zdoom/raze-gles
0
0
Fork 0
mirror of https://github.com/ZDoom/raze-gles.git synced 2025-01-12 11:10:39 +00:00
Code Issues Projects Releases Packages Wiki Activity

To avoid potential compiler bugs when inlining inline assembly, remove the asm variants of msqrtasm and always use the C version. Compilers generate equivalent or better assembly, and the function is used exactly once, in a routine called during startup.

Browse source 
git-svn-id: https://svn.eduke32.com/eduke32@5909 1a8010ca-5511-0410-912e-c29ae57300e0
This commit is contained in:
hendricks266 2016-10-23 19:47:36 +00:00
parent 95452a5a8a
commit 6c6fdec733
2 changed files with 21 additions and 70 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

21
polymer/eduke32/build/src/engine.c
View file

@ -7013,6 +7013,27 @@ static void dorotatesprite(int32_t sx, int32_t sy, int32_t z, int16_t a, int16_t
}*/
}
static uint32_t msqrtasm(uint32_t c)
{
uint32_t a = 0x40000000l, b = 0x20000000l;
do
{
if (c >= a)
{
c -= a;
a += b*4;
}
a -= b;
a >>= 1;
b >>= 2;
} while (b);
if (c >= a)
a++;
return a >> 1;
}
//
// initksqrt (internal)

70
polymer/eduke32/build/src/engine_priv.h
View file

@ -61,31 +61,6 @@ extern uint16_t ATTRIBUTE((used)) sqrtable[4096], ATTRIBUTE((used)) shlookup[409
}
}
static inline int32_t msqrtasm(int32_t c)
{
_asm
{
push ebx
mov ecx, c
mov eax, 0x40000000
mov ebx, 0x20000000
begit:
cmp ecx, eax
jl skip
sub ecx, eax
lea eax, [eax+ebx*4]
skip :
sub eax, ebx
shr eax, 1
shr ebx, 2
jnz begit
cmp ecx, eax
sbb eax, -1
shr eax, 1
pop ebx
}
}
static inline int32_t getclipmask(int32_t a, int32_t b, int32_t c, int32_t d)
{
_asm
@ -154,28 +129,6 @@ extern uint16_t ATTRIBUTE((used)) sqrtable[4096], ATTRIBUTE((used)) shlookup[409
: "=a" (__r) : "a" (__a) : "ebx", "ecx", "cc"); \
__r; })
// edx is blown by this code somehow?!
#define msqrtasm(c) \
({ int32_t __r, __c=(c); \
__asm__ __volatile__ ( \
"movl $0x40000000, %%eax\n\t" \
"movl $0x20000000, %%ebx\n\t" \
"0:\n\t" \
"cmpl %%eax, %%ecx\n\t" \
"jl 1f\n\t" \
"subl %%eax, %%ecx\n\t" \
"leal (%%eax,%%ebx,4), %%eax\n\t" \
"1:\n\t" \
"subl %%ebx, %%eax\n\t" \
"shrl $1, %%eax\n\t" \
"shrl $2, %%ebx\n\t" \
"jnz 0b\n\t" \
"cmpl %%eax, %%ecx\n\t" \
"sbbl $-1, %%eax\n\t" \
"shrl $1, %%eax" \
: "=a" (__r) : "c" (__c) : "edx","ebx", "cc"); \
__r; })
#define getclipmask(a,b,c,d) \
({ int32_t __a=(a), __b=(b), __c=(c), __d=(d); \
__asm__ __volatile__ ("sarl $31, %%eax; addl %%ebx, %%ebx; adcl %%eax, %%eax; " \
@ -229,29 +182,6 @@ extern uint16_t ATTRIBUTE((used)) sqrtable[4096], ATTRIBUTE((used)) shlookup[409
return a;
}
static inline int32_t msqrtasm(uint32_t c)
{
uint32_t a, b;
a = 0x40000000l; // mov eax, 0x40000000
b = 0x20000000l; // mov ebx, 0x20000000
do // begit:
{
if (c >= a) // cmp ecx, eax / jl skip
{
c -= a; // sub ecx, eax
a += b*4; // lea eax, [eax+ebx*4]
} // skip:
a -= b; // sub eax, ebx
a >>= 1; // shr eax, 1
b >>= 2; // shr ebx, 2
} while (b); // jnz begit
if (c >= a) // cmp ecx, eax
a++; // sbb eax, -1
a >>= 1; // shr eax, 1
return a;
}
static inline int32_t getclipmask(int32_t a, int32_t b, int32_t c, int32_t d)
{
// Ken did this