Re: [PING][PATCH] RISC-V: Standardize formatting of SFB ALU conditional move

["Maciej W. Rozycki" <macro@embecosm.com>]

On Thu, 18 Aug 2022, Kito Cheng wrote:

> OK, thanks for tweaking this!

 Committed now, thanks for your review!

 Would you mind sharing your opinion on my previous observation here:
<https://gcc.gnu.org/pipermail/gcc-patches/2022-July/598542.html>?

 I have since realised we have a `-mbranch-cost=' option letting the user 
set the threshold for choosing branches over alternative code sequences, 
so my concern is valid even for our tree unchanged and without the change 
just committed here applied.  Consequently the test case may fail.

 E.g. with:

RUNTESTFLAGS="--target_board remote-unmatched/-mbranch-cost=1 riscv.exp=pr105314.c"

I get:

PASS: gcc.target/riscv/pr105314.c   -O0  (test for excess errors)
FAIL: gcc.target/riscv/pr105314.c   -O0   scan-assembler-not \tbeq\t
PASS: gcc.target/riscv/pr105314.c   -O1  (test for excess errors)
FAIL: gcc.target/riscv/pr105314.c   -O1   scan-assembler-not \tbeq\t
[...]

		=== gcc Summary ===

# of expected passes		9
# of unexpected failures	9

because GCC legitimately chooses to emit branches as less costly in this 
case.

 I think we need to pacify the test case somehow if it does not match the 
criteria for PR105314, either by excluding the case from testing in that 
situation or by forcing it via command-line options to make it match the 
criteria (or indeed by verifying a branch is produced regardless).  Sadly 
Jakub chose not to chime in and it's not clear to me which approach would 
be the most appropriate here.

  Maciej