[AMDGPU] Strengthen preload intrinsics to noundef and nonnull

The various preloaded registers (workitem IDs, workgroup IDs, various
implicit pointers) always have a finite, inviriant, well-defined value
throughout a well-defined program.

In cases where the compiler infers or the user declares that some
implicit input will not be used (ex. via amdgcn-no-workitem-id-y), the
behavior of the entire program is undefined, since that misdeclaration
can cause arbitrary other preloaded-register intrinsics to access the
wrong register. This case is not expected to arise in practice, but
could occur when the no implicit argument attributes were not cleared
correctly in the prenence of external functions, indrect calls, or
other means of executing un-analyzable code. Failure to detect that
case would be a bug in the attributor.

This commit updates the documentation to reflect this long-standing
reality.

Then, on the basis that all implicit arguments are defined in all
correct programs, the intrinsics that return those values are
annototated with `noundef` and, in the case of implicit pointers,
`nonnull`.

This will prevent spurious calls to `freeze` in front-end
optimizations that destroy user-provided ranges on built-in IDs.

(While I'm here, this commit adds a test for `noundef` on kernel
arguments which is currently unimplemented)

Author: krzysz00

llvm/docs/AMDGPUUsage.rst

@@ -1395,8 +1395,11 @@ The AMDGPU backend supports the following LLVM IR attributes.
 
      "amdgpu-no-workitem-id-x"               Indicates the function does not depend on the value of the
                                              llvm.amdgcn.workitem.id.x intrinsic. If a function is marked with this
-                                             attribute, or reached through a call site marked with this attribute,
-                                             the value returned by the intrinsic is undefined. The backend can
+                                             attribute, or reached through a call site marked with this attribute, and
+                                             that intrinsic is called, the behavior of the program is undefined. (Whole-program
+                                             undefined behavior is used here because, for example, the absence of a required workitem
+                                             ID in the preloaded register set can mean that all other preloaded registers
+                                             are earlier than the compilation assumed they would be.) The backend can
                                              generally infer this during code generation, so typically there is no
                                              benefit to frontends marking functions with this.
 

llvm/include/llvm/IR/IntrinsicsAMDGPU.td

@@ -12,11 +12,16 @@
 
 def global_ptr_ty : LLVMQualPointerType<1>;
 
+// The amdgpu-no-* attributes (ex amdgpu-no-workitem-id-z) typically inferred
+// by the backend cause whole-program undefined behavior when violated, such as
+// by causing all other preload register intrinsics to return arbitrarily incorrect
+// values. Outside of such IR-level UB, these preloaded registers are always set
+// to a well-defined value and are thus `noundef`.
 class AMDGPUReadPreloadRegisterIntrinsic
-  : DefaultAttrsIntrinsic<[llvm_i32_ty], [], [IntrNoMem, IntrSpeculatable]>;
+  : DefaultAttrsIntrinsic<[llvm_i32_ty], [], [NoUndef<RetIndex>, IntrNoMem, IntrSpeculatable]>;
 
 class AMDGPUReadPreloadRegisterIntrinsicNamed<string name>
-  : DefaultAttrsIntrinsic<[llvm_i32_ty], [], [IntrNoMem, IntrSpeculatable]>, ClangBuiltin<name>;
+  : DefaultAttrsIntrinsic<[llvm_i32_ty], [], [NoUndef<RetIndex>, IntrNoMem, IntrSpeculatable]>, ClangBuiltin<name>;
 
 // Used to tag image and resource intrinsics with information used to generate
 // mem operands.
@@ -56,7 +61,7 @@ def int_r600_group_barrier : ClangBuiltin<"__builtin_r600_group_barrier">,
 def int_r600_implicitarg_ptr :
   ClangBuiltin<"__builtin_r600_implicitarg_ptr">,
   DefaultAttrsIntrinsic<[LLVMQualPointerType<7>], [],
-  [IntrNoMem, IntrSpeculatable]>;
+  [NoUndef<RetIndex>, IntrNoMem, IntrSpeculatable]>;
 
