[vectorIntrinsics] RFR: 8283709: Add x86 back-end implementation for bit BIT_COUNT operation [v9]

[Sandhya Viswanathan]

On Mon, 11 Apr 2022 10:58:39 GMT, Jatin Bhateja <jbhateja at openjdk.org> wrote:

>> Summary of changes:
>> 
>> - Patch re-uses existing C2 IR nodes and re-organizes LUT based JIT code sequence of VectorOperations.BIT_COUNT operation 
>>   for sub-word type (BYTE, SHORT) vectors over X86 targets supporting AVA2 and AVX512 features.
>> - Efficient single instruction POPCOUNT instruction is emitted for applicable targets. 
>> 
>> Kindly review and share you feedback.
>> 
>> Best Regards,
>> Jatin
>
> Jatin Bhateja has updated the pull request incrementally with one additional commit since the last revision:
> 
>   8283709: Review comments resolved.

Please make the changes suggested here. No need for re-review.

src/hotspot/cpu/x86/c2_MacroAssembler_x86.cpp line 4541:

> 4539:                                                       KRegister mask, bool merge, int vec_enc) {
> 4540:   assert(VM_Version::supports_avx512vl() || vec_enc == Assembler::AVX_512bit, "");
> 4541:   assert(UsePopCountInstruction, "");

Please remove this assert on UsePopCountInstruction.

src/hotspot/cpu/x86/stubGenerator_x86_32.cpp line 4127:

> 4125:     StubRoutines::x86::_vector_reverse_byte_perm_mask_short = generate_vector_reverse_byte_perm_mask_short("perm_mask_short");
> 4126: 
> 4127:     if (VM_Version::supports_avx2() && (!VM_Version::supports_avx512_vpopcntdq() || !UsePopCountInstruction)) {

Please remove the reference to UsePopCountInstruction here.

src/hotspot/cpu/x86/stubGenerator_x86_64.cpp line 7795:

> 7793:     StubRoutines::x86::_vector_reverse_byte_perm_mask_short = generate_vector_reverse_byte_perm_mask_short("perm_mask_short");
> 7794: 
> 7795:     if (VM_Version::supports_avx2() && (!VM_Version::supports_avx512_vpopcntdq() || !UsePopCountInstruction)) {

Please remove the reference to UsePopCountInstruction here.

src/hotspot/cpu/x86/vm_version_x86.cpp line 1705:

> 1703: 
> 1704:   // Use population count instruction if available.
> 1705:   if (supports_popcnt() || supports_avx512_vpopcntdq() || supports_avx512_bitalg()) {

The UsePopCountInstruction is only for scalar popcont. Extending it to Vector popcount is causing lot of confustion. Let us keep it for scalar only as below:
if (supports_popcnt()) {
 ...
}

src/hotspot/cpu/x86/x86.ad line 8635:

> 8633:               VM_Version::supports_avx512_bitalg()) ||
> 8634:              (is_non_subword_integral_type(Matcher::vector_element_basic_type(n->in(1))) &&
> 8635:               VM_Version::supports_avx512_vpopcntdq())));

Please replace by:
predicate(is_pop_count_instr_target(Matcher::vector_element_basic_type(n->in(1))));
Also no need to check for UsePopCountInstruction any where as it is only meant for scalar popcount.

src/hotspot/cpu/x86/x86.ad line 8660:

> 8658:               VM_Version::supports_avx512_bitalg()) ||
> 8659:              (is_non_subword_integral_type(Matcher::vector_element_basic_type(n->in(1))) &&
> 8660:               VM_Version::supports_avx512_vpopcntdq())));

Could be replaced by:
predicate(is_pop_count_instr_target(Matcher::vector_element_basic_type(n->in(1))));
Also no need to check for UsePopCountInstruction any where as it is only meant for scalar popcount.

src/hotspot/cpu/x86/x86.ad line 8677:

> 8675:   predicate(!UsePopCountInstruction ||
> 8676:             (!VM_Version::supports_avx512_vpopcntdq() && Matcher::vector_element_basic_type(n->in(1)) == T_INT) ||
> 8677:             (!VM_Version::supports_avx512_bitalg() && is_subword_type(Matcher::vector_element_basic_type(n->in(1)))));

Could be replaced by:
predicate((Matcher::vector_element_basic_type(n->in(1)) != T_LONG) 
 && !is_pop_count_instr_target(Matcher::vector_element_basic_type(n->in(1))));

src/hotspot/cpu/x86/x86.ad line 8691:

> 8689: 
> 8690: instruct vpopcountL_avx_reg(vec dst, vec src, vec xtmp1, vec xtmp2, vec xtmp3, rRegP rtmp) %{
> 8691:   predicate(!UsePopCountInstruction || !VM_Version::supports_avx512_vpopcntdq());

Please remove the reference to UsePopCountInstruction here.

-------------

Marked as reviewed by sviswanathan (Committer).

PR: https://git.openjdk.java.net/panama-vector/pull/185