[clang] [llvm] [clang][X86] Return Win64 __int128 indirectly (PR #210358)

via cfe-commits cfe-commits at lists.llvm.org
Tue Nov 3 23:03:00 PST 2026


Andarwinux wrote:

11800H
```
core clock ~ 4.432 GHz (shl+xor chain = 2 cycles)
100000000 calls per run, best of 7 runs

empty call                       (tput)        0 ps/call (  0.00 cyc)
A same-ptr in-place  dep chain   (lat )      908 ps/call (  4.02 cyc)
A same-ptr in-place  8 indep     (tput)      899 ps/call (  3.98 cyc)
A split in/out ping-pong        (lat )       711 ps/call (  3.15 cyc)
A split in/out 8 sinks          (tput)       884 ps/call (  3.92 cyc)
B ld/add-adc/pack indep          (tput)       11 ps/call (  0.05 cyc)
B + caller 16B store dep chain   (lat )     1997 ps/call (  8.85 cyc)
B + caller 2x8B store dep chain  (lat )     1994 ps/call (  8.84 cyc)
B split in/out ping-pong        (lat )      1995 ps/call (  8.84 cyc)

marginal cost of one dependent store->load round trip:
  chain1   1563 ps/call
  chain2   3217 ps/call  (+1654 ps)
  chain4   6260 ps/call  (+3043 ps for 2 more)
  chain8  13430 ps/call  (+7170 ps for 4 more => 1792 ps each)

sink=0  v=7000000000000
A in-place chain (7e8 increments)        got 7000000000000  want 7000000000000  OK
A split ping-pong lo                     got 7000000000000  want 7000000000000  OK
A split ping-pong hi                     got 6999999990000  want 6999999990000  OK
```

Indeed. looks like can't just trust llvm-mca.



https://github.com/llvm/llvm-project/pull/210358


More information about the cfe-commits mailing list