On Sun, Jan 1, 2023 at 5:53 PM Roger Sayle <ro...@nextmovesoftware.com> wrote:
>
>
> This patch addresses PR target/108229, which is a change in code
> generation during the STV pass, due to the recently approved patch
> to handle vec_select (reductions) in the vector unit. The recent
> change is innocent, but exposes a latent STV "gain" calculation issue
> that is benign (or closely balanced) on most microarchitectures.
>
> The issue is when STV considers converting PLUS with a MEM operand.
>
> On TARGET_64BIT (m=1):
> addq 24(%rdi), %rdx // 4 bytes
> or with -m32 (m=2)
> addl 24(%esi), %eax // 3 bytes
> adcl 28(%esi), %edx // 3 bytes
> is being converted by STV to
> vmovq 24(%rdi), %xmm5 // 5 bytes
> vpaddq %xmm5, %xmm4, %xmm4 // 4 bytes
>
> The current code in general_scalar_chain::compute_convert_gain
> considers that scalar unit addition is replaced with a vector
> unit addition (usually about the same cost), but doesn't consider
> anything special about MEM operands, assuming that a scalar load
> gains/costs nothing compared to a vector load. We can allow the
> backend slightly better fine tuning by including in the gain
> calculation that m scalar loads are being replaced by one vector
> load, and when optimizing for size including that we're increasing
> code size (e.g. an extra vmovq instruction for a MEM operand).
>
> This patch is a win on the CSiBE benchmark when compiled with -Os.
> Alas I've no new testcase as this is extremely sensitive to the
> backend microarchitecture parameterization (and it's dangerous to
> select parameters from the N=1 statistics of a single bugzilla PR).
>
> This patch has been tested on x86_64-pc-linux-gnu with make bootstrap
> and make -k check, both with and without --target_board=unix{-m32},
> with no new failures. Ok for mainline?
>
>
> 2023-01-01 Roger Sayle <ro...@nextmovesoftware.com>
>
> gcc/ChangeLog
> PR target/108229
> * config/i386/i386-features.cc
> (general_scalar_chain::compute_convert_gain) <case PLUS>: Consider
> the gain/cost of converting a MEM operand.
LGTM.
Thanks,
Uros.
>
> Thanks in advance,
> Roger
> --
>
