Revert "Zero top bits in INSERTQ/EXTRQ (#3217)"

This reverts commit f5336358ea.
This commit is contained in:
kalaposfos13 2025-07-09 12:56:24 +02:00 committed by GitHub
parent f5336358ea
commit 84a5bf1bf2
No known key found for this signature in database
GPG key ID: B5690EEEBB952194

View file

@ -163,9 +163,7 @@ static void GenerateEXTRQ(void* /* address */, const ZydisDecodedOperand* operan
mask = (1ULL << length) - 1; mask = (1ULL << length) - 1;
} }
if (length + index > 64) { ASSERT_MSG(length + index <= 64, "length + index must be less than or equal to 64.");
mask = 0xFFFF'FFFF'FFFF'FFFF;
}
// Get lower qword from xmm register // Get lower qword from xmm register
c.vmovq(scratch1, xmm_dst); c.vmovq(scratch1, xmm_dst);
@ -179,8 +177,8 @@ static void GenerateEXTRQ(void* /* address */, const ZydisDecodedOperand* operan
c.mov(scratch2, mask); c.mov(scratch2, mask);
c.and_(scratch1, scratch2); c.and_(scratch1, scratch2);
// Writeback to xmm register, extrq instruction says top 64-bits are undefined but zeroed on // Writeback to xmm register, extrq instruction says top 64-bits are undefined so we don't
// AMD CPUs // care to preserve them
c.vmovq(xmm_dst, scratch1); c.vmovq(xmm_dst, scratch1);
c.pop(scratch2); c.pop(scratch2);
@ -289,9 +287,7 @@ static void GenerateINSERTQ(void* /* address */, const ZydisDecodedOperand* oper
mask_value = (1ULL << length) - 1; mask_value = (1ULL << length) - 1;
} }
if (length + index > 64) { ASSERT_MSG(length + index <= 64, "length + index must be less than or equal to 64.");
mask_value = 0xFFFF'FFFF'FFFF'FFFF;
}
c.vmovq(scratch1, xmm_src); c.vmovq(scratch1, xmm_src);
c.vmovq(scratch2, xmm_dst); c.vmovq(scratch2, xmm_dst);
@ -311,9 +307,8 @@ static void GenerateINSERTQ(void* /* address */, const ZydisDecodedOperand* oper
// dst |= src // dst |= src
c.or_(scratch2, scratch1); c.or_(scratch2, scratch1);
// Insert scratch2 into low 64 bits of dst, upper 64 bits are undefined but zeroed on AMD // Insert scratch2 into low 64 bits of dst, upper 64 bits are unaffected
// CPUs c.vpinsrq(xmm_dst, xmm_dst, scratch2, 0);
c.vmovq(xmm_dst, scratch2);
c.pop(mask); c.pop(mask);
c.pop(scratch2); c.pop(scratch2);
@ -379,7 +374,7 @@ static void GenerateINSERTQ(void* /* address */, const ZydisDecodedOperand* oper
c.and_(scratch2, mask); c.and_(scratch2, mask);
c.or_(scratch2, scratch1); c.or_(scratch2, scratch1);
// Upper 64 bits are undefined in insertq but AMD CPUs zero them // Upper 64 bits are undefined in insertq
c.vmovq(xmm_dst, scratch2); c.vmovq(xmm_dst, scratch2);
c.pop(mask); c.pop(mask);
@ -640,7 +635,6 @@ static bool TryExecuteIllegalInstruction(void* ctx, void* code_address) {
lowQWordDst >>= index; lowQWordDst >>= index;
lowQWordDst &= mask; lowQWordDst &= mask;
memset((u8*)dst + sizeof(u64), 0, sizeof(u64));
memcpy(dst, &lowQWordDst, sizeof(lowQWordDst)); memcpy(dst, &lowQWordDst, sizeof(lowQWordDst));
Common::IncrementRip(ctx, 4); Common::IncrementRip(ctx, 4);
@ -681,7 +675,6 @@ static bool TryExecuteIllegalInstruction(void* ctx, void* code_address) {
lowQWordDst &= ~(mask << index); lowQWordDst &= ~(mask << index);
lowQWordDst |= lowQWordSrc << index; lowQWordDst |= lowQWordSrc << index;
memset((u8*)dst + sizeof(u64), 0, sizeof(u64));
memcpy(dst, &lowQWordDst, sizeof(lowQWordDst)); memcpy(dst, &lowQWordDst, sizeof(lowQWordDst));
Common::IncrementRip(ctx, 4); Common::IncrementRip(ctx, 4);