Alexander Monakov / is on Nostr

"9c38f8c7885b463ac517bc05d56edc21f921d4ad2ca8c16437a4df31b697b991f1e31aa510dbb1cab67bd19e38a108fbe13600ef69630983061ee5296d1d8cff"

Last Notes

2024-08-31 09:04:00

- reply

@npub18du…tnc7 those are atomic accesses, not plain; they receive required memory ordering as an optional parameter that defaults to seq_cst

2024-08-31 08:16:18

I'm almost ashamed to ask, but, anyway, if in C11 or C++11 you want plain loads/stores on a variable declared with an atomic type, can you do that? How?

2024-06-17 19:14:06

- reply

@npub1a3h…m44m this is approximately how far IEEE double representation of 1.5*π is off from the true value of 1.5*π

2024-01-05 08:24:22

- reply

@npub1ar8…rmzu and yet sometimes a quad-long is okay with g++ if you twist its arm in exactly the right way:
https://cdn.masto.host/mastodongamedevplace/media_attachments/files/111/702/374/026/464/618/original/08ac55ca9343213e.png

2023-12-01 21:26:43

- reply

@npub1pu3…8r76 and if ic_fetch_stall.ic_stall_back_pressure can be trusted, it's the backend, not frontend (the increase in this counter matches the cycle count difference)

2023-12-01 21:20:59

- reply

@npub1pu3…8r76 Zen 2 doesn't fuse this cmp imm,disp(*)+jne combo, and I can confirm with ex_ret_fus_brnch_inst and ex_ret_cops PMU counters

2023-12-01 19:54:43

Just when I thought Zen 2 uarch is mature enough that a 35% perf swing will be for an easily discernible reason and not some crazy complex microarchitectural bullshit :]
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112697
https://cdn.masto.host/mastodongamedevplace/media_attachments/files/111/506/908/101/538/254/original/a6f84d5b522e6e51.png

2023-04-18 09:16:28

A colleague's computer is glitching in a peculiar way: under load, or especially as it's cooling down after a CPU load, interrupts for incoming Ethernet packets don't fire. Happens with both on-board and a PCIe NIC, but not with a USB NIC. I think this hints at a hardware issue, but is there a way to check if the hardware interrupt line glitches out, or Linux leaves the interrupt disabled for too long?
(hmmm, doesn't USB Ethernet work via polling?)