Small fix to buffer load specialization pass to allow more specialization to happen. (#8653)

This allows us to further cleanup unnecessary copies in the target code
we generate.

Part of effort of #8652.

2 files changed, 7 insertions, 1 deletions

diff --git a/source/slang/slang-emit.cpp b/source/slang/slang-emit.cpp
index 804a44b81..7d72fb77f 100644
--- a/source/slang/slang-emit.cpp
+++ b/source/slang/slang-emit.cpp
@@ -1402,6 +1402,12 @@ Result linkAndOptimizeIR(
     // function parameters, reults, etc. is invalid.
     // We clean up the usages of resource values here.
     specializeResourceUsage(codeGenContext, irModule);
+
+    // Specialize calls to functions with values loaded from an immutable location,
+    // so that we directly load the value inside the callee, instead of loading the
+    // value outside of the callee and copy it in. This is necessary to avoid copying
+    // large values (e.g. arrays) in registers, where most of the elements are not
+    // actually used.
     specializeFuncsForBufferLoadArgs(codeGenContext, irModule);
 
     // Push `structuredBufferLoad` to the end of access chain to avoid loading unnecessary data.
diff --git a/source/slang/slang-ir-specialize-buffer-load-arg.cpp b/source/slang/slang-ir-specialize-buffer-load-arg.cpp
index a5a3dd2d9..c473c6047 100644
--- a/source/slang/slang-ir-specialize-buffer-load-arg.cpp
+++ b/source/slang/slang-ir-specialize-buffer-load-arg.cpp
@@ -89,7 +89,7 @@ struct FuncBufferLoadSpecializationCondition : FunctionCallSpecializeCondition
                 a = argLoad->getPtr();
 
                 // We can safely defer a load to the callee if the source dest is immutable.
-                if (isPointerToImmutableLocation(a))
+                if (isPointerToImmutableLocation(getRootAddr(a)))
                     continue;
 
                 // Otherwise, we check if there is no other instructions in between the load and the