RFR: 8287373: remove unnecessary paddings in generated code

[Evgeny Astigeevich]

On Thu, 28 Apr 2022 14:46:57 GMT, Boris Ulasevich <bulasevich at openjdk.org> wrote:

> The goal is to remove unnecessary paddings in generated code. The alignment of the [Stub Code] section is determined by the same value as the alignment of the [Entry Point] section: the CodeEntryAlignment parameter with default values 64B on AARCH, and 32B on AMD.
> 
> Large entry alignment values are questionable for entry section. For example, Arm Neoverse N1 Software Optimization Guide recommends to align subroutines to 32B, while static compilers uses an even smaller value of 16B. However, with this change, I suggest to apply different (and smaller) values for [Constants] and [Stub Code] section alignments. This makes overall code 2% smaller on AARCH.
> 
> The correctness of the changes is checked by jtreg. Performance tested by Renaissance and SpecJBB benchmarkds on AARCH and AMD.
> 
> Example. Dummy method disassembly on AARCH, before vs after:
> 
> [Verified Entry Point]                          |  [Verified Entry Point]
>   78c63b80:   nop                               |    7437e480:   nop
>   78c63b84:   sub     x9, sp, #0x20, lsl #12    |    7437e484:   sub     x9, sp, #0x20, lsl #12
>   78c63b88:   str     xzr, [x9]                 |    7437e488:   str     xzr, [x9]
>   78c63b8c:   sub     sp, sp, #0x20             |    7437e48c:   sub     sp, sp, #0x20
>   78c63b90:   stp     x29, x30, [sp, #16]       |    7437e490:   stp     x29, x30, [sp, #16]
>   78c63b94:   orr     w1, wzr, #0x10            |    7437e494:   orr     w1, wzr, #0x10
>   78c63b98:   bl      78343e00                  |    7437e498:   bl      73a61980
>   78c63b9c:   .inst   0x00000000 ; undefined    |    7437e49c:   .inst   0x00000000 ; undefined
>   78c63ba0:   .inst   0x00000000 ; undefined    |
>   78c63ba4:   .inst   0x00000000 ; undefined    |
>   78c63ba8:   .inst   0x00000000 ; undefined    |
>   78c63bac:   .inst   0x00000000 ; undefined    |
>   78c63bb0:   .inst   0x00000000 ; undefined    |
>   78c63bb4:   .inst   0x00000000 ; undefined    |
>   78c63bb8:   .inst   0x00000000 ; undefined    |
>   78c63bbc:   .inst   0x00000000 ; undefined    |
> [Stub Code]                                     |  [Stub Code]
>   78c63bc0:   ldr     x8, 78c63bc8              |    7437e4a0:   ldr     x8, 7437e4a8
>   78c63bc4:   br      x8                        |    7437e4a4:   br      x8
>   78c63bc8:   .inst   0x78343e00 ; undefined    |    7437e4a8:   .inst   0x73a61980 ; undefined
>   78c63bcc:   .inst    ; undefined              |    7437e4ac:   .inst    ; undefined
> [Exception Handler]                             |  [Exception Handler]
>   78c63bd0:   b       783ee080                  |    7437e4b0:   b       73b0c100
> [Deopt Handler Code]                            |  [Deopt Handler Code]
>   78c63bd4:   adr     x30, 78c63bd4             |    7437e4b4:   adr     x30, 7437e4b4
>   78c63bd8:   b       78343ac0                  |    7437e4b8:   b       73a61620
>   78c63bdc:   .inst   0x00000000 ; undefined    |    7437e4bc:   .inst   0x00000000 ; undefined

src/hotspot/cpu/aarch64/globals_aarch64.hpp line 40:

> 38: 
> 39: define_pd_global(uintx, CodeCacheSegmentSize,    64 COMPILER1_AND_COMPILER2_PRESENT(+64)); // Tiered compilation has large code-entry alignment.
> 40: define_pd_global(intx, CodeEntryAlignment,       64);

This change looks reasonable to me. I found the following in the N1 Opt Guide:

Consider aligning subroutine entry points and branch targets to 32B boundaries, within the bounds of the code-density requirements of the program. This will ensure that the subsequent fetch can maximize bandwidth following the taken branch by bringing in all useful instructions

src/hotspot/share/asm/codeBuffer.hpp line 259:

> 257:     // TODO: move InteriorEntryAlignment to common c1/c2 header
> 258:     int code_entry_aliginment = (int) COMPILER2_PRESENT(InteriorEntryAlignment) NOT_COMPILER2(CodeEntryAlignment);
> 259:     return MAX2((int)sizeof(jdouble), code_entry_aliginment);

I don't think the stub code section needs alignment more than 4 bytes. It is not performance critical code.

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

PR: https://git.openjdk.java.net/jdk/pull/8453