[llvm] [AMDGPU] precision error observed after SIPeepholeSDWA optimization for fp16 (PR #109395)

[Matt Arsenault via llvm-commits]

================
@@ -0,0 +1,265 @@
+%struct.rocfft_complex = type { half, half }
----------------
arsenm wrote:

This test is way too big, and is missing run lines 

https://github.com/llvm/llvm-project/pull/109395