[llvm] ISel/AArch64/SVE: custom lower vector ISD::[L]LRINT (PR #89035)

[David Sherwood via llvm-commits]

================
@@ -0,0 +1,492 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=aarch64 -mattr=+sve | FileCheck %s
+
+define <vscale x 1 x i64> @lrint_v1f16(<vscale x 1 x half> %x) {
+; CHECK-LABEL: lrint_v1f16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.d
+; CHECK-NEXT:    fcvtzs z0.d, p0/m, z0.h
----------------
david-arm wrote:

I don't think this looks right because the man page for the lrint libm routine says:

`The lrint() and llrint() functions return the integral value nearest to x (according to the prevailing rounding mode)`

but fcvtzs always rounds to zero. I think it needs to be a sequence of frintx, fcvts like the scalar and NEON variants.

https://github.com/llvm/llvm-project/pull/89035