别名分析如何与关键字__restrict
和__noalias
一起使用? 它是否认为它们是没有锯齿的证据? 或者根据指针简单计算自己的结果?别名分析和_restrict关键字 - C
望着LLVM的别名分析,在许多紧密循环做负载-binop店序列的结果,下面的瓶颈看到:即使 输入和输出指针用__restrict
,别名分析仍然充分标假设它们为'MayAlias
',并使循环结束时的商店依赖于循环中的所有负载。
例如
void _BitwiseOr_(unsigned char * __restrict * __restrict src1Addr, unsigned char * __restrict * __restrict src2Addr, unsigned char * __restrict * __restrict destAddr, unsigned int width) {
uchar16 * __restrict src1 = (uchar16 * __restrict) *src1Addr;
uchar16 * __restrict src2 = (uchar16 * __restrict) *src2Addr;
uchar16 * __restrict dest = (uchar16 * __restrict) *destAddr;
for (unsigned int i = 0; i < width; i += 4) {
*dest++ = *src1++ | *src2++;
*dest++ = *src1++ | *src2++;
*dest++ = *src1++ | *src2++;
*dest++ = *src1++ | *src2++;
}
}
。
define void @_BitwiseOr_(i8** noalias nocapture readonly %src1Addr, i8** noalias nocapture readonly %src2Addr, i8** noalias nocapture readonly %destAddr, i32 %width) local_unnamed_addr #0 {
entry:
%0 = bitcast i8** %src1Addr to <16 x i8>**
%1 = load <16 x i8>*, <16 x i8>** %0, align 4, !tbaa !2
%2 = bitcast i8** %src2Addr to <16 x i8>**
%3 = load <16 x i8>*, <16 x i8>** %2, align 4, !tbaa !2
%4 = bitcast i8** %destAddr to <16 x i8>**
%5 = load <16 x i8>*, <16 x i8>** %4, align 4, !tbaa !2
%6 = load <16 x i8>, <16 x i8>* %1, align 8, !tbaa !6
%7 = load <16 x i8>, <16 x i8>* %3, align 8, !tbaa !6
%or = or <16 x i8> %7, %6
store <16 x i8> %or, <16 x i8>* %5, align 8, !tbaa !6
ret void
}
别名分析回答,唯一的NO别名之间:
i8** src1addr - i8** src2addr,
i8** src1addr - i8** destaddr,
i8** src2addr - i8** destaddr,
i8** destaddr - i8** destaddr,
i8** src1addr - i8** src1addr,
i8** src2addr - i8** src2addr
为何不使用__restrict
关键字的“好处”? 是否有可能使其工作?
以上是通过用铿锵编译:
-cc1 -S -disable-free -main-file-name file.cpp -mllvm -disable-block-placement -funroll-loops -mllvm -unroll-allow-partial -mllvm -tail-merge-size=71 -mllvm -tail-dup-size=70 -fmath-errno -v -gcodeview -dwarf-column-info -coverage-file file.s -O3 -Wall -Werror=implicit-function-declaration -std=c++14 -fdeprecated-macro -fno-dwarf-directory-asm -ferror-limit 19 -fmessage-length 0 -ffreestanding -fallow-half-arguments-and-returns -fobjc-runtime=gcc -fdiagnostics-show-option -vectorize-loops -vectorize-slp -mllvm -no-phi-elim-live-out-early-exit -mllvm -use-cfl-aa=anders -mllvm -use-cfl-aa-in-codegen=anders -mllvm -debug -mllvm -da-delinearize -mllvm -mllvm -enable-tbaa -mllvm -enable-scoped-noalias -mllvm -evaluate-aa-metadata -mllvm -print-all-alias-modref-info
你还期待什么其他的“别名”?你提供的列表似乎涵盖了所有的输入? – Joky
我还应该期待%1 - %3,%3 - %5 aka src1 - src2等(检查更新后的源代码) – eternalStudent