当我编译这个简单的单词XOR示例时,Clang 3.4不会对它进行矢量化。为什么?它似乎比http://llvm.org/docs/Vectorizers.html#features中的示例更简单,并且我们在单个单词之间没有数据依赖性。所以从理论上讲,这可以进行矢量化。
#include <stdio.h>
void do_xor(const unsigned int num1[5], const unsigned int num2[5]) {
unsigned int num3[5];
// word-wise xor
for (int i = 0; i < 5; ++i)
num3[i] = num1[i] ^ num2[i];
for (int i = 4; i >= 0; --i)
printf("%08x", num3[i]);
printf("\n");
}
使用clang -O3 -fslp-vectorize-aggressive -emit-llvm .....
(强制向量化)我们最终得到(剥离的非功能部分):
define void @do_xor(i32* nocapture readonly %num1, i32* nocapture readonly %num2) #0 {
.preheader4:
%0 = load i32* %num1, align 4, !tbaa !1
%1 = load i32* %num2, align 4, !tbaa !1
%2 = xor i32 %1, %0
%3 = getelementptr inbounds i32* %num1, i64 1
%4 = load i32* %3, align 4, !tbaa !1
%5 = getelementptr inbounds i32* %num2, i64 1
%6 = load i32* %5, align 4, !tbaa !1
%7 = xor i32 %6, %4
%8 = getelementptr inbounds i32* %num1, i64 2
%9 = load i32* %8, align 4, !tbaa !1
%10 = getelementptr inbounds i32* %num2, i64 2
%11 = load i32* %10, align 4, !tbaa !1
%12 = xor i32 %11, %9
%13 = getelementptr inbounds i32* %num1, i64 3
%14 = load i32* %13, align 4, !tbaa !1
%15 = getelementptr inbounds i32* %num2, i64 3
%16 = load i32* %15, align 4, !tbaa !1
%17 = xor i32 %16, %14
%18 = getelementptr inbounds i32* %num1, i64 4
%19 = load i32* %18, align 4, !tbaa !1
%20 = getelementptr inbounds i32* %num2, i64 4
%21 = load i32* %20, align 4, !tbaa !1
%22 = xor i32 %21, %19
%23 = tail call i32 (i8*, ...)* @printf(i8* getelementptr inbounds ([5 x i8]* @.str, i64 0, i64 0), i32 %22) #2
%24 = tail call i32 (i8*, ...)* @printf(i8* getelementptr inbounds ([5 x i8]* @.str, i64 0, i64 0), i32 %17) #2
%25 = tail call i32 (i8*, ...)* @printf(i8* getelementptr inbounds ([5 x i8]* @.str, i64 0, i64 0), i32 %12) #2
%26 = tail call i32 (i8*, ...)* @printf(i8* getelementptr inbounds ([5 x i8]* @.str, i64 0, i64 0), i32 %7) #2
%27 = tail call i32 (i8*, ...)* @printf(i8* getelementptr inbounds ([5 x i8]* @.str, i64 0, i64 0), i32 %2) #2
%putchar = tail call i32 @putchar(i32 10) #2
ret void
}
没有生成矢量代码。为什么呢?
答案 0 :(得分:0)
我试图连续增加int数组的大小(循环的大小),如果big-int是&gt; = 14个单词,似乎向量化开始。
但我仍然不明白为什么前端有这个门槛。我仍然认为前端应该总是矢量化,后端应该决定如何降低矢量,因为它有关于目标的信息(例如,支持的矢量,指令的成本模型,......)