_try_store_flag_constants.
So, I've restricted it to just a REG (perhaps SUBREG of REG might be ok too)
next to what has been handled previously.
2022-04-26 Jakub Jelinek <jakub@redhat.com>
PR rtl-optimization/105314
* ifcvt.cc (noce_try_store_flag_mask): Don't require that the non-zero
operand is equal to if_info->x, instead use the non-zero operand
as one of the operands of AND with if_info->x as target.
* gcc.target/riscv/pr105314.c: New test.
Diffstat (limited to 'gcc/ifcvt.cc')
-rw-r--r-- | gcc/ifcvt.cc | 6 |
1 files changed, 3 insertions, 3 deletions
diff --git a/gcc/ifcvt.cc b/gcc/ifcvt.cc index 22960a6..b983e87 100644 --- a/gcc/ifcvt.cc +++ b/gcc/ifcvt.cc @@ -1678,10 +1678,10 @@ noce_try_store_flag_mask (struct noce_if_info *if_info) reversep = 0; if ((if_info->a == const0_rtx - && rtx_equal_p (if_info->b, if_info->x)) + && (REG_P (if_info->b) || rtx_equal_p (if_info->b, if_info->x))) || ((reversep = (noce_reversed_cond_code (if_info) != UNKNOWN)) && if_info->b == const0_rtx - && rtx_equal_p (if_info->a, if_info->x))) + && (REG_P (if_info->a) || rtx_equal_p (if_info->a, if_info->x)))) { start_sequence (); target = noce_emit_store_flag (if_info, @@ -1689,7 +1689,7 @@ noce_try_store_flag_mask (struct noce_if_info *if_info) reversep, -1); if (target) target = expand_simple_binop (GET_MODE (if_info->x), AND, - if_info->x, + reversep ? if_info->a : if_info->b, target, if_info->x, 0, OPTAB_WIDEN); |