IPRA: Don't assume called function is first call operand

[arsenm]

Summary

Fixes not finding the called global for AMDGPU
call pseudoinstructions. The call instruction can't
directly refer to a global, so we use a pseudo mirroring
the actual instruction's operands with an extra operand
appended to refer to the function.

[arsenm]

Use a dyn_cast. It might be possible to see a call to something that isn't directly a Function

[arsenm]

Add tests

[MatzeB]

Looks like an incomplete patch (I only see a changed testcase).

[MatzeB]

I guess the sourcecode changes are in the earlier revision, they LGTM (I'd tend to remove the TODO comment).

[arsenm]

r311637

test/CodeGen/AMDGPU/ipra.ll

-; RUN: llc -mtriple=amdgcn-amd-amdhsa -verify-machineinstrs -enable-ipra < %s | FileCheck -check-prefix=GCN %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -verify-machineinstrs -enable-ipra -amdgpu-sroa=0 < %s | FileCheck -check-prefix=GCN %s
 
 ; Kernels are not called, so there is no call preserved mask.
 ; GCN-LABEL: {{^}}kernel:
 ; GCN: flat_store_dword
 define amdgpu_kernel void @kernel(i32 addrspace(1)* %out) #0 {
 entry:
   store i32 0, i32 addrspace(1)* %out
   ret void
 }
 
-attributes #0 = { nounwind }
+; GCN-LABEL: {{^}}func:
+define internal void @func() #1 {
+  call void asm sideeffect "", "~{v0},~{v1},~{v2},~{v3},~{v4},~{v5},~{v6},~{v7}"() #0
+  ret void
+}
+
+; GCN-LABEL: {{^}}kernel_call:
+; GCN-NOT: buffer_store
+; GCN-NOT: buffer_load
+; GCN-NOT: readlane
+; GCN-NOT: writelane
+; GCN: flat_load_dword v8
+; GCN: s_swappc_b64
+; GCN-NOT: buffer_store
+; GCN-NOT: buffer_load
+; GCN-NOT: readlane
+; GCN-NOT: writelane
+; GCN: flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, v8
+define amdgpu_kernel void @kernel_call() #0 {
+  %vgpr = load volatile i32, i32 addrspace(1)* undef
+  call void @func()
+  store volatile i32 %vgpr, i32 addrspace(1)* undef
+  ret void
+}
+
+; GCN-LABEL: {{^}}func_regular_call:
+; GCN-NOT: buffer_store
+; GCN-NOT: buffer_load
+; GCN-NOT: readlane
+; GCN-NOT: writelane
+; GCN: flat_load_dword v8
+; GCN: s_swappc_b64
+; GCN-NOT: buffer_store
+; GCN-NOT: buffer_load
+; GCN-NOT: readlane
+; GCN-NOT: writelane
+; GCN: flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, v8
+define void @func_regular_call() #1 {
+  %vgpr = load volatile i32, i32 addrspace(1)* undef
+  call void @func()
+  store volatile i32 %vgpr, i32 addrspace(1)* undef
+  ret void
+}
+
+; GCN-LABEL: {{^}}func_tail_call:
+; GCN: s_waitcnt
+; GCN-NEXT: s_getpc_b64 s[6:7]
+; GCN-NEXT: s_add_u32 s6,
+; GCN-NEXT: s_addc_u32 s7,
+; GCN-NEXT: s_setpc_b64 s[6:7]
+define void @func_tail_call() #1 {
+  tail call void @func()
+  ret void
+}
+
+attributes #0 = { nounwind norecurse }
+attributes #1 = { nounwind norecurse noinline }