Added a few more integral SIMD extensions for x86/x86_64 (SSE).
Rationale:
ART vectorizer needs a couple of extra SIMD operations before
sending out the larger general CL.
Test: assembler_x86[_64]_test
Bug: 34083438
Change-Id: Id156283424ad311e6109b360efcd409c671cd5b7
diff --git a/compiler/utils/x86/assembler_x86.cc b/compiler/utils/x86/assembler_x86.cc
index 5a466e1..6eab302 100644
--- a/compiler/utils/x86/assembler_x86.cc
+++ b/compiler/utils/x86/assembler_x86.cc
@@ -642,39 +642,6 @@
}
-void X86Assembler::psrldq(XmmRegister reg, const Immediate& shift_count) {
- DCHECK(shift_count.is_uint8());
-
- AssemblerBuffer::EnsureCapacity ensured(&buffer_);
- EmitUint8(0x66);
- EmitUint8(0x0F);
- EmitUint8(0x73);
- EmitXmmRegisterOperand(3, reg);
- EmitUint8(shift_count.value());
-}
-
-
-void X86Assembler::psrlq(XmmRegister reg, const Immediate& shift_count) {
- DCHECK(shift_count.is_uint8());
-
- AssemblerBuffer::EnsureCapacity ensured(&buffer_);
- EmitUint8(0x66);
- EmitUint8(0x0F);
- EmitUint8(0x73);
- EmitXmmRegisterOperand(2, reg);
- EmitUint8(shift_count.value());
-}
-
-
-void X86Assembler::punpckldq(XmmRegister dst, XmmRegister src) {
- AssemblerBuffer::EnsureCapacity ensured(&buffer_);
- EmitUint8(0x66);
- EmitUint8(0x0F);
- EmitUint8(0x62);
- EmitXmmRegisterOperand(dst, src);
-}
-
-
void X86Assembler::addsd(XmmRegister dst, XmmRegister src) {
AssemblerBuffer::EnsureCapacity ensured(&buffer_);
EmitUint8(0xF2);
@@ -828,6 +795,51 @@
}
+void X86Assembler::paddb(XmmRegister dst, XmmRegister src) {
+ AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+ EmitUint8(0x66);
+ EmitUint8(0x0F);
+ EmitUint8(0xFC);
+ EmitXmmRegisterOperand(dst, src);
+}
+
+
+void X86Assembler::psubb(XmmRegister dst, XmmRegister src) {
+ AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+ EmitUint8(0x66);
+ EmitUint8(0x0F);
+ EmitUint8(0xF8);
+ EmitXmmRegisterOperand(dst, src);
+}
+
+
+void X86Assembler::paddw(XmmRegister dst, XmmRegister src) {
+ AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+ EmitUint8(0x66);
+ EmitUint8(0x0F);
+ EmitUint8(0xFD);
+ EmitXmmRegisterOperand(dst, src);
+}
+
+
+void X86Assembler::psubw(XmmRegister dst, XmmRegister src) {
+ AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+ EmitUint8(0x66);
+ EmitUint8(0x0F);
+ EmitUint8(0xF9);
+ EmitXmmRegisterOperand(dst, src);
+}
+
+
+void X86Assembler::pmullw(XmmRegister dst, XmmRegister src) {
+ AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+ EmitUint8(0x66);
+ EmitUint8(0x0F);
+ EmitUint8(0xD5);
+ EmitXmmRegisterOperand(dst, src);
+}
+
+
void X86Assembler::paddd(XmmRegister dst, XmmRegister src) {
AssemblerBuffer::EnsureCapacity ensured(&buffer_);
EmitUint8(0x66);
@@ -856,6 +868,24 @@
}
+void X86Assembler::paddq(XmmRegister dst, XmmRegister src) {
+ AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+ EmitUint8(0x66);
+ EmitUint8(0x0F);
+ EmitUint8(0xD4);
+ EmitXmmRegisterOperand(dst, src);
+}
+
+
+void X86Assembler::psubq(XmmRegister dst, XmmRegister src) {
+ AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+ EmitUint8(0x66);
+ EmitUint8(0x0F);
+ EmitUint8(0xFB);
+ EmitXmmRegisterOperand(dst, src);
+}
+
+
void X86Assembler::cvtsi2ss(XmmRegister dst, Register src) {
AssemblerBuffer::EnsureCapacity ensured(&buffer_);
EmitUint8(0xF3);
@@ -1186,6 +1216,141 @@
}
+void X86Assembler::punpcklbw(XmmRegister dst, XmmRegister src) {
+ AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+ EmitUint8(0x66);
+ EmitUint8(0x0F);
+ EmitUint8(0x60);
+ EmitXmmRegisterOperand(dst, src);
+}
+
+
+void X86Assembler::punpcklwd(XmmRegister dst, XmmRegister src) {
+ AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+ EmitUint8(0x66);
+ EmitUint8(0x0F);
+ EmitUint8(0x61);
+ EmitXmmRegisterOperand(dst, src);
+}
+
+
+void X86Assembler::punpckldq(XmmRegister dst, XmmRegister src) {
+ AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+ EmitUint8(0x66);
+ EmitUint8(0x0F);
+ EmitUint8(0x62);
+ EmitXmmRegisterOperand(dst, src);
+}
+
+
+void X86Assembler::punpcklqdq(XmmRegister dst, XmmRegister src) {
+ AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+ EmitUint8(0x66);
+ EmitUint8(0x0F);
+ EmitUint8(0x6C);
+ EmitXmmRegisterOperand(dst, src);
+}
+
+
+void X86Assembler::psllw(XmmRegister reg, const Immediate& shift_count) {
+ DCHECK(shift_count.is_uint8());
+ AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+ EmitUint8(0x66);
+ EmitUint8(0x0F);
+ EmitUint8(0x71);
+ EmitXmmRegisterOperand(6, reg);
+ EmitUint8(shift_count.value());
+}
+
+
+void X86Assembler::pslld(XmmRegister reg, const Immediate& shift_count) {
+ DCHECK(shift_count.is_uint8());
+ AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+ EmitUint8(0x66);
+ EmitUint8(0x0F);
+ EmitUint8(0x72);
+ EmitXmmRegisterOperand(6, reg);
+ EmitUint8(shift_count.value());
+}
+
+
+void X86Assembler::psllq(XmmRegister reg, const Immediate& shift_count) {
+ DCHECK(shift_count.is_uint8());
+ AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+ EmitUint8(0x66);
+ EmitUint8(0x0F);
+ EmitUint8(0x73);
+ EmitXmmRegisterOperand(6, reg);
+ EmitUint8(shift_count.value());
+}
+
+
+void X86Assembler::psraw(XmmRegister reg, const Immediate& shift_count) {
+ DCHECK(shift_count.is_uint8());
+ AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+ EmitUint8(0x66);
+ EmitUint8(0x0F);
+ EmitUint8(0x71);
+ EmitXmmRegisterOperand(4, reg);
+ EmitUint8(shift_count.value());
+}
+
+
+void X86Assembler::psrad(XmmRegister reg, const Immediate& shift_count) {
+ DCHECK(shift_count.is_uint8());
+ AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+ EmitUint8(0x66);
+ EmitUint8(0x0F);
+ EmitUint8(0x72);
+ EmitXmmRegisterOperand(4, reg);
+ EmitUint8(shift_count.value());
+}
+
+
+void X86Assembler::psrlw(XmmRegister reg, const Immediate& shift_count) {
+ DCHECK(shift_count.is_uint8());
+ AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+ EmitUint8(0x66);
+ EmitUint8(0x0F);
+ EmitUint8(0x71);
+ EmitXmmRegisterOperand(2, reg);
+ EmitUint8(shift_count.value());
+}
+
+
+void X86Assembler::psrld(XmmRegister reg, const Immediate& shift_count) {
+ DCHECK(shift_count.is_uint8());
+ AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+ EmitUint8(0x66);
+ EmitUint8(0x0F);
+ EmitUint8(0x72);
+ EmitXmmRegisterOperand(2, reg);
+ EmitUint8(shift_count.value());
+}
+
+
+void X86Assembler::psrlq(XmmRegister reg, const Immediate& shift_count) {
+ DCHECK(shift_count.is_uint8());
+ AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+ EmitUint8(0x66);
+ EmitUint8(0x0F);
+ EmitUint8(0x73);
+ EmitXmmRegisterOperand(2, reg);
+ EmitUint8(shift_count.value());
+}
+
+
+void X86Assembler::psrldq(XmmRegister reg, const Immediate& shift_count) {
+ DCHECK(shift_count.is_uint8());
+ AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+ EmitUint8(0x66);
+ EmitUint8(0x0F);
+ EmitUint8(0x73);
+ EmitXmmRegisterOperand(3, reg);
+ EmitUint8(shift_count.value());
+}
+
+
void X86Assembler::fldl(const Address& src) {
AssemblerBuffer::EnsureCapacity ensured(&buffer_);
EmitUint8(0xDD);
diff --git a/compiler/utils/x86/assembler_x86.h b/compiler/utils/x86/assembler_x86.h
index 4343e2e..2999599 100644
