[llvm] [AMDGPU] Handle gfx1251 wmma hazard (PR #203145)

[Changpeng Fang via llvm-commits]

================
@@ -2091,11 +2091,27 @@ static bool isCoexecutableVALUInst(const MachineInstr &MI) {
 //
 // Category 3: SWMMAC with Latency 16
 //   SWMMAC_IU8
-static unsigned
-getWMMAHazardInstInCategory(const MachineInstr &MI, const SIInstrInfo *TII,
-                            const TargetSchedModel &SchedModel) {
+//
+// Category 4: 16 Pass GFX1251 WMMA with latency 16
+//   V_WMMA_*_16X16X32_{F16,BF16}
+//   V_WMMA_{F32,F16}_16X16X64_{FP8,BF8}*
+//   V_WMMA_F32_16x16x128_F8F6F4 (F4 only)
+//   V_SWMMAC_*_16X16X64_{F16,BF16}
+//   V_SWMMAC_{F32,F16}_16X16X128_{FP8,BF8}*
+//
+// Category: 32 Pass GFX1251 WMMA with latency 32
----------------
changpeng wrote:

> This should be Category 5, no?

You are right. Will update the comment later

https://github.com/llvm/llvm-project/pull/203145