ShuriZma
/
xenia
mirror of https://github.com/xenia-project/xenia.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #86 from hlide/master 

EXTRACT_I8, EXTRACT_I16, EXTRACT_I32, EXTRACT_F32: properly modified the...
This commit is contained in:
Ben Vanik 2014-05-31 08:19:03 -07:00
parent e45fc3dc56 0598df1d9e
commit 8be043fbd4
1 changed files with 30 additions and 22 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

52
src/alloy/backend/x64/x64_sequences.cc
View File

@ -3275,7 +3275,7 @@ EMITTER_OPCODE_TABLE(
// OPCODE_MUL_ADD // OPCODE_MUL_ADD
// ============================================================================ // ============================================================================
// d = 1 * 2 + 3 // d = 1 * 2 + 3
// $0 = $1×$0 + $2 // $0 = $1x$0 + $2
// TODO(benvanik): use other forms (132/213/etc) to avoid register shuffling. // TODO(benvanik): use other forms (132/213/etc) to avoid register shuffling.
// dest could be src2 or src3 - need to ensure it's not before overwriting dest // dest could be src2 or src3 - need to ensure it's not before overwriting dest
// perhaps use other 132/213/etc // perhaps use other 132/213/etc
@ -3338,7 +3338,7 @@ EMITTER_OPCODE_TABLE(
// OPCODE_MUL_SUB // OPCODE_MUL_SUB
// ============================================================================ // ============================================================================
// d = 1 * 2 - 3 // d = 1 * 2 - 3
// $0 = $2×$0 - $3 // $0 = $2x$0 - $3
// TODO(benvanik): use other forms (132/213/etc) to avoid register shuffling. // TODO(benvanik): use other forms (132/213/etc) to avoid register shuffling.
// dest could be src2 or src3 - need to ensure it's not before overwriting dest // dest could be src2 or src3 - need to ensure it's not before overwriting dest
// perhaps use other 132/213/etc // perhaps use other 132/213/etc
@ -4436,11 +4436,12 @@ EMITTER(EXTRACT_I8, MATCH(I<OPCODE_EXTRACT, I8<>, V128<>, I8<>>)) {
} else { } else {
XEASSERTALWAYS(); XEASSERTALWAYS();
// TODO(benvanik): try out hlide's version: // TODO(benvanik): try out hlide's version:
// mov eax, 0x80808003 // e.mov(e.eax, 0x80808003);
// xor al, i.src2.cvt8() // e.xor(e.al, i.src2);
// vmovd xmm0, eax // e.and(e.al, 15);
// vpshufb xmm0, i.src1, xmm0 // e.vmovd(e.xmm0, e.eax);
// vmovd i.dest.reg().cvt32(), xmm0 // e.vpshufb(e.xmm0, i.src1, e.xmm0);
// e.vmovd(i.dest.reg().cvt32(), e.xmm0);
} }
} }
}; };
@ -4450,14 +4451,13 @@ EMITTER(EXTRACT_I16, MATCH(I<OPCODE_EXTRACT, I16<>, V128<>, I8<>>)) {
e.vpextrw(i.dest.reg().cvt32(), i.src1, VEC128_W(i.src2.constant())); e.vpextrw(i.dest.reg().cvt32(), i.src1, VEC128_W(i.src2.constant()));
} else { } else {
// TODO(benvanik): try out hlide's version: // TODO(benvanik): try out hlide's version:
// xor eax, eax // e.mov(e.eax, 7);
// mov al, i.src2 // eax = [i, 0, 0, 0] // e.and(e.al, i.src2); // eax = [i&7, 0, 0, 0]
// xor eax, 0x80800203 // e.imul(e.eax, 0x00000202); // [(i&7)*2, (i&7)*2, 0, 0]
// imul eax, eax, 0x00000202 // [i*2, i*2, 0, 0] supposedly that 0<= i < 8 // e.xor(e.eax, 0x80800203); // [((i&7)*2)^3, ((i&7)*2)^2, 0x80, 0x80]
// add eax,0x80800100 // [i*2+0b00, i*2+0b01, 0x80, 0x80] // e.vmovd(e.xmm0, e.eax);
// vmovd xmm0, eax // e.vpshufb(e.xmm0, i.src1, e.xmm0);
// vpshufb xmm0, i.src1, xmm0 // e.vmovd(i.dest.reg().cvt32(), e.xmm0);
// vmovd i.dest.reg().cvt32(), xmm0
XEASSERTALWAYS(); XEASSERTALWAYS();
} }
} }
@ -4474,13 +4474,13 @@ EMITTER(EXTRACT_I32, MATCH(I<OPCODE_EXTRACT, I32<>, V128<>, I8<>>)) {
e.vpextrd(i.dest, i.src1, VEC128_D(i.src2.constant())); e.vpextrd(i.dest, i.src1, VEC128_D(i.src2.constant()));
} else { } else {
// TODO(benvanik): try out hlide's version: // TODO(benvanik): try out hlide's version:
// xor eax, eax // e.mov(e.eax, 3);
// mov al, i.src2 // eax = [i, 0, 0, 0] // e.and(e.al, i.src2); // eax = [(i&3), 0, 0, 0]
// imul eax, eax, 0x04040404 // [i*4, i*4, i*4, i*4] supposedly that 0<= i < 4 // e.imul(e.eax, 0x04040404); // [(i&3)*4, (i&3)*4, (i&3)*4, (i&3)*4]
// xor/add eax, 0x00010203 // [i*4+0b00, i*4+0b01, i*4+0b10, i*4+0b11] // e.add(e.eax, 0x00010203); // [((i&3)*4)+3, ((i&3)*4)+2, ((i&3)*4)+1, ((i&3)*4)+0]
// vmovd xmm0, eax // e.vmovd(e.xmm0, e.eax);
// vpshufb xmm0, i.src1, xmm0 // e.vpshufb(e.xmm0, i.src1, e.xmm0);
// vmovd i.dest.reg().cvt32(), xmm0 // e.vmovd(i.dest.reg().cvt32(), e.xmm0);
// Get the desired word in xmm0, then extract that. // Get the desired word in xmm0, then extract that.
e.xor(e.rax, e.rax); e.xor(e.rax, e.rax);
e.mov(e.al, i.src2); e.mov(e.al, i.src2);
@ -4500,6 +4500,14 @@ EMITTER(EXTRACT_F32, MATCH(I<OPCODE_EXTRACT, F32<>, V128<>, I8<>>)) {
e.vextractps(i.dest, i.src1, VEC128_F(i.src2.constant())); e.vextractps(i.dest, i.src1, VEC128_F(i.src2.constant()));
} else { } else {
XEASSERTALWAYS(); XEASSERTALWAYS();
// TODO(benvanik): try out hlide's version:
// e.mov(e.eax, 3);
// e.and(e.al, i.src2); // eax = [(i&3), 0, 0, 0]
// e.imul(e.eax, 0x04040404); // [(i&3)*4, (i&3)*4, (i&3)*4, (i&3)*4]
// e.add(e.eax, 0x00010203); // [((i&3)*4)+3, ((i&3)*4)+2, ((i&3)*4)+1, ((i&3)*4)+0]
// e.vmovd(e.xmm0, e.eax);
// e.vpshufb(e.xmm0, i.src1, e.xmm0);
// e.vmovd(i.dest, e.xmm0);
} }
} }
}; };