RFR: 8309502: RISC-V: String.indexOf intrinsic may produce misaligned memory loads

[Vladimir Kempik]

On Mon, 5 Jun 2023 20:52:01 GMT, Vladimir Kempik <vkempik at openjdk.org> wrote:

> Please review this attempt to remove misaligned loads in String.indexOf intrinsic on RISC-V
> 
> Initialy found these misaligned loads when profiling finagle-http test from renaissance suite.
> The majority of trp_lam events (about 66k per finagle-http round) came at line 706 (https://github.com/openjdk/jdk/pull/14320/files#diff-35eb1d2f1e2f0514dd46bd7fbad49ff2c87703d5a3041a6433956df00a3fe6e6L706)
> The other two produced about 100 events combined.
> Later I've found this can partially be reproduced with StringIndexOf.advancedWithMediumSub.
> Numbers on hifive before and after applying the patch:
> 
> 
> Benchmark                                                  Mode  Cnt       Score      Error  Units
> StringIndexOf.advancedWithMediumSub                        avgt   25   47031.406 ±  144.005  ns/op
> 
> 
> After:
> 
> Benchmark                                                 Mode  Cnt       Score     Error  Units
> StringIndexOf.advancedWithMediumSub                       avgt   25    4256.830 ±  23.075  ns/op
> 
> 
> Testing: tier1/tier2 is clean on hifive.

See, this part of algo is misaligned

 bind(CH1_LOOP);
    add(tmp3, haystack, hlen_neg);
    (this->*load_2chr)(ch2, Address(tmp3), noreg);
    beq(ch1, ch2, MATCH);
    add(hlen_neg, hlen_neg, haystack_chr_size);
    blez(hlen_neg, CH1_LOOP);

this becomes:

CH1_LOOP: add t4, a1, a2
lhu t1, 0(t4)
beg t0, t1, 0xMATCH
addi a2, a2, 1
blez a2, CH1_LOOP


so we load halfword on each iteration from address (a1+a2), and while a1 is constant, a2 is incrementing by 1 each step, so every other load is misaligned.

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

PR Comment: https://git.openjdk.org/jdk/pull/14320#issuecomment-1580138574