rampitec wrote: After some digging I believe with this bug fixed we are fine now. Since we are passing all bf16 inputs as i16 we can only inline small integers, and inline integer 1 shall be the same as using 1 in an input register I believe. Although we are missing a potential optimization, say we could fold 'i16 0x3f80' as inline constant 1.0, and a pair of these as 1.0 with opsel should we know this is really a bf16 operand.
https://github.com/llvm/llvm-project/pull/72709 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
