fabi/Xenia-Canary
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[x64] Fix vector mask issue and add missing tests

Browse Source
This commit is contained in:
Herman S.
2025-10-14 14:49:56 +09:00
parent 435ea98a5a
commit 15f61a1a10
7 changed files with 340 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
src/xenia/cpu/backend/x64/x64_seq_vector.cc
View File

@@ -1111,6 +1111,7 @@ struct VECTOR_SHL_V128
e.L(looper);
e.movzx(e.ecx, e.word[e.rsp + stack_offset_src2 + e.rdx]);
e.and_(e.cl, 0xF); // Mask shift count to 4 bits (0-15) for word shifts
e.shl(e.word[e.rsp + stack_offset_src1 + e.rdx], e.cl);
@@ -1205,6 +1206,7 @@ struct VECTOR_SHL_V128
e.L(looper);
e.mov(e.ecx, e.dword[e.rsp + stack_offset_src2 + e.rdx]);
e.and_(e.cl, 0x1F); // Mask shift count to 5 bits (0-31) for dword shifts
e.shl(e.dword[e.rsp + stack_offset_src1 + e.rdx], e.cl);
@@ -1313,6 +1315,7 @@ struct VECTOR_SHR_V128
// movzx is to eliminate any possible dep on previous value of rcx at start
// of loop
e.movzx(e.ecx, e.byte[e.rsp + stack_offset_src2 + e.rdx]);
e.and_(e.cl, 7); // Mask shift count to 3 bits (0-7) for byte shifts
// maybe using a memory operand as the left side isn't the best idea lol,
// still better than callnativesafe though agners docs have no timing info
// on shx [m], cl so shrug
@@ -1392,7 +1395,7 @@ struct VECTOR_SHR_V128
e.L(looper);
e.movzx(e.ecx, e.word[e.rsp + stack_offset_src2 + e.rdx]);
e.and_(e.cl, 0xF); // Mask shift count to 4 bits (0-15) for word shifts
e.shr(e.word[e.rsp + stack_offset_src1 + e.rdx], e.cl);
e.add(e.edx, 2);
@@ -1492,6 +1495,7 @@ struct VECTOR_SHR_V128
e.L(looper);
e.mov(e.ecx, e.dword[e.rsp + stack_offset_src2 + e.rdx]);
e.and_(e.cl, 0x1F); // Mask shift count to 5 bits (0-31) for dword shifts
e.shr(e.dword[e.rsp + stack_offset_src1 + e.rdx], e.cl);
e.add(e.edx, 4);
@@ -1612,6 +1616,7 @@ struct VECTOR_SHA_V128
// movzx is to eliminate any possible dep on previous value of rcx at start
// of loop
e.movzx(e.ecx, e.byte[e.rsp + stack_offset_src2 + e.rdx]);
e.and_(e.cl, 7); // Mask shift count to 3 bits (0-7) for byte shifts
// maybe using a memory operand as the left side isn't the best idea lol,
// still better than callnativesafe though agners docs have no timing info
// on shx [m], cl so shrug
@@ -1691,7 +1696,7 @@ struct VECTOR_SHA_V128
e.L(looper);
e.movzx(e.ecx, e.word[e.rsp + stack_offset_src2 + e.rdx]);
e.and_(e.cl, 0xF); // Mask shift count to 4 bits (0-15) for word shifts
e.sar(e.word[e.rsp + stack_offset_src1 + e.rdx], e.cl);
e.add(e.edx, 2);
@@ -1775,6 +1780,7 @@ struct VECTOR_SHA_V128
e.L(looper);
e.mov(e.ecx, e.dword[e.rsp + stack_offset_src2 + e.rdx]);
e.and_(e.cl, 0x1F); // Mask shift count to 5 bits (0-31) for dword shifts
e.sar(e.dword[e.rsp + stack_offset_src1 + e.rdx], e.cl);
e.add(e.edx, 4);