[llvm] [Offload] Full AMD support for olMemFill (PR #154958)

[Joseph Huber via llvm-commits]

================
@@ -2611,26 +2629,73 @@ struct AMDGPUDeviceTy : public GenericDeviceTy, AMDGenericDeviceTy {
   Error dataFillImpl(void *TgtPtr, const void *PatternPtr, int64_t PatternSize,
                      int64_t Size,
                      AsyncInfoWrapperTy &AsyncInfoWrapper) override {
-    hsa_status_t Status;
+    // Fast case, where we can use the 4 byte hsa_amd_memory_fill
+    if (Size % 4 == 0 &&
+        (PatternSize == 4 || PatternSize == 2 || PatternSize == 1)) {
+      uint32_t Pattern;
+      if (PatternSize == 1) {
+        auto *Byte = reinterpret_cast<const uint8_t *>(PatternPtr);
+        Pattern = *Byte | *Byte << 8 | *Byte << 16 | *Byte << 24;
+      } else if (PatternSize == 2) {
+        auto *Word = reinterpret_cast<const uint16_t *>(PatternPtr);
+        Pattern = *Word | (*Word << 16);
+      } else if (PatternSize == 4) {
+        Pattern = *reinterpret_cast<const uint32_t *>(PatternPtr);
+      } else {
+        // Shouldn't be here if the pattern size is outwith those values
+        std::terminate();
----------------
jhuber6 wrote:

Is this an assertion? Use `llvm_unreachable` for stuff like this.

https://github.com/llvm/llvm-project/pull/154958