[llvm] [AArch64] Keep floating-point conversion in SIMD (PR #147707)

[David Sherwood via llvm-commits]

================
@@ -6598,6 +6598,18 @@ def : Pat<(f64 (AArch64frsqrts (f64 FPR64:$Rn), (f64 FPR64:$Rm))),
 def : Pat<(v2f64 (AArch64frsqrts (v2f64 FPR128:$Rn), (v2f64 FPR128:$Rm))),
           (FRSQRTSv2f64 FPR128:$Rn, FPR128:$Rm)>;
 
+let HasOneUse = 1 in {
+def fp_to_uint_oneuse : PatFrag<(ops node:$src0), (fp_to_uint $src0)>;
+def fp_to_sint_oneuse : PatFrag<(ops node:$src0), (fp_to_sint $src0)>;
+}
+
+class StoreMaybeAssertZext<PatFrag op> : PatFrags<(ops node:$val, node:$ptr),
----------------
david-arm wrote:

Given this is specifically related to truncating stores, perhaps worth renaming to TruncStoreMaybeAssertZext?

Also, is it possible to add one for assertsext? I don't know if that comes up in practice.

https://github.com/llvm/llvm-project/pull/147707