LLVM Begins Preparing For Intel Sierra Forest & Grand Ridge CPUs
The first of the Sierra Forest / Grand Ridge enablement work landing overnight in LLVM 16 is adding WRMSRNS instructions and MSRLIST instructions -- both new to Sierra Forest and Grand Ridge.
The MSRLIST instructions are RDMSRLIST for reading a list of model specific registers and WRMSRLIST for writing a list of model specific registers. These new RDMSRLIST/WRMSRLIST instructions allow supplying a list of up to 64 model specific registers at a time by this instruction compared to RDMSR/WRMSR instructions that are an MSR at a time.
Also added so far to LLVM is WRMSRNS. The WRMSRNS with Grand Ridge and Sierra Forest is a non-serializing write to a model specific register. Rather than WRMSR being a serialized instruction, WRMSRNS is a non-serializing write but otherwise behaves the same.
Last month's Intel programming reference manual update also noted several more instructions coming with Sierra Forest and Ridge Ridge, including CMPCCXADD, AVX-IFMA, AVX-NE-CONVERT, and AVX-VNNI-INT8. So far those additional instructions haven't landed in LLVM Git nor the "-march=sierraforest" target but given Intel's usual punctuality around the compiler patches they will probably come soon. These parts likely won't be out until 2024 meanwhile LLVM Git has yet to see any AMD Zen 4 (znver4) target and tuned support added for these already shipping processors.