[AMDGPU] Respect `nobuiltin` when converting `printf`

[jhuber6]

Summary

The AMDGPU backend uses a pass to transform calls to the printf
function to a built-in verision for either HIP or OpenCL. Currently this
does not respect -fno-builtin and is always emitted. This allows the
user to turn off this functionality as is standard for these types of
built-in transformations. The motivation behind this change is to allow
the libc project to provide a linkable version of the printf
function in the future.

[sameerds]

llvm/lib/Target/AMDGPU/AMDGPUPrintfRuntimeBinding.cpp

441

Attribute::Builtin is documented as "only valid on a callsite", so there shouldn't be any need to check it here?

llvm/test/CodeGen/AMDGPU/printf_nobuiltin.ll

19

There should be another test where "-fno-builtin" is passed or equivalently, printf is declared with nobuiltin, but then an over-riding "builtin" is specified on a call to printf.

[jhuber6]

Changing logic for nobuiltin and builtin

[sameerds]

llvm/lib/Target/AMDGPU/AMDGPUPrintfRuntimeBinding.cpp

443–445

But doesn't (!CI->isNoBuiltin()) cover all of this? It calls hasFnAttrImpl(), which in turn checks both attributes in the positive and negative sense, including on the called function. The overall effect is as the function comment says: "check whether this call should not be treated as a builtin".

[jhuber6]

llvm/lib/Target/AMDGPU/AMDGPUPrintfRuntimeBinding.cpp

443–445

You're right, totally forgot that it checks the callee attributes as well.

[jhuber6]

Remove redundant check

llvm/lib/Target/AMDGPU/AMDGPUPrintfRuntimeBinding.cpp

bool AMDGPUPrintfRuntimeBindingImpl::run(Module &M) {
   if (TT.getArch() == Triple::r600)
     return false;
 
   auto PrintfFunction = M.getFunction("printf");
   if (!PrintfFunction || !PrintfFunction->isDeclaration())
     return false;
 
   for (auto &U : PrintfFunction->uses()) {
     if (auto *CI = dyn_cast<CallInst>(U.getUser())) {
-      if (CI->isCallee(&U))
+      if (CI->isCallee(&U) && !CI->isNoBuiltin())
         Printfs.push_back(CI);
     }
   }
 
   if (Printfs.empty())
     return false;
 
   TD = &M.getDataLayout();
 
   return lowerPrintfForGpu(M);
 }

llvm/test/CodeGen/AMDGPU/printf_builtin.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 2
+; RUN: opt -mtriple=amdgcn-- -passes=amdgpu-printf-runtime-binding -mcpu=gfx90a -S < %s | FileCheck --check-prefix=GCN %s
+
+@str = private unnamed_addr constant [10 x i8] c"string %d\00", align 1
+
+define void @caller(i32 %n) {
+; GCN-LABEL: define void @caller
+; GCN-SAME: (i32 [[N:%.*]]) #[[ATTR0:[0-9]+]] {
+; GCN-NEXT:  entry:
+; GCN-NEXT:    [[CALL:%.*]] = call i32 (ptr, ...) @printf(ptr @str, i32 [[N]])
+; GCN-NEXT:    ret void
+;
+entry:
+  %call = call i32 (ptr, ...) @printf(ptr @str, i32 %n)
+  ret void
+}
+
+define void @caller_builtin(i32 %n) {
+; GCN-LABEL: define void @caller_builtin
+; GCN-SAME: (i32 [[N:%.*]]) #[[ATTR0]] {
+; GCN-NEXT:  entry:
+; GCN-NEXT:    [[PRINTF_ALLOC_FN:%.*]] = call ptr addrspace(1) @__printf_alloc(i32 8)
+; GCN-NEXT:    br label [[ENTRY_SPLIT:%.*]]
+; GCN:       entry.split:
+; GCN-NEXT:    [[TMP0:%.*]] = icmp ne ptr addrspace(1) [[PRINTF_ALLOC_FN]], null
+; GCN-NEXT:    br i1 [[TMP0]], label [[TMP1:%.*]], label [[TMP2:%.*]]
+; GCN:       1:
+; GCN-NEXT:    [[PRINTBUFFID:%.*]] = getelementptr i8, ptr addrspace(1) [[PRINTF_ALLOC_FN]], i32 0
+; GCN-NEXT:    [[PRINTBUFFIDCAST:%.*]] = bitcast ptr addrspace(1) [[PRINTBUFFID]] to ptr addrspace(1)
+; GCN-NEXT:    store i32 1, ptr addrspace(1) [[PRINTBUFFIDCAST]], align 4
+; GCN-NEXT:    [[PRINTBUFFGEP:%.*]] = getelementptr i8, ptr addrspace(1) [[PRINTF_ALLOC_FN]], i32 4
+; GCN-NEXT:    store i32 [[N]], ptr addrspace(1) [[PRINTBUFFGEP]], align 4
+; GCN-NEXT:    br label [[TMP2]]
+; GCN:       2:
+; GCN-NEXT:    ret void
+;
+entry:
+  %call = call i32 (ptr, ...) @printf(ptr @str, i32 %n) #0
+  ret void
+}
+
+declare i32 @printf(ptr, ...) #1
+
+attributes #0 = { builtin }
+attributes #1 = { nobuiltin }

