[llvm] [AMDGPU] Preserve `noundef` and `range` during kernel argument loads (PR #118395)

[Krzysztof Drewniak via llvm-commits]

================
@@ -255,6 +255,25 @@ define amdgpu_kernel void @kern_i32(i32 %arg0) {
   ret void
 }
 
+define amdgpu_kernel void @kern_range_noundef_i32(i32 noundef range(i32 0, 8) %arg0) {
+; HSA-LABEL: @kern_range_noundef_i32(
+; HSA-NEXT:    [[KERN_RANGE_NOUNDEF_I32_KERNARG_SEGMENT:%.*]] = call nonnull align 16 dereferenceable(264) ptr addrspace(4) @llvm.amdgcn.kernarg.segment.ptr()
+; HSA-NEXT:    [[ARG0_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, ptr addrspace(4) [[KERN_RANGE_NOUNDEF_I32_KERNARG_SEGMENT]], i64 0
+; HSA-NEXT:    [[ARG0_LOAD:%.*]] = load i32, ptr addrspace(4) [[ARG0_KERNARG_OFFSET]], align 16, !range [[RNG2:![0-9]+]], !invariant.load [[META1]], !noundef [[META1]]
+; HSA-NEXT:    store volatile i32 [[ARG0_LOAD]], ptr addrspace(1) poison, align 4
+; HSA-NEXT:    ret void
+;
+; MESA-LABEL: @kern_range_noundef_i32(
+; MESA-NEXT:    [[KERN_RANGE_NOUNDEF_I32_KERNARG_SEGMENT:%.*]] = call nonnull align 16 dereferenceable(260) ptr addrspace(4) @llvm.amdgcn.kernarg.segment.ptr()
+; MESA-NEXT:    [[ARG0_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, ptr addrspace(4) [[KERN_RANGE_NOUNDEF_I32_KERNARG_SEGMENT]], i64 36
+; MESA-NEXT:    [[ARG0_LOAD:%.*]] = load i32, ptr addrspace(4) [[ARG0_KERNARG_OFFSET]], align 4, !range [[RNG2:![0-9]+]], !invariant.load [[META1]], !noundef [[META1]]
+; MESA-NEXT:    store volatile i32 [[ARG0_LOAD]], ptr addrspace(1) poison, align 4
+; MESA-NEXT:    ret void
+;
+  store volatile i32 %arg0, ptr addrspace(1) poison
----------------
krzysz00 wrote:

Is there a good idiom here for a side effect that keeps the argument alive? I remember someone recently complaining about stores to undef

https://github.com/llvm/llvm-project/pull/118395