Optimized dataset read (#211)

* Optimized dataset read

There was a false dependency on readReg2 and readReg3 (caused by `xor rbp, rax` instruction) when reading dataset item (see design.md - 4.6.2 Loop execution, steps 5 and 7). This change uses `ma` register to read dataset item before the whole `rbp` (`ma` and `mx`) is changed, so superscalar and out-of-order CPU can start executing it earlier.

Results: https://i.imgur.com/Bpeq9mx.png

~1% speedup on modern Intel/AMD CPUs.

* ARMv8: optimized dataset read

Break dependency from readReg2 and readReg3.

* Fixed light mode hashing
This commit is contained in:
SChernykh
2021-05-22 13:54:50 +02:00
committed by GitHub
parent c12097400b
commit 3c8c7ee097
5 changed files with 20 additions and 18 deletions
+3 -2
View File
@@ -307,6 +307,9 @@ literal_v14: .fill 2,8,0
literal_v15: .fill 2,8,0
DECL(randomx_program_aarch64_vm_instructions_end):
# Calculate dataset pointer for dataset read
# Do it here to break false dependency from readReg2 and readReg3 (see next line)
lsr x10, x9, 32
# mx ^= r[readReg2] ^ r[readReg3];
eor x9, x9, x18
@@ -324,8 +327,6 @@ DECL(randomx_program_aarch64_cacheline_align_mask1):
# mx <-> ma
ror x9, x9, 32
# Calculate dataset pointer for dataset read
mov w10, w9
DECL(randomx_program_aarch64_cacheline_align_mask2):
# Actual mask will be inserted by JIT compiler
and x10, x10, 1