[llvm] [SPIRV] Do not use OpTypeRuntimeArray in Kernel env. (PR #149522)

[Dmitry Sidorov via llvm-commits]

================
@@ -828,9 +828,11 @@ SPIRVType *SPIRVGlobalRegistry::getOpTypeArray(uint32_t NumElems,
          "Invalid array element type");
   SPIRVType *SpvTypeInt32 = getOrCreateSPIRVIntegerType(32, MIRBuilder);
   SPIRVType *ArrayType = nullptr;
-  if (NumElems != 0) {
-    Register NumElementsVReg =
-        buildConstantInt(NumElems, MIRBuilder, SpvTypeInt32, EmitIR);
+  const SPIRVSubtarget &ST =
+      cast<SPIRVSubtarget>(MIRBuilder.getMF().getSubtarget());
+  if (NumElems != 0 || !ST.isShader()) {
+    Register NumElementsVReg = buildConstantInt(
+        NumElems ? NumElems : 1, MIRBuilder, SpvTypeInt32, EmitIR);
----------------
MrSidims wrote:

> Is this the correct solution?

That's interesting Q. I lean to say: yes. This replacement will legalize few GEPs that would be UB otherwise, but since they were already UB - such replacement should be fine.

> Are runtime arrays allowed in OpenCL?
> If not, should the compiler refuse to lower a runtime array to Kernel SPIR-V?

OpenCL allows unbounded arrays as kernel parameters and I believe GEP to unbound array kernel argument happens in the test provided by @maarquitos14 (though I'm having troubles to get a similar IR in my experiments).


https://github.com/llvm/llvm-project/pull/149522