RE: [PATCH] [arm] adjust expectations for armv8_2-fp16-move-[12].c

[Kyrylo Tkachov via Gcc-patches]

> -----Original Message-----
> From: Alexandre Oliva <ol...@adacore.com>
> Sent: Friday, February 17, 2023 7:12 AM
> To: gcc-patches@gcc.gnu.org
> Cc: ni...@redhat.com; Richard Earnshaw <richard.earns...@arm.com>
> Subject: [PATCH] [arm] adjust expectations for armv8_2-fp16-move-[12].c
> 
> 
> Commit 3a7ba8fd0cda387809e4902328af2473662b6a4a, a patch for
> tree-ssa-sink, enabled the removal of basic blocks in ways that
> affected the generated code for both of these tests, deviating from
> the expectations of the tests.
> 
> The simplest case is that of -2, in which the edge unsplitting ends up
> enabling a conditional return rather than a conditional branch to a
> set-and-return block.  That looks like an improvement to me, but the
> condition in which the branch or the return takes place can be
> reasonably reversed (and, with the current code, it is), I've relaxed
> the pattern in the test so as to accept reversed and unreversed
> conditions applied to return or branch opcodes.
> 
> The situation in -1 is a little more elaborate: conditional branches
> based on FP compares in test_select_[78] are initially expanded with
> CCFPE compare-and-cbranch on G{T,E}, but when ce2 turns those into a
> cmove, because now we have a different fallthrough block, the
> condition is reversed, and that lands us with a compare-and-cmove
> sequence that needs CCFP for UNL{E,T}.  The insn output reverses the
> condition and swaps the cmove input operands, so the vcmp and vsel
> insns come out the same except for the missing 'e' (for the compare
> mode) in vcmp, so, since such reversals could have happened to any of
> the tests depending on legitimate basic block layout, I've combined
> the vcmp and vcmpe counts.
> 
> I see room for improving cmove sequence generation, e.g. trying direct
> and reversed conditions and selecting the cheapest one (which would
> require CCFP conditions to be modeled as more expensive than CCFPE),
> or for some other machine-specific (peephole2?) optimization to turn
> CCFP-requiring compare and cmove into CCFPE compare and swapped-
> inputs
> cmove, but I haven't tried that.
> 
> Regstrapped on x86_64-linux-gnu.
> Tested on arm-vxworks7 (gcc-12) and arm-eabi (trunk).  Ok to install?

The changes in the patch are okay for now. We can look at other improvements 
separately.
Thanks,
Kyrill

> 
> for  gcc/testsuite/ChangeLog
> 
>       * gcc.target/arm/armv8_2-fp16-move-1.c: Combine vcmp and vcmpe
>       expected counts into a single pattern.
>       * gcc.target/arm/armv8_2-fp16-move-2.c: Accept conditional
>       return and reversed conditions.
> ---
>  gcc/testsuite/gcc.target/arm/armv8_2-fp16-move-1.c |    3 +--
>  gcc/testsuite/gcc.target/arm/armv8_2-fp16-move-2.c |    2 +-
>  2 files changed, 2 insertions(+), 3 deletions(-)
> 
> diff --git a/gcc/testsuite/gcc.target/arm/armv8_2-fp16-move-1.c
> b/gcc/testsuite/gcc.target/arm/armv8_2-fp16-move-1.c
> index 009bb8d1575a4..444c4a3353555 100644
> --- a/gcc/testsuite/gcc.target/arm/armv8_2-fp16-move-1.c
> +++ b/gcc/testsuite/gcc.target/arm/armv8_2-fp16-move-1.c
> @@ -196,5 +196,4 @@ test_compare_5 (__fp16 a, __fp16 b)
>  /* { dg-final { scan-assembler-not {vcmp\.f16} } }  */
>  /* { dg-final { scan-assembler-not {vcmpe\.f16} } }  */
> 
> -/* { dg-final { scan-assembler-times {vcmp\.f32} 4 } }  */
> -/* { dg-final { scan-assembler-times {vcmpe\.f32} 8 } }  */
> +/* { dg-final { scan-assembler-times {vcmpe?\.f32} 12 } }  */
> diff --git a/gcc/testsuite/gcc.target/arm/armv8_2-fp16-move-2.c
> b/gcc/testsuite/gcc.target/arm/armv8_2-fp16-move-2.c
> index fcb857f29ff15..dff57ac8147c2 100644
> --- a/gcc/testsuite/gcc.target/arm/armv8_2-fp16-move-2.c
> +++ b/gcc/testsuite/gcc.target/arm/armv8_2-fp16-move-2.c
> @@ -8,4 +8,4 @@ test_select (__fp16 a, __fp16 b, __fp16 c)
>  {
>    return (a < b) ? b : c;
>  }
> -/* { dg-final { scan-assembler "bmi" } } */
> +/* { dg-final { scan-assembler "bx?(mi|pl)" } } */
> 
> --
> Alexandre Oliva, happy hacker                https://FSFLA.org/blogs/lxo/
>    Free Software Activist                       GNU Toolchain Engineer
> Disinformation flourishes because many people care deeply about injustice
> but very few check the facts.  Ask me about <https://stallmansupport.org>