[IR] Relax convergence requirements on call (#135794)

Before this commit, having a convergence token on a non-convergent call
was considered to be an error.
This commit relaxes this requirement and allows convergence tokens to be
present on non-convergent calls.

When such token is present, they have no effect as the underlying call
is non-convergent.
This allows passes like DCE to strip `convergent` attribute from
functions for which all convergent operations have been stripped. When
this happens, a convergence token can still exist in the call-site,
causing the verifier to complain.

Alternatives have been considered in #134863 and #134844.

llvm/include/llvm/IR/GenericConvergenceVerifierImpl.h

@@ -102,9 +102,6 @@ void GenericConvergenceVerifier<ContextT>::visit(const InstructionT &I) {
     SeenFirstConvOp = true;
 
   if (TokenDef || ConvOp != CONV_NONE) {
-    Check(isConvergent(I),
-          "Convergence control token can only be used in a convergent call.",
-          {Context.print(&I)});
     Check(ConvergenceKind != UncontrolledConvergence,
           "Cannot mix controlled and uncontrolled convergence in the same "
           "function.",

llvm/test/Assembler/convergence-control.ll

@@ -24,6 +24,13 @@ define void @mixed2() {
   ret void
 }
 
+; convergence control token can be used on non-convergent calls,
+; but it has no effect.
+define void @mixed3() {
+  %t05_tok1 = call token @llvm.experimental.convergence.anchor()
+  call void @g() [ "convergencectrl"(token %t05_tok1) ]
+  ret void
+}
 
 define void @region_nesting1(i1 %arg) convergent {
 A:

llvm/test/Transforms/ADCE/convergence.ll

@@ -0,0 +1,93 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals all --version 5
+; RUN: opt %s -passes=adce -S | FileCheck %s
+
+define i32 @foo(i32 %a) #0 {
+; CHECK-LABEL: define i32 @foo(
+; CHECK-SAME: i32 [[A:%.*]]) #[[ATTR0:[0-9]+]] {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    ret i32 [[A]]
+;
+entry:
+  %tk = call token @llvm.experimental.convergence.entry()
+  ret i32 %a
+}
+
+define void @bar() #0 {
+; CHECK-LABEL: define void @bar(
+; CHECK-SAME: ) #[[ATTR0]] {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    ret void
+;
+entry:
+  %tk = call token @llvm.experimental.convergence.anchor()
+  ret void
+}
+
+define void @baz() #0 {
+; CHECK-LABEL: define void @baz(
+; CHECK-SAME: ) #[[ATTR0]] {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    br label %[[HEADER:.*]]
+; CHECK:       [[HEADER]]:
+; CHECK-NEXT:    br i1 true, label %[[BODY:.*]], label %[[EXIT:.*]]
+; CHECK:       [[BODY]]:
+; CHECK-NEXT:    br label %[[HEADER]]
+; CHECK:       [[EXIT]]:
+; CHECK-NEXT:    ret void
+;
+entry:
+  %tk0 = call token @llvm.experimental.convergence.entry()
+  br label %header
+
+header:
+  %tk1 = call token @llvm.experimental.convergence.loop() [ "convergencectrl"(token %tk0) ]
+  br i1 true, label %body, label %exit
+
+body:
+  br label %header
+
+exit:
+  ret void
+}
+
+define void @indirect_inner() #0 {
+; CHECK-LABEL: define void @indirect_inner(
+; CHECK-SAME: ) #[[ATTR0]] {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    ret void
+;
+entry:
+  %tk0 = call token @llvm.experimental.convergence.entry()
+  ret void
+}
+
+define void @indirect() #0 {
+; CHECK-LABEL: define void @indirect(
+; CHECK-SAME: ) #[[ATTR0]] {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    [[TK0:%.*]] = call token @llvm.experimental.convergence.entry()
+; CHECK-NEXT:    [[VAR:%.*]] = alloca ptr, align 8
+; CHECK-NEXT:    store ptr @indirect_inner, ptr [[VAR]], align 8
+; CHECK-NEXT:    [[PTR:%.*]] = load ptr, ptr [[VAR]], align 8
+; CHECK-NEXT:    call void [[PTR]]() #[[ATTR0]] [ "convergencectrl"(token [[TK0]]) ]
+; CHECK-NEXT:    ret void
+;
+entry:
+  %tk0 = call token @llvm.experimental.convergence.entry()
+  %var = alloca ptr, align 8
+  store ptr @indirect_inner, ptr %var, align 8
+  %ptr = load ptr, ptr %var, align 8
+  call void %ptr() convergent [ "convergencectrl"(token %tk0) ]
+  ret void
+}
+
+declare token @llvm.experimental.convergence.entry() #1
+declare token @llvm.experimental.convergence.anchor() #1
+declare token @llvm.experimental.convergence.loop() #1
+
+attributes #0 = { convergent }
+attributes #1 = { convergent nocallback nofree nosync nounwind willreturn memory(none) }
+;.
+; CHECK: attributes #[[ATTR0]] = { convergent }
+; CHECK: attributes #[[ATTR1:[0-9]+]] = { convergent nocallback nofree nosync nounwind willreturn memory(none) }
+;.

