プログラミング/gcc/Auto-vectorization - futada wiki

gcc Auto-vectorization

---

---

環境

• gcc version 4.8.2 (Ubuntu 4.8.2-19ubuntu1)

ベクトルに定数を加算

コード(vec.cpp)

#include <stdio.h>
 
void doFilter(const short* __restrict src,
              short* __restrict dst,
              const size_t n)
{
  src = (const short*)__builtin_assume_aligned(src, 16);
  dst = (short*)__builtin_assume_aligned(dst, 16);
 
  for(size_t i = 0; i < n; ++i){
    const short v = src[i] + 10;
    dst[i] = v;
  }
}

コンパイル

$ g++ -m32 -S -O3 -ftree-vectorizer-verbose=1 -fverbose-asm -msse2 ./vec.cpp
 
Analyzing loop at ./vec.cpp:10
 
 
Vectorizing loop at ./vec.cpp:10
 
./vec.cpp:10: note: === vect_do_peeling_for_loop_bound ===Setting upper bound of nb iterations for epilogue loop to 6
 
./vec.cpp:10: note: LOOP VECTORIZED.
./vec.cpp:3: note: vectorized 1 loops in function.
 
./vec.cpp:10: note: Completely unroll loop 6 times

生成されたコード(vec.s)の抜粋

movl	28(%esp), %eax	# n, n
	movl	20(%esp), %ebx	# src, src
	movl	24(%esp), %ecx	# dst, dst
	testl	%eax, %eax	# n
	je	.L1	#,
	movl	%eax, %edi	# n, bnd.4
	shrl	$3, %edi	#, bnd.4 (nは3ビット右シフトしてediに入れておく)
	leal	0(,%edi,8), %edx	#, i
	testl	%edx, %edx	# i
	je	.L10	#,
	cmpl	$7, %eax	#, n
	jbe	.L10	#,
	xorl	%esi, %esi	# ivtmp.22
	xorl	%ebp, %ebp	# D.2999 (ループ変数edpを0にする)
	movdqa	.LC0, %xmm0	#, tmp162
.L9:
	movdqa	(%ebx,%esi), %xmm1	# MEM[base: src_4, index: ivtmp.22_41, offset: 0B], vect_var_.13
	addl	$1, %ebp	#, D.2999
	paddw	%xmm0, %xmm1	# tmp162, vect_var_.13
	movdqa	%xmm1, (%ecx,%esi)	# vect_var_.13, MEM[base: dst_6, index: ivtmp.22_41, offset: 0B]
	addl	$16, %esi	#, ivtmp.22
	cmpl	%ebp, %edi	# D.2999, bnd.4 (ループ変数とn >> 3の比較)
	ja	.L9	#,