llvm/test/CodeGen/AMDGPU/printf_nobuiltin.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 2
+; RUN: opt -mtriple=amdgcn-- -passes=amdgpu-printf-runtime-binding -mcpu=gfx90a -S < %s | FileCheck --check-prefix=GCN %s
+
+@str = private unnamed_addr constant [10 x i8] c"string %d\00", align 1
+
+define void @caller(i32 %n) {
+; GCN-LABEL: define void @caller
+; GCN-SAME: (i32 [[N:%.*]]) #[[ATTR0:[0-9]+]] {
+; GCN-NEXT:  entry:
+; GCN-NEXT:    [[CALL:%.*]] = call i32 (ptr, ...) @printf(ptr @str, i32 [[N]]) #[[ATTR2:[0-9]+]]
+; GCN-NEXT:    ret void
+;
+entry:
+  %call = call i32 (ptr, ...) @printf(ptr @str, i32 %n) #0
+  ret void
+}
+
+define void @caller_builtin(i32 %n) {
+; GCN-LABEL: define void @caller_builtin
+; GCN-SAME: (i32 [[N:%.*]]) #[[ATTR0]] {
+; GCN-NEXT:  entry:
+; GCN-NEXT:    [[PRINTF_ALLOC_FN:%.*]] = call ptr addrspace(1) @__printf_alloc(i32 8)
+; GCN-NEXT:    br label [[ENTRY_SPLIT:%.*]]
+; GCN:       entry.split:
+; GCN-NEXT:    [[TMP0:%.*]] = icmp ne ptr addrspace(1) [[PRINTF_ALLOC_FN]], null
+; GCN-NEXT:    br i1 [[TMP0]], label [[TMP1:%.*]], label [[TMP2:%.*]]
+; GCN:       1:
+; GCN-NEXT:    [[PRINTBUFFID:%.*]] = getelementptr i8, ptr addrspace(1) [[PRINTF_ALLOC_FN]], i32 0
+; GCN-NEXT:    [[PRINTBUFFIDCAST:%.*]] = bitcast ptr addrspace(1) [[PRINTBUFFID]] to ptr addrspace(1)
+; GCN-NEXT:    store i32 1, ptr addrspace(1) [[PRINTBUFFIDCAST]], align 4
+; GCN-NEXT:    [[PRINTBUFFGEP:%.*]] = getelementptr i8, ptr addrspace(1) [[PRINTF_ALLOC_FN]], i32 4
+; GCN-NEXT:    store i32 [[N]], ptr addrspace(1) [[PRINTBUFFGEP]], align 4
+; GCN-NEXT:    br label [[TMP2]]
+; GCN:       2:
+; GCN-NEXT:    ret void
+;
+entry:
+  %call = call i32 (ptr, ...) @printf(ptr @str, i32 %n)
+  ret void
+}
+
+declare i32 @printf(ptr, ...)
+
+attributes #0 = { nobuiltin }