 def int_r600_rat_store_typed :
   // 1st parameter: Data
@@ -144,39 +149,44 @@ defm int_amdgcn_workgroup_id : AMDGPUReadPreloadRegisterIntrinsic_xyz_named
 
 def int_amdgcn_dispatch_ptr :
   DefaultAttrsIntrinsic<[LLVMQualPointerType<4>], [],
-  [Align<RetIndex, 4>, IntrNoMem, IntrSpeculatable]>;
+  [Align<RetIndex, 4>, NoUndef<RetIndex>, NonNull<RetIndex>, IntrNoMem, IntrSpeculatable]>;
 
 def int_amdgcn_queue_ptr :
   ClangBuiltin<"__builtin_amdgcn_queue_ptr">,
   DefaultAttrsIntrinsic<[LLVMQualPointerType<4>], [],
-  [Align<RetIndex, 4>, IntrNoMem, IntrSpeculatable]>;
+  [Align<RetIndex, 4>, NoUndef<RetIndex>, NonNull<RetIndex>, IntrNoMem, IntrSpeculatable]>;
 
 def int_amdgcn_kernarg_segment_ptr :
   ClangBuiltin<"__builtin_amdgcn_kernarg_segment_ptr">,
   DefaultAttrsIntrinsic<[LLVMQualPointerType<4>], [],
-  [Align<RetIndex, 4>, IntrNoMem, IntrSpeculatable]>;
+  [Align<RetIndex, 4>, NoUndef<RetIndex>, NonNull<RetIndex>, IntrNoMem, IntrSpeculatable]>;
 
 def int_amdgcn_implicitarg_ptr :
   ClangBuiltin<"__builtin_amdgcn_implicitarg_ptr">,
   DefaultAttrsIntrinsic<[LLVMQualPointerType<4>], [],
-  [Align<RetIndex, 4>, IntrNoMem, IntrSpeculatable]>;
+  [Align<RetIndex, 4>, NoUndef<RetIndex>, NonNull<RetIndex>, IntrNoMem, IntrSpeculatable]>;
 
+// Returns the amount of LDS statically allocated for this program.
+// This is no longer guaranteed to be a compile-time constant due to linking
+// support.
 def int_amdgcn_groupstaticsize :
   ClangBuiltin<"__builtin_amdgcn_groupstaticsize">,
-  DefaultAttrsIntrinsic<[llvm_i32_ty], [], [IntrNoMem, IntrSpeculatable]>;
+  DefaultAttrsIntrinsic<[llvm_i32_ty], [], [NoUndef<RetIndex>, IntrNoMem, IntrSpeculatable]>;
 
 def int_amdgcn_dispatch_id :
   ClangBuiltin<"__builtin_amdgcn_dispatch_id">,
-  DefaultAttrsIntrinsic<[llvm_i64_ty], [], [IntrNoMem, IntrSpeculatable]>;
+  DefaultAttrsIntrinsic<[llvm_i64_ty], [], [NoUndef<RetIndex>, IntrNoMem, IntrSpeculatable]>;
 
 // For internal use. Coordinates LDS lowering between IR transform and backend.
 def int_amdgcn_lds_kernel_id :
-  DefaultAttrsIntrinsic<[llvm_i32_ty], [], [IntrNoMem, IntrSpeculatable]>;
+  DefaultAttrsIntrinsic<[llvm_i32_ty], [], [NoUndef<RetIndex>, IntrNoMem, IntrSpeculatable]>;
 
 def int_amdgcn_implicit_buffer_ptr :
   ClangBuiltin<"__builtin_amdgcn_implicit_buffer_ptr">,
   DefaultAttrsIntrinsic<[LLVMQualPointerType<4>], [],
-  [Align<RetIndex, 4>, IntrNoMem, IntrSpeculatable]>;
+  [Align<RetIndex, 4>, Dereferenceable<RetIndex, 16>,
+    NoUndef<RetIndex>, NonNull<RetIndex>,
+    IntrNoMem, IntrSpeculatable]>;
 
