amd/llvm: Mark pointer function arguments as 32-byte aligned.

Otherwise LLVM does not see the pointers as allowing speculative
loads.

The pipeline-db results are pretty wild, but mostly what is to be
expected from allowing more code movement in LLVM:

Totals from affected shaders:
SGPRS: 157728 -> 168336 (6.73 %)
VGPRS: 158628 -> 158664 (0.02 %)
Spilled SGPRs: 10845 -> 24753 (128.24 %)
Spilled VGPRs: 13 -> 13 (0.00 %)
Private memory VGPRs: 0 -> 0 (0.00 %)
Scratch size: 8 -> 8 (0.00 %) dwords per thread
Code Size: 17189180 -> 17313712 (0.72 %) bytes
LDS: 204 -> 204 (0.00 %) blocks
Max Waves: 5700 -> 5687 (-0.23 %)
Wait states: 0 -> 0 (0.00 %)

This gives some boosts for shaders we can move a descriptor load
outside a loop.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/3159>

diff --git a/src/amd/llvm/ac_llvm_build.c b/src/amd/llvm/ac_llvm_build.c
index 6d5d1f7bab27d0ba9fc54830ed4c1b8dd8d24aa6..8be8433c997e207ee15312981ce51a02811dab3e 100644 (file)
--- a/src/amd/llvm/ac_llvm_build.c
+++ b/src/amd/llvm/ac_llvm_build.c
@@ -4988,6 +4988,7 @@ ac_build_main(const struct ac_shader_args *args,
                if (LLVMGetTypeKind(LLVMTypeOf(P)) == LLVMPointerTypeKind) {
                        ac_add_function_attr(ctx->context, main_function, i + 1, AC_FUNC_ATTR_NOALIAS);
                        ac_add_attr_dereferenceable(P, UINT64_MAX);
+                       ac_add_attr_alignment(P, 32);
                }
        }
 

diff --git a/src/amd/llvm/ac_llvm_helper.cpp b/src/amd/llvm/ac_llvm_helper.cpp
index f5383344dd4f3c37348449d0eb28c61dd6d3699d..779085b7e7893666d4b70157c88c6e08efd2d174 100644 (file)
--- a/src/amd/llvm/ac_llvm_helper.cpp
+++ b/src/amd/llvm/ac_llvm_helper.cpp
@@ -45,6 +45,18 @@ void ac_add_attr_dereferenceable(LLVMValueRef val, uint64_t bytes)
    A->addAttr(llvm::Attribute::getWithDereferenceableBytes(A->getContext(), bytes));
 }
 
+void ac_add_attr_alignment(LLVMValueRef val, uint64_t bytes)
+{
+#if LLVM_VERSION_MAJOR >= 10
+       llvm::Argument *A = llvm::unwrap<llvm::Argument>(val);
+       A->addAttr(llvm::Attribute::getWithAlignment(A->getContext(), llvm::Align(bytes)));
+#else
+       /* Avoid unused parameter warnings. */
+       (void)val;
+       (void)bytes;
+#endif
+}
+
 bool ac_is_sgpr_param(LLVMValueRef arg)
 {
        llvm::Argument *A = llvm::unwrap<llvm::Argument>(arg);

diff --git a/src/amd/llvm/ac_llvm_util.h b/src/amd/llvm/ac_llvm_util.h
index 1d90a6a1a03394530435a601602bb4ce9360103c..4a01eceadecaf67f3b6a45fadc23aaac15d44bd6 100644 (file)
--- a/src/amd/llvm/ac_llvm_util.h
+++ b/src/amd/llvm/ac_llvm_util.h
@@ -95,6 +95,7 @@ struct ac_llvm_compiler {
 
 const char *ac_get_llvm_processor_name(enum radeon_family family);
 void ac_add_attr_dereferenceable(LLVMValueRef val, uint64_t bytes);
+void ac_add_attr_alignment(LLVMValueRef val, uint64_t bytes);
 bool ac_is_sgpr_param(LLVMValueRef param);
 void ac_add_function_attr(LLVMContextRef ctx, LLVMValueRef function,
                           int attr_idx, enum ac_func_attr attr);