From: Matt Arsenault <Matthew.Arsenault@amd.com>
Date: Tue, 21 May 2019 23:23:10 +0000 (+0000)
Subject: AMDGPU: Assume call pseudos are convergent
X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=972131f95cbdecca06302766a52f0f8b496ee393;p=llvm

AMDGPU: Assume call pseudos are convergent

There should probably be nonconvergent versions, but my guess is it
doesn't matter in practice.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@361331 91177308-0d34-0410-b5e6-96231b3b80d8
---

diff --git a/lib/Target/AMDGPU/SIInstructions.td b/lib/Target/AMDGPU/SIInstructions.td
index 94c91ad7f2f..74302693239 100644
--- a/lib/Target/AMDGPU/SIInstructions.td
+++ b/lib/Target/AMDGPU/SIInstructions.td
@@ -383,6 +383,8 @@ def SI_CALL_ISEL : SPseudoInstSI <
   let isCall = 1;
   let SchedRW = [WriteBranch];
   let usesCustomInserter = 1;
+  // TODO: Should really base this on the call target
+  let isConvergent = 1;
 }
 
 // Wrapper around s_swappc_b64 with extra $callee parameter to track
@@ -393,6 +395,8 @@ def SI_CALL : SPseudoInstSI <
   let isCall = 1;
   let UseNamedOperandTable = 1;
   let SchedRW = [WriteBranch];
+  // TODO: Should really base this on the call target
+  let isConvergent = 1;
 }
 
 // Tail call handling pseudo
@@ -406,6 +410,8 @@ def SI_TCRETURN : SPseudoInstSI <(outs),
   let isBarrier = 1;
   let UseNamedOperandTable = 1;
   let SchedRW = [WriteBranch];
+  // TODO: Should really base this on the call target
+  let isConvergent = 1;
 }
 
 
diff --git a/test/CodeGen/AMDGPU/tail-duplication-convergent.ll b/test/CodeGen/AMDGPU/tail-duplication-convergent.ll
new file mode 100644
index 00000000000..70790311480
--- /dev/null
+++ b/test/CodeGen/AMDGPU/tail-duplication-convergent.ll
@@ -0,0 +1,105 @@
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -O2 -tail-dup-size=1000 -tail-dup-placement-threshold=1000 -enable-tail-merge=0 < %s | FileCheck -enable-var-scope -check-prefix=GCN %s
+
+; Need to to trigger tail duplication this during
+; MachineBlockPlacement, since calls aren't tail duplicated pre-RA.
+
+declare void @nonconvergent_func() #0
+declare void @convergent_func() #1
+declare void @llvm.amdgcn.s.barrier() #1
+
+; barrier shouldn't be duplicated.
+
+; GCN-LABEL: {{^}}taildup_barrier:
+; GCN: s_barrier
+; GCN-NOT: s_barrier
+define void @taildup_barrier(i32 addrspace(1)* %a, i32 addrspace(1)* %b, i1 %cond) #0 {
+entry:
+  br i1 %cond, label %bb1, label %bb2
+
+bb1:
+  store i32 0, i32 addrspace(1)* %a
+  br label %call
+
+bb2:
+  store i32 1, i32 addrspace(1)* %a
+  br label %call
+
+call:
+  call void @llvm.amdgcn.s.barrier()
+  br label %ret
+
+ret:
+  ret void
+}
+
+; GCN-LABEL: {{^}}taildup_convergent_call:
+; GCN: s_swappc_b64
+; GCN-NOT: s_swappc_b64
+define void @taildup_convergent_call(i32 addrspace(1)* %a, i32 addrspace(1)* %b, i1 %cond) #1 {
+entry:
+  br i1 %cond, label %bb1, label %bb2
+
+bb1:
+  store i32 0, i32 addrspace(1)* %a
+  br label %call
+
+bb2:
+  store i32 1, i32 addrspace(1)* %a
+  br label %call
+
+call:
+  call void @convergent_func()
+  br label %ret
+
+ret:
+  ret void
+}
+
+; TODO: Currently there is only one convergent call pseudo, but this
+; theoretically could use a nonconvergent variant.
+; GCN-LABEL: {{^}}taildup_nonconvergent_call:
+; GCN: s_swappc_b64
+; GCN-NOT: s_swappc_b64
+define void @taildup_nonconvergent_call(i32 addrspace(1)* %a, i32 addrspace(1)* %b, i1 %cond) #1 {
+entry:
+  br i1 %cond, label %bb1, label %bb2
+
+bb1:
+  store i32 0, i32 addrspace(1)* %a
+  br label %call
+
+bb2:
+  store i32 1, i32 addrspace(1)* %a
+  br label %call
+
+call:
+  call void @nonconvergent_func()
+  br label %ret
+
+ret:
+  ret void
+}
+
+; GCN-LABEL: {{^}}taildup_convergent_tailcall:
+; GCN: s_setpc_b64
+; GCN-NOT: s_setpc_b64
+define void @taildup_convergent_tailcall(i32 addrspace(1)* %a, i32 addrspace(1)* %b, i1 %cond) #1 {
+entry:
+  br i1 %cond, label %bb1, label %bb2
+
+bb1:
+  store i32 0, i32 addrspace(1)* %a
+  br label %call
+
+bb2:
+  store i32 1, i32 addrspace(1)* %a
+  br label %call
+
+call:
+  tail call void @convergent_func()
+  ret void
+}
+
+
+attributes #0 = { nounwind }
+attributes #1 = { nounwind convergent }