Disable load/store pairing for quads on -mcpu=exynos-m1 as it is not as efficient as the non paired ld/st.
Details
Details
Diff Detail
Diff Detail
- Repository
- rL LLVM
Event Timeline
Comment Actions
You might want to also add a test that shows these type of loads are not clustered by the MI scheduler. You should add the test to test/CodeGen/AArch64/arm64-ldp-cluster.ll
llvm/test/CodeGen/AArch64/arm64-ldp-cluster.ll | ||
---|---|---|
102 ↗ | (On Diff #53418) | Since you only disable clustering for quad load/store, you may want to add check for EXYNOS in above other test cases. |
llvm/test/CodeGen/AArch64/arm64-ldp-cluster.ll | ||
---|---|---|
102 ↗ | (On Diff #53418) | I tend to agree. |
Comment Actions
Added check for EXYNOS in the other test cases (made also the new Exynos test case do the CHECK but can undo it if not seen useful)