[amdgpu][lds] Raise an explicit unimplemented error on absolute address LDS variables

[JonChesterfield]

Summary

These aren't implemented. They could be at moderate implementation
complexity. Raising an error is better than silently miscompiling.

Patching now because the patch at D155125 is a step towards using this metadata
more extensively as part of the lowering path and that will interact badly with
input variables with this annotation.

Lowering user defined variables at specific addresses would drop this error,
put them at the requested position in the frame during this pass, and then
use the same codegen that will be used for the kernel specific struct shortly.

[jmmartinez]

Looks good to me. If you think it's relevant to privde a better diagnostic with the souce location I'd welcome that change.

I was thinking about using LLVMContext::diagnose + DiagnosticInfoUnsupported but it seems there is no alternative for GlobalVariables (only for Functions) and adding support for it is a whole other patch.

[JonChesterfield]

In D155132#4496383, @jmmartinez wrote:

I was thinking about using LLVMContext::diagnose + DiagnosticInfoUnsupported but it seems there is no alternative for GlobalVariables (only for Functions) and adding support for it is a whole other patch.

Interesting. That probably is something LLVM should have available. In this case, I'm hoping we'll implement this (put the variable at the requested location in the kernels that can reach it) as opposed to improving the diagnostics, though that may be unimplementable in the context of separate compilation (which is also not yet implemented).

[arsenm]

llvm/lib/Target/AMDGPU/AMDGPULowerModuleLDSPass.cpp

358

LLVMContext::emitError is a lazier way than going through the full diagnostic. You still need to produce some result though, so usually just replace with poison/undef

[arsenm]

llvm/lib/Target/AMDGPU/AMDGPULowerModuleLDSPass.cpp

358

I just looked at the implementation and it is emitting a diagnostic, but assuming it's related to inlineasm for some reason

llvm/lib/Target/AMDGPU/AMDGPULowerModuleLDSPass.cpp

static void getUsesOfLDSByFunction(CallGraph const &CG, Module &M,
 
     // Get uses from the current function, excluding uses by called functions
     // Two output variables to avoid walking the globals list twice
     for (auto &GV : M.globals()) {
       if (!AMDGPU::isLDSVariableToLower(GV)) {
         continue;
       }
 
+      if (GV.isAbsoluteSymbolRef()) {
+        report_fatal_error(
+            "LDS variables with absolute addresses are unimplemented.");
+      }
+
       for (User *V : GV.users()) {
         if (auto *I = dyn_cast<Instruction>(V)) {
           Function *F = I->getFunction();
           if (isKernelLDS(F)) {
             kernels[F].insert(&GV);
           } else {
             functions[F].insert(&GV);
           }

llvm/test/CodeGen/AMDGPU/lds-reject-absolute-addresses.ll

+; RUN: not --crash opt -S -mtriple=amdgcn-- -amdgpu-lower-module-lds < %s 2>&1 | FileCheck %s
+; RUN: not --crash opt -S -mtriple=amdgcn-- -passes=amdgpu-lower-module-lds < %s 2>&1 | FileCheck %s
+
+@var1 = addrspace(3) global i32 undef, !absolute_symbol !0
+
+; CHECK: LLVM ERROR: LDS variables with absolute addresses are unimplemented.
+define amdgpu_kernel void @kern() {
+  %val0 = load i32, ptr addrspace(3) @var1
+  %val1 = add i32 %val0, 4
+  store i32 %val1, ptr addrspace(3) @var1
+  ret void
+}
+
+!0 = !{i64 0, i64 1}
+