llvm-project
[AMDGPU] In promote-alloca, if index is dynamic, sandwich load with bitcasts to reduce excessive codegen
#171253

Merged

[AMDGPU] In promote-alloca, if index is dynamic, sandwich load with bitcasts to reduce excessive codegen #171253

choikwa merged 13 commits into llvm:main from choikwa:alloca-bitcast

choikwa requested a review from

arsenm 21 days ago

choikwa requested a review from

nhaehnle 21 days ago

choikwa requested a review from

ruiling 21 days ago

choikwa requested a review from

Pierre-vh 21 days ago

choikwa requested a review from

copilot-pull-request-reviewer 21 days ago

llvmbot added backend:AMDGPU

copilot-pull-request-reviewer commented on 2025-12-09

nhaehnle commented on 2025-12-12

[AMDGPU] In promote-alloca, if index is dynamic, sandwich load with b…

55266345

Address feedback, add unaligned testcase. Now checks if VectorTy is m…

8ae99ccd

choikwa force pushed from fffe8614 to 8ae99ccd 15 days ago

nhaehnle commented on 2025-12-15

don't use gep, compare against SubVecTy and Load's alignment

25b622f7

formatting

0ef9d81c

more formatting

9b60c8e3

nhaehnle commented on 2025-12-16

ruiling commented on 2025-12-16

Address feedback, use LShr instead of UDiv, simplify checks

2bd4cd6d

space

a5f1a4b6

ruiling approved these changes on 2025-12-18

ruiling commented on 2025-12-18

choikwa changed the title ~~[AMDGPU] In promote-alloca, if index is dynamic, sandwich load with bitcasts to reduce number of extractelements as they have large expansion in the backend.~~ [AMDGPU] In promote-alloca, if index is dynamic, sandwich load with bitcasts to reduce excessive codegen 12 days ago

add check for pow2 on TotalNumElts, needed for correct shift amt calc…

2eee4b64

fmt

aada9522

add non-pow2 alloca NumElts testcase

a6521f5d

TotalNumElts / NewNumElts = SubVecTy->getNumElts

4fb61fd2

arsenm commented on 2025-12-18

Use TypeSize

dcfba6e9

ruiling approved these changes on 2025-12-19

add tests varying NumElts

6133aed4

choikwa merged 5897f276 into main 10 days ago

Reviewers

ruiling

arsenm

nhaehnle

copilot-pull-request-reviewer

Pierre-vh

Assignees

No one assigned

Labels

backend:AMDGPU

Milestone

No milestone

llvm-project [AMDGPU] In promote-alloca, if index is dynamic, sandwich load with bitcasts to reduce excessive codegen #171253 Merged

[AMDGPU] In promote-alloca, if index is dynamic, sandwich load with bitcasts to reduce excessive codegen #171253

llvm-project
[AMDGPU] In promote-alloca, if index is dynamic, sandwich load with bitcasts to reduce excessive codegen
#171253

Merged