Fixing vslo and vsro.

This commit is contained in:
Ben Vanik 2015-08-17 08:15:43 -07:00
parent 359e5b578a
commit 55068f230e
3 changed files with 85 additions and 5 deletions

View File

@ -1470,8 +1470,9 @@ XEEMITTER(vsldoi128, VX128_5(4, 16), VX128_5)(PPCHIRBuilder& f, InstrData& i) {
int InstrEmit_vslo_(PPCHIRBuilder& f, uint32_t vd, uint32_t va, uint32_t vb) {
// (VD) <- (VA) << (VB.b[F] & 0x78) (by octet)
// TODO(benvanik): flag for shift-by-octet as optimization.
Value* sh =
f.And(f.Extract(f.LoadVR(vb), 15, INT8_TYPE), f.LoadConstantInt8(0x78));
Value* sh = f.Shr(
f.And(f.Extract(f.LoadVR(vb), 15, INT8_TYPE), f.LoadConstantInt8(0x78)),
3);
Value* v = f.Permute(f.LoadVectorShl(sh), f.LoadVR(va), f.LoadZeroVec128(),
INT8_TYPE);
f.StoreVR(vd, v);
@ -1622,9 +1623,10 @@ XEEMITTER(vsrh, 0x10000244, VX)(PPCHIRBuilder& f, InstrData& i) {
int InstrEmit_vsro_(PPCHIRBuilder& f, uint32_t vd, uint32_t va, uint32_t vb) {
// (VD) <- (VA) >> (VB.b[F] & 0x78) (by octet)
// TODO(benvanik): flag for shift-by-octet as optimization.
Value* sh =
f.And(f.Extract(f.LoadVR(vb), 15, INT8_TYPE), f.LoadConstantInt8(0x78));
Value* v = f.Permute(f.LoadVectorShr(sh), f.LoadVR(va), f.LoadZeroVec128(),
Value* sh = f.Shr(
f.And(f.Extract(f.LoadVR(vb), 15, INT8_TYPE), f.LoadConstantInt8(0x78)),
3);
Value* v = f.Permute(f.LoadVectorShr(sh), f.LoadZeroVec128(), f.LoadVR(va),
INT8_TYPE);
f.StoreVR(vd, v);
return 0;

View File

@ -0,0 +1,39 @@
test_vslo_1:
#_ REGISTER_IN v3 [00112233, 44556677, 8899AABB, CCDDEEFF]
#_ REGISTER_IN v4 [04040404, 04040404, 04040404, 04040404]
vslo v3, v3, v4
blr
#_ REGISTER_OUT v3 [00112233, 44556677, 8899AABB, CCDDEEFF]
#_ REGISTER_OUT v4 [04040404, 04040404, 04040404, 04040404]
test_vslo_2:
#_ REGISTER_IN v3 [00112233, 44556677, 8899AABB, CCDDEEFF]
#_ REGISTER_IN v4 [08080808, 08080808, 08080808, 08080808]
vslo v3, v3, v4
blr
#_ REGISTER_OUT v3 [11223344, 55667788, 99AABBCC, DDEEFF00]
#_ REGISTER_OUT v4 [08080808, 08080808, 08080808, 08080808]
test_vslo_3:
#_ REGISTER_IN v3 [00112233, 44556677, 8899AABB, CCDDEEFF]
#_ REGISTER_IN v4 [12121212, 12121212, 12121212, 12121212]
vslo v3, v3, v4
blr
#_ REGISTER_OUT v3 [22334455, 66778899, AABBCCDD, EEFF0000]
#_ REGISTER_OUT v4 [12121212, 12121212, 12121212, 12121212]
test_vslo_4:
#_ REGISTER_IN v3 [00112233, 44556677, 8899AABB, CCDDEEFF]
#_ REGISTER_IN v4 [80808080, 80808080, 80808080, 80808080]
vslo v3, v3, v4
blr
#_ REGISTER_OUT v3 [00112233, 44556677, 8899AABB, CCDDEEFF]
#_ REGISTER_OUT v4 [80808080, 80808080, 80808080, 80808080]
test_vslo_5:
#_ REGISTER_IN v3 [00112233, 44556677, 8899AABB, CCDDEEFF]
#_ REGISTER_IN v4 [FFFFFFFF, FFFFFFFF, FFFFFFFF, FFFFFFFF]
vslo v3, v3, v4
blr
#_ REGISTER_OUT v3 [FF000000, 00000000, 00000000, 00000000]
#_ REGISTER_OUT v4 [FFFFFFFF, FFFFFFFF, FFFFFFFF, FFFFFFFF]

View File

@ -0,0 +1,39 @@
test_vsro_1:
#_ REGISTER_IN v3 [00112233, 44556677, 8899AABB, CCDDEEFF]
#_ REGISTER_IN v4 [04040404, 04040404, 04040404, 04040404]
vsro v3, v3, v4
blr
#_ REGISTER_OUT v3 [00112233, 44556677, 8899AABB, CCDDEEFF]
#_ REGISTER_OUT v4 [04040404, 04040404, 04040404, 04040404]
test_vsro_2:
#_ REGISTER_IN v3 [00112233, 44556677, 8899AABB, CCDDEEFF]
#_ REGISTER_IN v4 [08080808, 08080808, 08080808, 08080808]
vsro v3, v3, v4
blr
#_ REGISTER_OUT v3 [00001122, 33445566, 778899AA, BBCCDDEE]
#_ REGISTER_OUT v4 [08080808, 08080808, 08080808, 08080808]
test_vsro_3:
#_ REGISTER_IN v3 [00112233, 44556677, 8899AABB, CCDDEEFF]
#_ REGISTER_IN v4 [12121212, 12121212, 12121212, 12121212]
vsro v3, v3, v4
blr
#_ REGISTER_OUT v3 [00000011, 22334455, 66778899, AABBCCDD]
#_ REGISTER_OUT v4 [12121212, 12121212, 12121212, 12121212]
test_vsro_4:
#_ REGISTER_IN v3 [00112233, 44556677, 8899AABB, CCDDEEFF]
#_ REGISTER_IN v4 [80808080, 80808080, 80808080, 80808080]
vsro v3, v3, v4
blr
#_ REGISTER_OUT v3 [00112233, 44556677, 8899AABB, CCDDEEFF]
#_ REGISTER_OUT v4 [80808080, 80808080, 80808080, 80808080]
test_vsro_5:
#_ REGISTER_IN v3 [00112233, 44556677, 8899AABB, CCDDEEFF]
#_ REGISTER_IN v4 [FFFFFFFF, FFFFFFFF, FFFFFFFF, FFFFFFFF]
vsro v3, v3, v4
blr
#_ REGISTER_OUT v3 [00000000, 00000000, 00000000, 00000000]
#_ REGISTER_OUT v4 [FFFFFFFF, FFFFFFFF, FFFFFFFF, FFFFFFFF]