 // Set EXEC to the 64-bit value given.
 // This is always moved to the beginning of the basic block.
@@ -199,7 +209,7 @@ def int_amdgcn_init_exec_from_input : Intrinsic<[],
 
 def int_amdgcn_wavefrontsize :
   ClangBuiltin<"__builtin_amdgcn_wavefrontsize">,
-  DefaultAttrsIntrinsic<[llvm_i32_ty], [], [IntrNoMem, IntrSpeculatable]>;
+  DefaultAttrsIntrinsic<[llvm_i32_ty], [], [NoUndef<RetIndex>, IntrNoMem, IntrSpeculatable]>;
 
 // Represent a relocation constant.
 def int_amdgcn_reloc_constant : DefaultAttrsIntrinsic<
@@ -1923,8 +1933,8 @@ def int_amdgcn_s_setreg :
 // s_getpc_b64 instruction returns a zero-extended value.
 def int_amdgcn_s_getpc :
   ClangBuiltin<"__builtin_amdgcn_s_getpc">,
-  DefaultAttrsIntrinsic<[llvm_i64_ty], [], [IntrNoMem, IntrSpeculatable,
-                                IntrWillReturn]>;
+  DefaultAttrsIntrinsic<[llvm_i64_ty], [], [NoUndef<RetIndex>, IntrNoMem,
+                                IntrSpeculatable, IntrWillReturn]>;
 
 // __builtin_amdgcn_interp_mov <param>, <attr_chan>, <attr>, <m0>
 // param values: 0 = P10, 1 = P20, 2 = P0
@@ -2044,7 +2054,7 @@ def int_amdgcn_ps_live : DefaultAttrsIntrinsic <
 // Query currently live lanes.
 // Returns true if lane is live (and not a helper lane).
 def int_amdgcn_live_mask : DefaultAttrsIntrinsic <[llvm_i1_ty],
-  [], [IntrReadMem, IntrInaccessibleMemOnly]
+  [], [NoUndef<RetIndex>, IntrReadMem, IntrInaccessibleMemOnly]
 >;
 
 def int_amdgcn_mbcnt_lo :
@@ -2517,7 +2527,7 @@ def int_amdgcn_mov_dpp8 :
 def int_amdgcn_s_get_waveid_in_workgroup :
   ClangBuiltin<"__builtin_amdgcn_s_get_waveid_in_workgroup">,
   Intrinsic<[llvm_i32_ty], [],
-    [IntrNoMem, IntrHasSideEffects, IntrWillReturn, IntrNoCallback, IntrNoFree]>;
+    [NoUndef<RetIndex>, IntrNoMem, IntrHasSideEffects, IntrWillReturn, IntrNoCallback, IntrNoFree]>;
 
 class AMDGPUAtomicRtn<LLVMType vt, LLVMType pt = llvm_anyptr_ty> : Intrinsic <
   [vt],
@@ -2751,7 +2761,7 @@ def int_amdgcn_global_load_tr_b128 : AMDGPULoadIntrinsic<global_ptr_ty>;
 
 // i32 @llvm.amdgcn.wave.id()
 def int_amdgcn_wave_id :
-  DefaultAttrsIntrinsic<[llvm_i32_ty], [], [IntrNoMem, IntrSpeculatable]>;
+  DefaultAttrsIntrinsic<[llvm_i32_ty], [], [NoUndef<RetIndex>, IntrNoMem, IntrSpeculatable]>;
 
 //===----------------------------------------------------------------------===//
 // Deep learning intrinsics.

llvm/lib/Target/AMDGPU/AMDGPULowerKernelArguments.cpp

@@ -120,8 +120,6 @@ static bool lowerKernelArguments(Function &F, const TargetMachine &TM) {
   CallInst *KernArgSegment =
       Builder.CreateIntrinsic(Intrinsic::amdgcn_kernarg_segment_ptr, {}, {},
                               nullptr, F.getName() + ".kernarg.segment");
-
-  KernArgSegment->addRetAttr(Attribute::NonNull);
   KernArgSegment->addRetAttr(
       Attribute::getWithDereferenceableBytes(Ctx, TotalKernArgSize));