Emulator-Archive/ppsspp
1
0
Fork 0
mirror of https://github.com/hrydgard/ppsspp.git synced 2024-11-27 07:20:49 +00:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

x86jit: Correct vh2f NAN handling (#16275)

Browse Source 
* x86jit: Correct vh2f NAN handling.

Allows another test to pass.

* x86jit: Reuse MAccessibleDisp().
This commit is contained in:
Unknown W. Brackets 2022-10-23 01:09:29 -07:00 committed by GitHub
parent bd81aa328a
commit 813bfded92
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
4 changed files with 47 additions and 29 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
Common/x64Emitter.h
View File

@ -278,6 +278,19 @@ inline OpArg MRegSum(X64Reg base, X64Reg offset)
return MComplex(base, offset, 1, 0);
}
template <typename T>
inline bool Accessible(const T *t1, const T *t2) {
ptrdiff_t diff = (const uint8_t *)t1 - (const uint8_t *)t2;
return diff > -0x7FFFFFE0 && diff < 0x7FFFFFE0;
}
template <typename T>
inline OpArg MAccessibleDisp(X64Reg r, const T *tbase, const T *t) {
_assert_(Accessible(tbase, t));
ptrdiff_t diff = (const uint8_t *)t - (const uint8_t *)tbase;
return MDisp(r, (int)diff);
}
inline OpArg Imm8 (u8 imm) {return OpArg(imm, SCALE_IMM8);}
inline OpArg Imm16(u16 imm) {return OpArg(imm, SCALE_IMM16);} //rarely used
inline OpArg Imm32(u32 imm) {return OpArg(imm, SCALE_IMM32);}

