From 65ea6fd48af95fb62c56fea9cb2414bd2eaaa8a0 Mon Sep 17 00:00:00 2001
From: Dmitry Babokin <babokin@gmail.com>
Date: Wed, 13 Nov 2013 13:15:01 +0400
Subject: [PATCH] Reasoning to use sse4 bitcode file

---
 builtins.cpp | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/builtins.cpp b/builtins.cpp
index 2c9703c6..2afd92d9 100644
--- a/builtins.cpp
+++ b/builtins.cpp
@@ -943,6 +943,15 @@ DefineStdlib(SymbolTable *symbolTable, llvm::LLVMContext *ctx, llvm::Module *mod
         switch (g->target->getVectorWidth()) {
         case 4:
             if (g->target->getDataTypeWidth() == 32) {
+                // Note here that for avx1-i32x4 we are using bitcode file for
+                // sse4-i32x4. This is intentional and good enough.
+                // AVX target implies appropriate target-feature attrbute,
+                // which forces LLVM to generate AVX code, even for SSE4
+                // intrinsics. Except that the only "missing" feature in sse4
+                // target is implemenation of __masked_[store|load]_[i32|i64]
+                // using maskmov instruction. But it's not very popular
+                // intrinsics, so we assume the implementation to be good
+                // enough at the moment.
                 if (runtime32) {
                     EXPORT_MODULE(builtins_bitcode_sse4_32bit);
                 }