[llvm] [AMDGPU] Using feature to define vmem cacheline size (PR #207884)

via llvm-commits llvm-commits at lists.llvm.org
Wed Jul 8 16:51:41 PDT 2026


================
@@ -2149,6 +2158,7 @@ def FeatureISAVersion12 : FeatureSet<
    FeatureRealTrue16Insts,
    FeatureMTBUFInsts,
    FeatureFormattedMUBUFInsts,
+   FeatureDataCacheLineSize128
----------------
Shoreshen wrote:

Hi @arsenm , the current default is 0. The loop prefetch pass will generate llvm.prefecth if not 0. I will check for pre gfx12 ones.

>From the description of `getCachelineSize` saying that it does not have any use before gfx12, setting the default to 0 will not effect pre-gfx12 target..

https://github.com/llvm/llvm-project/pull/207884


More information about the llvm-commits mailing list