OrdinaryMagician/vkdoom_m
OrdinaryMagician/vkdoom_m
1
Fork 0
Code Activity

- removed the separate SSE2 version of the node builder's ClassifyLine function and all code associated with it.

Browse source 
Like everything else related to doing standard math with SSE2 vs. x87, there's nothing to be gained here with anything but first generation SSE2 systems which are irrelevant these days.

Taking 'thespir2.wad' from https://forum.zdoom.org/viewtopic.php?f=1&t=10655 the SSE2 version is reproducably ~3% slower than the x87 version on my Core i7, which quite closely mirrors all my previous tests since 2007.
Overall this just looks like an optimization not worth doing.
This commit is contained in:
Christoph Oelckers 2017-02-26 12:47:16 +01:00
commit 90b8dbb096
5 changed files with 2 additions and 311 deletions
Download patch file Download diff file Expand all files Collapse all files

92
src/nodebuild.cpp
View file

@ -1062,95 +1062,3 @@ void FNodeBuilder::PrintSet (int l, DWORD set)
}
Printf (PRINT_LOG, "*\n");
}
#ifdef BACKPATCH
#ifdef _WIN32
extern "C" {
__declspec(dllimport) int __stdcall VirtualProtect(void *, unsigned long, unsigned long, unsigned long *);
}
#define PAGE_EXECUTE_READWRITE 64
#else
#include <sys/mman.h>
#include <limits.h>
#include <unistd.h>
#endif
#ifdef __GNUC__
extern "C" int ClassifyLineBackpatch (node_t &node, const FSimpleVert *v1, const FSimpleVert *v2, int sidev[2])
#else
static int *CallerOffset;
int ClassifyLineBackpatchC (node_t &node, const FSimpleVert *v1, const FSimpleVert *v2, int sidev[2])
#endif
{
// Select the routine based on SSE2 availability and patch the caller so that
// they call that routine directly next time instead of going through here.
int *calleroffset;
int diff;
int (*func)(node_t &, const FSimpleVert *, const FSimpleVert *, int[2]);
#ifdef __GNUC__
calleroffset = (int *)__builtin_return_address(0);
#else
calleroffset = CallerOffset;
#endif
// printf ("Patching for SSE %d @ %p %d\n", SSELevel, calleroffset, *calleroffset);
#ifndef DISABLE_SSE
if (CPU.bSSE2)
{
func = ClassifyLineSSE2;
diff = int((char *)ClassifyLineSSE2 - (char *)calleroffset);
}
else
#endif
{
func = ClassifyLine2;
diff = int((char *)ClassifyLine2 - (char *)calleroffset);
}
calleroffset--;
// Patch the caller.
#ifdef _WIN32
unsigned long oldprotect;
if (VirtualProtect (calleroffset, 4, PAGE_EXECUTE_READWRITE, &oldprotect))
#else
// must make this page-aligned for mprotect
long pagesize = sysconf(_SC_PAGESIZE);
char *callerpage = (char *)((intptr_t)calleroffset & ~(pagesize - 1));
size_t protectlen = (intptr_t)calleroffset + sizeof(void*) - (intptr_t)callerpage;
int ptect;
if (!(ptect = mprotect(callerpage, protectlen, PROT_READ|PROT_WRITE|PROT_EXEC)))
#endif
{
*calleroffset = diff;
#ifdef _WIN32
VirtualProtect (calleroffset, sizeof(void*), oldprotect, &oldprotect);
#else
mprotect(callerpage, protectlen, PROT_READ|PROT_EXEC);
#endif
}
// And return by calling the real function.
return func (node, v1, v2, sidev);
}
#ifndef __GNUC__
// The ClassifyLineBackpatch() function here is a stub that uses inline assembly and nakedness
// to retrieve the return address of the stack before sending control to the real
// ClassifyLineBackpatchC() function. Since BACKPATCH shouldn't be defined on 64-bit builds,
// we're okay that VC++ can't do inline assembly on that target.
extern "C" __declspec(noinline) __declspec(naked) int ClassifyLineBackpatch (node_t &node, const FSimpleVert *v1, const FSimpleVert *v2, int sidev[2])
{
// We store the return address in a global, so as not to need to mess with the parameter list.
__asm
{
mov eax, [esp]
mov CallerOffset, eax
jmp ClassifyLineBackpatchC
}
}
#endif
#endif