projects / platform / upstream / llvm.git / commit
? search:
summary | shortlog | log | commit | commitdiff | tree
(parent: d7e6575) | patch
[X86][Atom] Fix vector fp<->int resource/throughputs
authorSimon Pilgrim <llvm-dev@redking.me.uk>
Wed, 7 Jul 2021 14:27:14 +0000 (15:27 +0100)
committerSimon Pilgrim <llvm-dev@redking.me.uk>
Wed, 7 Jul 2021 15:52:34 +0000 (16:52 +0100)
commitded8866f4a4600c69f3d1d6f5e0c9ec7bd9ade45
treee18d7992ec22aa504ab269ec68d857b2c4d2698btree | snapshot
parentd7e65757ed8f2dfe17b71cf169a3c83a155d1c2bcommit | diff
[X86][Atom] Fix vector fp<->int resource/throughputs

Match whats documented in the Intel AOM - almost all the conversion instructions requires BOTH ports (apart from the MMX cvtpi2ps/cvtpi2ps instructions which we already override) - this was being incorrectly modelled as EITHER port.

Now that we can use in-order models in llvm-mca, the atom model is a good "worst case scenario" analysis for x86.
llvm/lib/Target/X86/X86ScheduleAtom.td diff | blob | history
llvm/test/tools/llvm-mca/X86/Atom/resources-sse1.s diff | blob | history
llvm/test/tools/llvm-mca/X86/Atom/resources-sse2.s diff | blob | history
Domain: System / Toolchain;