llvm/test/Transforms/BDCE/convergence.ll

@@ -0,0 +1,62 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals all --version 5
+; RUN: opt %s -passes=bdce -S | FileCheck %s
+
+define i32 @foo(i32 %a) #0 {
+; CHECK-LABEL: define i32 @foo(
+; CHECK-SAME: i32 [[A:%.*]]) #[[ATTR0:[0-9]+]] {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    ret i32 [[A]]
+;
+entry:
+  %tk0 = call token @llvm.experimental.convergence.entry()
+  ret i32 %a
+}
+
+define void @bar() #0 {
+; CHECK-LABEL: define void @bar(
+; CHECK-SAME: ) #[[ATTR0]] {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    ret void
+;
+entry:
+  %tk0 = call token @llvm.experimental.convergence.anchor()
+  ret void
+}
+
+define void @baz() #0 {
+; CHECK-LABEL: define void @baz(
+; CHECK-SAME: ) #[[ATTR0]] {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    br label %[[HEADER:.*]]
+; CHECK:       [[HEADER]]:
+; CHECK-NEXT:    br i1 true, label %[[BODY:.*]], label %[[EXIT:.*]]
+; CHECK:       [[BODY]]:
+; CHECK-NEXT:    br label %[[HEADER]]
+; CHECK:       [[EXIT]]:
+; CHECK-NEXT:    ret void
+;
+entry:
+  %tk0 = call token @llvm.experimental.convergence.entry()
+  br label %header
+
+header:
+  %tk1 = call token @llvm.experimental.convergence.loop() [ "convergencectrl"(token %tk0) ]
+  br i1 true, label %body, label %exit
+
+body:
+  br label %header
+
+exit:
+  ret void
+}
+
+declare token @llvm.experimental.convergence.entry() #1
+declare token @llvm.experimental.convergence.anchor() #1
+declare token @llvm.experimental.convergence.loop() #1
+
+attributes #0 = { convergent }
+attributes #1 = { convergent nocallback nofree nosync nounwind willreturn memory(none) }
+;.
+; CHECK: attributes #[[ATTR0]] = { convergent }
+; CHECK: attributes #[[ATTR1:[0-9]+]] = { convergent nocallback nofree nosync nounwind willreturn memory(none) }
+;.

llvm/test/Transforms/FunctionAttrs/convergent.ll

@@ -1,4 +1,4 @@
-; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --function-signature --check-attributes
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --function-signature --check-attributes --check-globals
 ; RUN: opt -passes=function-attrs -S < %s | FileCheck %s
 
 define i32 @nonleaf() convergent {
@@ -129,3 +129,43 @@ define i32 @noopt_friend() convergent {
   %a = call i32 @noopt()
   ret i32 0
 }
+
+
+; A function which is stripped of its convergent attribute, even
+; if used in a controlled convergence call.
+; This should be OK.
+define i32 @leaf_noconvergent_used() convergent {
+; CHECK: Function Attrs: mustprogress nofree norecurse nosync nounwind willreturn memory(none)
+; CHECK-LABEL: define {{[^@]+}}@leaf_noconvergent_used
+; CHECK-SAME: () #[[ATTR0]] {
+; CHECK-NEXT:    ret i32 0
+;
+  ret i32 0
+}
+
+define i32 @nonleaf_convergent() convergent {
+; CHECK: Function Attrs: convergent mustprogress nofree norecurse nosync nounwind willreturn memory(none)
+; CHECK-LABEL: define {{[^@]+}}@nonleaf_convergent
+; CHECK-SAME: () #[[ATTR7:[0-9]+]] {
+; CHECK-NEXT:    [[TMP1:%.*]] = call token @llvm.experimental.convergence.entry()
+; CHECK-NEXT:    [[TMP2:%.*]] = call i32 @leaf_noconvergent_used() [ "convergencectrl"(token [[TMP1]]) ]
+; CHECK-NEXT:    ret i32 0
+;
+  %1 = call token @llvm.experimental.convergence.entry()
+  %2 = call i32 @leaf_noconvergent_used() [ "convergencectrl"(token %1) ]
+  ret i32 0
+}
+
+
+declare token @llvm.experimental.convergence.entry() #1
+;.
+; CHECK: attributes #[[ATTR0]] = { mustprogress nofree norecurse nosync nounwind willreturn memory(none) }
+; CHECK: attributes #[[ATTR1]] = { convergent }
+; CHECK: attributes #[[ATTR2]] = { norecurse }
+; CHECK: attributes #[[ATTR3:[0-9]+]] = { convergent nocallback nounwind }
+; CHECK: attributes #[[ATTR4]] = { convergent norecurse nounwind }
+; CHECK: attributes #[[ATTR5]] = { nofree nosync nounwind memory(none) }
+; CHECK: attributes #[[ATTR6]] = { convergent noinline optnone }
+; CHECK: attributes #[[ATTR7]] = { convergent mustprogress nofree norecurse nosync nounwind willreturn memory(none) }
+; CHECK: attributes #[[ATTR8:[0-9]+]] = { convergent nocallback nofree nosync nounwind willreturn memory(none) }
+;.

llvm/test/Verifier/convergencectrl-invalid.ll

@@ -20,14 +20,6 @@ define void @wrong_token() {
   ret void
 }
 
-; CHECK: Convergence control token can only be used in a convergent call.
-; CHECK-NEXT  call void @g(){{.*}}%t05_tok1
-define void @missing.attribute() {
-  %t05_tok1 = call token @llvm.experimental.convergence.anchor()
-  call void @g() [ "convergencectrl"(token %t05_tok1) ]
-  ret void
-}
-
 ; CHECK: The 'convergencectrl' bundle requires exactly one token use.
 ; CHECK-NEXT:  call void @g()
 define void @multiple_tokens() {