48
Core/MIPS/x86/CompVFPU.cpp
View File

@ -1605,13 +1605,25 @@ void Jit::Comp_Vh2f(MIPSOpcode op) {
#define SSE_CONST4(name, val) alignas(16) static const u32 name[4] = { (val), (val), (val), (val) }
SSE_CONST4(mask_nosign, 0x7fff);
SSE_CONST4(nan_mantissa, 0x800003ff);
SSE_CONST4(magic, (254 - 15) << 23);
SSE_CONST4(was_infnan, 0x7bff);
SSE_CONST4(exp_infnan, 255 << 23);
// TODO: Fix properly
if (!RipAccessible(mask_nosign)) {
DISABLE;
OpArg mask_nosign_arg, nan_mantissa_arg, magic_arg, was_infnan_arg, exp_infnan_arg;
if (RipAccessible(mask_nosign)) {
mask_nosign_arg = M(&mask_nosign[0]);
nan_mantissa_arg = M(&nan_mantissa[0]);
magic_arg = M(&magic[0]);
was_infnan_arg = M(&was_infnan[0]);
exp_infnan_arg = M(&exp_infnan[0]);
} else {
MOV(PTRBITS, R(TEMPREG), ImmPtr(&mask_nosign[0]));
mask_nosign_arg = MAccessibleDisp(TEMPREG, &mask_nosign[0], &mask_nosign[0]);
nan_mantissa_arg = MAccessibleDisp(TEMPREG, &mask_nosign[0], &nan_mantissa[0]);
magic_arg = MAccessibleDisp(TEMPREG, &mask_nosign[0], &magic[0]);
was_infnan_arg = MAccessibleDisp(TEMPREG, &mask_nosign[0], &was_infnan[0]);
exp_infnan_arg = MAccessibleDisp(TEMPREG, &mask_nosign[0], &exp_infnan[0]);
}
#undef SSE_CONST4
@ -1649,30 +1661,36 @@ void Jit::Comp_Vh2f(MIPSOpcode op) {
// OK, 16 bits in each word.
// Let's go. Deep magic here.
MOVAPS(XMM1, R(XMM0));
ANDPS(XMM0, M(&mask_nosign[0])); // xmm0 = expmant. not rip accessible but bailing above
ANDPS(XMM0, mask_nosign_arg); // xmm0 = expmant
XORPS(XMM1, R(XMM0)); // xmm1 = justsign = expmant ^ xmm0
MOVAPS(tempR, R(XMM0));
PCMPGTD(tempR, M(&was_infnan[0])); // xmm2 = b_wasinfnan. not rip accessible but bailing above
PSLLD(XMM0, 13);
MULPS(XMM0, M(magic)); /// xmm0 = scaled
MULPS(XMM0, magic_arg); /// xmm0 = scaled
PSLLD(XMM1, 16); // xmm1 = sign
ANDPS(tempR, M(&exp_infnan[0])); // not rip accessible but bailing above
ORPS(XMM1, R(tempR));
ORPS(XMM0, R(XMM1));
// Now create a NAN mask, adding in the sign.
ORPS(XMM1, R(tempR)); // xmm1 = sign + original mantissa.
ANDPS(XMM1, nan_mantissa_arg); // xmm1 = original mantissa
PCMPGTD(tempR, was_infnan_arg); // xmm2 = b_wasinfnan
ORPS(XMM1, exp_infnan_arg); // xmm1 = infnan result
ANDPS(XMM1, R(tempR)); // xmm1 = infnan result OR zero if not infnan
ANDNPS(tempR, R(XMM0)); // tempR = result OR zero if infnan
ORPS(XMM1, R(tempR));
fpr.MapRegsV(dregs, outsize, MAP_NOINIT | MAP_DIRTY);
// TODO: Could apply D-prefix in parallel here...
MOVSS(fpr.V(dregs[0]), XMM0);
SHUFPS(XMM0, R(XMM0), _MM_SHUFFLE(3, 3, 2, 1));
MOVSS(fpr.V(dregs[1]), XMM0);
MOVSS(fpr.V(dregs[0]), XMM1);
SHUFPS(XMM1, R(XMM1), _MM_SHUFFLE(3, 3, 2, 1));
MOVSS(fpr.V(dregs[1]), XMM1);
if (sz != V_Single) {
SHUFPS(XMM0, R(XMM0), _MM_SHUFFLE(3, 3, 2, 1));
MOVSS(fpr.V(dregs[2]), XMM0);
SHUFPS(XMM0, R(XMM0), _MM_SHUFFLE(3, 3, 2, 1));
MOVSS(fpr.V(dregs[3]), XMM0);
SHUFPS(XMM1, R(XMM1), _MM_SHUFFLE(3, 3, 2, 1));
MOVSS(fpr.V(dregs[2]), XMM1);
SHUFPS(XMM1, R(XMM1), _MM_SHUFFLE(3, 3, 2, 1));
MOVSS(fpr.V(dregs[3]), XMM1);
}
ApplyPrefixD(dregs, outsize);

13
GPU/Software/DrawPixelX86.cpp
View File

@ -31,19 +31,6 @@ using namespace Gen;
namespace Rasterizer {
template <typename T>
static bool Accessible(const T *t1, const T *t2) {
ptrdiff_t diff = (const uint8_t *)t1 - (const uint8_t *)t2;
return diff > -0x7FFFFFE0 && diff < 0x7FFFFFE0;
}
template <typename T>
static OpArg MAccessibleDisp(X64Reg r, const T *tbase, const T *t) {
_assert_(Accessible(tbase, t));
ptrdiff_t diff = (const uint8_t *)t - (const uint8_t *)tbase;
return MDisp(r, (int)diff);
}
SingleFunc PixelJitCache::CompileSingle(const PixelFuncID &id) {
// Setup the reg cache and disallow spill for arguments.
regCache_.SetupABI({

2
test.py
View File

@ -69,6 +69,7 @@ tests_good = [
"cpu/cpu_alu/cpu_branch",
"cpu/cpu_alu/cpu_branch2",
"cpu/vfpu/colors",
"cpu/vfpu/convert",
"cpu/vfpu/gum",
"cpu/vfpu/matrix",
"cpu/vfpu/vavg",
@ -368,7 +369,6 @@ tests_good = [
tests_next = [
# These are the next tests up for fixing. These run by default.
"cpu/fpu/fcr",
"cpu/vfpu/convert",
"cpu/vfpu/prefixes",
"cpu/vfpu/vector",
"cpu/vfpu/vregs",