RFR: 8320379: C2: Sort spilling/unspilling sequence for better ld/st merging into ldp/stp on AArch64

Andrew Haley aph at openjdk.org
Tue Nov 21 10:27:59 UTC 2023


On Tue, 21 Nov 2023 07:15:15 GMT, Fei Gao <fgao at openjdk.org> wrote:

> Macro-assembler on aarch64 can merge adjacent loads or stores into ldp/stp.[[1]](https://github.com/openjdk/jdk/blob/a95062b39a431b4937ab6e9e73de4d2b8ea1ac49/src/hotspot/cpu/aarch64/macroAssembler_aarch64.cpp#L2079)
> 
> For example, it can merge:
> 
> str     w20, [sp, #16]
> str     w10, [sp, #20]
> 
> into
> 
> stp     w20, w10, [sp, #16]
> 
> 
> But C2 may generate a sequence like:
> 
> str     x21, [sp, #8]
> str     w20, [sp, #16]
> str     x19, [sp, #24] <---
> str     w10, [sp, #20] <--- Before sorting
> str     x11, [sp, #40]
> str     w13, [sp, #48]
> str     x16, [sp, #56]
> 
> We can't do any merging for non-adjacent loads or stores.
> 
> The patch is to sort the spilling or unspilling sequence in the order of offset during instruction scheduling and bundling phase. After that, we can get a new sequence:
> 
> str     x21, [sp, #8]
> str     w20, [sp, #16]
> str     w10, [sp, #20] <---
> str     x19, [sp, #24] <--- After sorting
> str     x11, [sp, #40]
> str     w13, [sp, #48]
> str     x16, [sp, #56]
> 
> 
> Then macro-assembler can do ld/st merging:
> 
> str     x21, [sp, #8]
> stp     w20, w10, [sp, #16] <--- Merged
> str     x19, [sp, #24]
> str     x11, [sp, #40]
> str     w13, [sp, #48]
> str     x16, [sp, #56]
> 
> 
> To justify the patch, we run `HelloWorld.java`
> 
> public class HelloWorld {
>     public static void main(String [] args) {
>         System.out.println("Hello World!");
>     }
> }
> 
> with `java -Xcomp -XX:-TieredCompilation HelloWorld`.
> 
> Before the patch, macro-assembler can do ld/st merging for 3688 times. After the patch, the number of ld/st merging increases to 3871 times, by ~5 %.
> 
> Tested tier1~3 on x86 and AArch64.

This is a good idea, although the real-world gains are small. I'd wonder if this was worth doing for non-AArch64 ports, although even on others sorting the accesses into order might help.

src/hotspot/share/opto/output.cpp line 2300:

> 2298:   }
> 2299: 
> 2300:   return false;

Suggestion:

  return 0; // Not comparable.

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

PR Comment: https://git.openjdk.org/jdk/pull/16754#issuecomment-1820636451
PR Review Comment: https://git.openjdk.org/jdk/pull/16754#discussion_r1400350878


More information about the hotspot-compiler-dev mailing list