[英]Do NVCC and NVRTC support tail call optimization?
I am doing a smallish functional language in F# that will compile to C++ (for about the fourth time) and am wondering whether Nvidia's compilers have this feature now. 我正在用F#做一种较小的函数语言,它将编译为C ++(大约第四次),并且想知道Nvidia的编译器现在是否具有此功能。 I am hoping they do as it would save me the effort of having to implement tuples, and a Google search turns up nothing.
我希望他们能这样做,因为这可以节省我必须执行元组的工作,而Google搜索什么也没发现。
However, neither does a search for tail call
in the Cuda User Guide, so I guess it is unlikely. 但是,《 Cuda用户指南》中没有搜索
tail call
的功能,因此我想这不太可能。
Short answer: YES! 简短的回答:是的!
Long answer: this code 长答案:此代码
__device__ int i;
__device__ int f2(int val)
{
if((val % 6) == 0)
return val;
return f2(val + 1);
}
__global__ void f1(int x)
{
i = f2(x);
}
int main()
{
return 0;
}
Compiled with: 编译:
nvcc -keep -O2 bla.cu
nvcc -keep -O2 bla.cu
Generates: bla.ptx 生成:bla.ptx
//
// Generated by NVIDIA NVVM Compiler
//
// Compiler Build ID: CL-21124049
// Cuda compilation tools, release 8.0, V8.0.44
// Based on LLVM 3.4svn
//
.version 5.0
.target sm_20
.address_size 64
// .globl _Z2f1i
.global .align 4 .u32 i;
.visible .entry _Z2f1i(
.param .u32 _Z2f1i_param_0
)
{
.reg .pred %p<2>;
.reg .b32 %r<10>;
ld.param.u32 %r9, [_Z2f1i_param_0];
BB0_1:
mov.u32 %r1, %r9;
mul.hi.s32 %r4, %r1, 715827883;
shr.u32 %r5, %r4, 31;
add.s32 %r6, %r4, %r5;
mul.lo.s32 %r7, %r6, 6;
sub.s32 %r8, %r1, %r7;
add.s32 %r9, %r1, 1;
setp.ne.s32 %p1, %r8, 0;
@%p1 bra BB0_1;
st.global.u32 [i], %r1;
ret;
}
From nvdisasm of cubin: 来自库宾的nvdisasm:
//--------------------- .text._Z2f1i --------------------------
.section .text._Z2f1i,"ax",@progbits
.sectioninfo @"SHI_REGISTERS=6"
.align 4
.global _Z2f1i
.type _Z2f1i,@function
.size _Z2f1i,(.L_13 - _Z2f1i)
.other _Z2f1i,@"STO_CUDA_ENTRY STV_DEFAULT"
_Z2f1i:
.text._Z2f1i:
/*0000*/ MOV R1, c[0x1][0x100];
/*0008*/ MOV R0, c[0x0][0x20];
/*0010*/ NOP;
/*0018*/ NOP;
/*0020*/ NOP;
/*0028*/ NOP;
/*0030*/ NOP;
/*0038*/ NOP;
.L_2:
/*0040*/ IMUL.HI R2, R0, c[0x10][0x0];
/*0048*/ IMAD.U32.U32.HI R2, R2, 0x2, R2;
/*0050*/ IMAD R2, -R2, 0x6, R0;
/*0058*/ ISETP.NE.AND P0, PT, R2, RZ, PT;
/*0060*/ MOV R2, R0;
/*0068*/ IADD R0, R0, 0x1;
/*0070*/ @P0 BRA `(.L_2);
/*0078*/ MOV R4, c[0xe][0x0];
/*0080*/ MOV R5, c[0xe][0x4];
/*0088*/ ST.E [R4], R2;
/*0090*/ EXIT;
.L_13:
In both cases there is predicated branch, but no call. 在这两种情况下,都有谓词分支,但没有调用。
声明:本站的技术帖子网页,遵循CC BY-SA 4.0协议,如果您需要转载,请注明本站网址或者原文地址。任何问题请咨询:yoyou2525@163.com.