RFR: 8351950: C2: masked vector MIN/MAX AVX512: SIGFPE / no valid evex tuple_table entry

[Jatin Bhateja]

On Tue, 6 May 2025 16:40:05 GMT, Emanuel Peter <epeter at openjdk.org> wrote:

>> PR adds missing EVEX compressed displacement attributes used for computing the scale factor (N) of compressed displacement.
>> AVX512 memory operand instructions use compressed disp8 encoding if the displacement is a multiple of scale (N), which depends on Vector Length, embedded broadcasting, and lane size.  Please refer to section 2.7.5 of Intel SDM for more details.
>> 
>> e.g., Consider two instructions, one with displacement 0x10203040 and the other with displacement 0x40, instruction operates over full 64-byte vector hence scale N = 64. Displacement of latter instruction is a multiple of scale, thus can be represented by 1 byte displacement encoding, while the former requires 4 bytes to represent displacement in instruction encoding.
>> 
>> 
>> 1) vpternlogq $0xff,0x10203040(%r20,%r21,8),%zmm23,%zmm24
>>     EVEX        OP   MR   SIB       DISP       IMM
>> --------------|----|----|----|---------------|-----|
>> 62 6b c1 40     25   84   ec     40 30 20 10     ff
>> 
>> 2) vpternlogq $0xff,0x40(%r20,%r21,8),%zmm23,%zmm24
>> For full vector width operation, scalar matches with vector size, hence scale N = 64
>> effective displacement / compressed DISP8 = OFFSET(64) / 64 = 0x1 
>>     EVEX       OP   MR SIB    DISP     IMM
>> -------------|----|---|---|-----------|---|
>> 62 6b c1 40    25  44   ec      01     ff 
>> 
>> 
>> Kindly review and share your feedback.
>> 
>> Best Regards,
>> Jatin
>
> src/hotspot/cpu/x86/assembler_x86.cpp line 11542:
> 
>> 11540:   assert(vector_len == AVX_512bit || VM_Version::supports_avx512vl(), "");
>> 11541:   InstructionAttr attributes(vector_len, /* vex_w */ true,/* legacy_mode */ false, /* no_mask_reg */ false,/* uses_vl */ true);
>> 11542:   attributes.set_address_attributes(/* tuple_type */ EVEX_FV,/* input_size_in_bits */ EVEX_NObit);
> 
> @jatin-bhateja How is this `fma` case related to the `min / max` cases that were reported? I did also not find a test below.

Hi @eme64 , For tuple_type Fully Vector (FV) scale factor (N) does not take into account the lane size, thus EVEX_NObit is right argument here, using EVEX_32bit will not cause functional correctness as lane size is anyways ignored, but EVEX_NObit better conveys our intent.

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

PR Review Comment: https://git.openjdk.org/jdk/pull/25021#discussion_r2076011064