-
Enhancement
-
Resolution: Withdrawn
-
P4
-
repo-riscv-port
-
None
-
riscv
-
linux
This patch support RISC-V RVC extension. It can introduce:
* 21% code size reduction in template interpreter generated code
* 20%~25% code size reduction in C1 generated code, evaluated by a common SpringBoot program
* 15%~20% code size reduction in C2 generated code, evaluated by a common SpringBoot program
In my observation, the code size footprint could be reduced to nearly a level of the AArch64 back-end. About the performance, there seems a stable ~0.7% performance improvement on SPECjbb2015 on one HiFive Unleashed board, considering the code density increase. I think the performance aspect might be a speculative behavior on different hardware implementations because C910's performance might be better than that, but HiFive Unleashed may be more general.
Things about this patch:
* If an instruction is compressible, then we will implicitly emit a 16-bit compressed instruction instead of the 32-bit instruction in Assembler.
* About the `_nc` postfix of some of Assembler instructions: we know a bunch of places should be reserved for patching, where we cannot change them into compressed instructions. `_nc` is short for `not compressible` - with this, those instructions should keep their origin 4-byte form and remain uncompressed.
* There are things not easy to compress like MachBranchNodes. Please see the comments in the code - currently this patch does not support this. We will support this improvement in patches coming afterward.
The macros after their expansion might be like:
```
void andr(Register Rd, Register Rs1, Register Rs2) {
+ {
+ Register src = noreg;
+ if (UseRVC && Rs1->is_compressed_valid() && Rs2->is_compressed_valid() &&
+ ((src = Rs1, Rs2 == Rd) || (src = Rs2, Rs1 == Rd))) {
+ and_c(Rd, src);
+ return;
+ }
+ }
unsigned insn = 0;
patch((address)&insn, 6, 0, 0b0110011);
patch((address)&insn, 14, 12, 0b111);
patch((address)&insn, 31, 25, 0b0000000);
patch_reg((address)&insn, 7, Rd);
patch_reg((address)&insn, 15, Rs1);
patch_reg((address)&insn, 20, Rs2);
emit(insn);
};
```
For further information, please see comments in `assembler_riscv_cext.hpp`.
* 21% code size reduction in template interpreter generated code
* 20%~25% code size reduction in C1 generated code, evaluated by a common SpringBoot program
* 15%~20% code size reduction in C2 generated code, evaluated by a common SpringBoot program
In my observation, the code size footprint could be reduced to nearly a level of the AArch64 back-end. About the performance, there seems a stable ~0.7% performance improvement on SPECjbb2015 on one HiFive Unleashed board, considering the code density increase. I think the performance aspect might be a speculative behavior on different hardware implementations because C910's performance might be better than that, but HiFive Unleashed may be more general.
Things about this patch:
* If an instruction is compressible, then we will implicitly emit a 16-bit compressed instruction instead of the 32-bit instruction in Assembler.
* About the `_nc` postfix of some of Assembler instructions: we know a bunch of places should be reserved for patching, where we cannot change them into compressed instructions. `_nc` is short for `not compressible` - with this, those instructions should keep their origin 4-byte form and remain uncompressed.
* There are things not easy to compress like MachBranchNodes. Please see the comments in the code - currently this patch does not support this. We will support this improvement in patches coming afterward.
The macros after their expansion might be like:
```
void andr(Register Rd, Register Rs1, Register Rs2) {
+ {
+ Register src = noreg;
+ if (UseRVC && Rs1->is_compressed_valid() && Rs2->is_compressed_valid() &&
+ ((src = Rs1, Rs2 == Rd) || (src = Rs2, Rs1 == Rd))) {
+ and_c(Rd, src);
+ return;
+ }
+ }
unsigned insn = 0;
patch((address)&insn, 6, 0, 0b0110011);
patch((address)&insn, 14, 12, 0b111);
patch((address)&insn, 31, 25, 0b0000000);
patch_reg((address)&insn, 7, Rd);
patch_reg((address)&insn, 15, Rs1);
patch_reg((address)&insn, 20, Rs2);
emit(insn);
};
```
For further information, please see comments in `assembler_riscv_cext.hpp`.
- relates to
-
JDK-8278994 riscv: RVC support
- Resolved
- links to
-
Review openjdk/riscv-port/24