tree 27f6dbd5c1b01b73f2ed004cd6fc0f8d94d6ae06
parent 97c9020805824ce88f36ce66c80489b2df5db74a
author Quinn Dawkins <quinn.dawkins@gmail.com> 1764774175 -0500
committer GitHub <noreply@github.com> 1764774175 +0000
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsFcBAABCAAQBQJpMFEfCRC1aQ7uu5UhlAAA4HQQAIyvnRp9m1vs0SZwuZU5j8ue
 nl9R1ruYCwb0coYMWarvLNy3EM8EYo7NBIcLLrov/ZOEQZaSRcecja+vaH6IZ2lO
 Gh6Jwe68KqQ1HSTB0CyNLIy8gg2Hry2ZAof1LrKIYktMTzNSwlK3tZnT2RoIcdkF
 tcuD5wyur4fUVQ4XvjgWb8WbGSaInyCl2gEPK0OUtgIqOxQTd/erzrSM85K3aepp
 jCWJ1iDc7qUVnr7LjWTwbnunR6qnx4rRGpj0oJdNImH9pjPdFm+kKYcWp+YsQOVo
 vxs4mKtO24GszStDGpDK1mOE5jHSdTDJTcvp2CbWH+98B6sv+McLtJlbrGyC+x4I
 ay0/1Ujkh8WMeKugM8h/biXrIXUATvK+X2Y3YIxFC/T+IiclunGwOaYwlgm1Taby
 ybfaqLYd412H3cAEf8E0UOaMIyaLzXyG555ZTkZzsHnVpPeoKF28M73fFx+E9MZa
 m1k+jOirOemPEVW4LPHoSc2opgTulkiyl/KO61VYlsXnLMEnxoeudVQjuQYsIRgk
 zk/Kr2IZrF37KU/IXoxJ3mffn3kQaewIYydeDckS6WN8HZepvszoR2+bEILMxXG/
 sj+Qe2sWLSbhxMgWI+lnbeWYXc2E9oaHVsRwlCNYTlSE2SWxnIUexiKugIyJJEb4
 JpDJAq/9T7379aOvXQqB
 =kMZk
 -----END PGP SIGNATURE-----
 

[Codegen] Support dynamic offsets in collapse_shape fusion to interface stores (#22800)

Previously, `FoldCollapseShapeIntoInterfaceTensorStore` only handled
static offsets when fusing `tensor.collapse_shape` into dispatch tensor
stores. This change extends the pattern to also accept dynamic offsets
when:

1. The offset comes from an `affine.apply` operation
2. The affine expression is provably divisible by the inner dimension
product

This also fixes an issue where offset computation was done at the
subspan insertion point, which could fail when the offset value is
defined in a nested region (e.g., inside a loop). The offset
affine.apply is now created at the store's insertion point where
dominance is guaranteed.