[llvm] [AMDGPU] Preserve call info when lowering SI_CALL_ISEL (PR #192200)

[Shilei Tian via llvm-commits]

================
@@ -0,0 +1,15 @@
+; RUN: llc -emit-call-site-info -stop-after=dead-mi-elimination -mtriple=amdgcn-- -mcpu=gfx900 -o - %s | FileCheck %s
+
+; CHECK-LABEL: name:            basic_call
+; CHECK: $sgpr30_sgpr31 = SI_CALL {{.*}}, @foo, csr_amdgpu
+
+define i32 @basic_call(i32 %src) #0 {
+  %token = call token @llvm.experimental.convergence.entry()
+  %result = call i32 @foo(i32 %src) [ "convergencectrl"(token %token) ]
+  ret i32 %result
+}
+
+declare i32 @foo(i32) #0
+declare token @llvm.experimental.convergence.entry()
+
+attributes #0 = { convergent nounwind readnone }
----------------
shiltian wrote:

trim the attribute if it is necessary, or just drop it if not?

https://github.com/llvm/llvm-project/pull/192200