Performance degradation when changing the "iovs_len" from 1 to 0 in "fd_write" calling

[hungryzzz]

Symptom Description

Hi, I ran the good.wasm and bad.wasm attached in the code.zip, and surprisingly found that when I change the third parameter (iovs_len) of fd_write function call from 1 to 0, the execution time of Wamr became 4x longer. But intuitively, when copying a 0 length data to the fd, the execution time may be shorter.

The execution times in the WasmEdge (AOT) and Wamr(AOT) of both good.wasm and bad.wasm are as follows (measured by time tool):

For good.wasm:

• WasmEdge: 0.62s
• Wamr: 0.65s

For bad.wasm:

• WasmEdge: 0.62s
• Wamr: 2.56s

Repreduce Steps

# run on Wamr
wamrc -o $filename-wamr-aot.wasm $filename.wasm
iwasm $filename-wamr-aot.wasm > case.out

# run on WasmEdge
wasmedgec $filename.wasm $filename-wasmedge-aot.wasm
wasmedge $filename-wasmedge-aot.wasm > case.out

Environment

• Wamr commit: ee056d8
• Operating system: Linux ringzzz-OptiPlex-Micro-Plus-7010 6.5.0-18-generic
• Architecture: Intel(R) Core(TM) i5-13500

code.zip

[yamt]

i guess it's because of the bh_log call

wasm-micro-runtime/core/iwasm/common/wasm_memory.c

Lines 615 to 616 in ee056d8

	LOG_WARNING(
	"wasm_runtime_malloc failed: memory hasn't been initialized.\n");

can you try #4452?

[hungryzzz]

Hi, I have tried #4452, and the performance degradation disappeared.
Thank you very much!

[yamt]

Hi, I have tried #4452, and the performance degradation disappeared. Thank you very much!

thank you for the confirmation.