--- a/compiler/utils/x86/assembler_x86.h
+++ b/compiler/utils/x86/assembler_x86.h
@@ -408,14 +408,9 @@
void movsd(const Address& dst, XmmRegister src);
void movsd(XmmRegister dst, XmmRegister src);
- void psrlq(XmmRegister reg, const Immediate& shift_count);
- void punpckldq(XmmRegister dst, XmmRegister src);
-
void movhpd(XmmRegister dst, const Address& src);
void movhpd(const Address& dst, XmmRegister src);
- void psrldq(XmmRegister reg, const Immediate& shift_count);
-
void addsd(XmmRegister dst, XmmRegister src);
void addsd(XmmRegister dst, const Address& src);
void subsd(XmmRegister dst, XmmRegister src);
@@ -436,10 +431,20 @@
void movdqa(const Address& dst, XmmRegister src); // store aligned
void movdqu(const Address& dst, XmmRegister src); // store unaligned
- void paddd(XmmRegister dst, XmmRegister src); // no addr variant (for now)
+ void paddb(XmmRegister dst, XmmRegister src); // no addr variant (for now)
+ void psubb(XmmRegister dst, XmmRegister src);
+
+ void paddw(XmmRegister dst, XmmRegister src);
+ void psubw(XmmRegister dst, XmmRegister src);
+ void pmullw(XmmRegister dst, XmmRegister src);
+
+ void paddd(XmmRegister dst, XmmRegister src);
void psubd(XmmRegister dst, XmmRegister src);
void pmulld(XmmRegister dst, XmmRegister src);
+ void paddq(XmmRegister dst, XmmRegister src);
+ void psubq(XmmRegister dst, XmmRegister src);
+
void cvtsi2ss(XmmRegister dst, Register src);
void cvtsi2sd(XmmRegister dst, Register src);
@@ -489,6 +494,24 @@
void shufps(XmmRegister dst, XmmRegister src, const Immediate& imm);
void pshufd(XmmRegister dst, XmmRegister src, const Immediate& imm);
+ void punpcklbw(XmmRegister dst, XmmRegister src);
+ void punpcklwd(XmmRegister dst, XmmRegister src);
+ void punpckldq(XmmRegister dst, XmmRegister src);
+ void punpcklqdq(XmmRegister dst, XmmRegister src);
+
+ void psllw(XmmRegister reg, const Immediate& shift_count);
+ void pslld(XmmRegister reg, const Immediate& shift_count);
+ void psllq(XmmRegister reg, const Immediate& shift_count);
+
+ void psraw(XmmRegister reg, const Immediate& shift_count);
+ void psrad(XmmRegister reg, const Immediate& shift_count);
+ // no psraq
+
+ void psrlw(XmmRegister reg, const Immediate& shift_count);
+ void psrld(XmmRegister reg, const Immediate& shift_count);
+ void psrlq(XmmRegister reg, const Immediate& shift_count);
+ void psrldq(XmmRegister reg, const Immediate& shift_count);
+
void flds(const Address& src);
void fstps(const Address& dst);
void fsts(const Address& dst);
diff --git a/compiler/utils/x86/assembler_x86_test.cc b/compiler/utils/x86/assembler_x86_test.cc
index c6ab893..a74bea2 100644
--- a/compiler/utils/x86/assembler_x86_test.cc
+++ b/compiler/utils/x86/assembler_x86_test.cc
@@ -122,18 +122,6 @@
DriverStr(expected, "movntl");
}
-TEST_F(AssemblerX86Test, psrlq) {
- GetAssembler()->psrlq(x86::XMM0, CreateImmediate(32));
- const char* expected = "psrlq $0x20, %xmm0\n";
- DriverStr(expected, "psrlq");
-}
-
-TEST_F(AssemblerX86Test, punpckldq) {
- GetAssembler()->punpckldq(x86::XMM0, x86::XMM1);
- const char* expected = "punpckldq %xmm1, %xmm0\n";
- DriverStr(expected, "punpckldq");
-}
-
TEST_F(AssemblerX86Test, LoadLongConstant) {
GetAssembler()->LoadLongConstant(x86::XMM0, 51);
const char* expected =
@@ -521,6 +509,26 @@
DriverStr(RepeatFF(&x86::X86Assembler::divpd, "divpd %{reg2}, %{reg1}"), "divpd");
}
+TEST_F(AssemblerX86Test, PAddB) {
+ DriverStr(RepeatFF(&x86::X86Assembler::paddb, "paddb %{reg2}, %{reg1}"), "paddb");
+}
+
+TEST_F(AssemblerX86Test, PSubB) {
+ DriverStr(RepeatFF(&x86::X86Assembler::psubb, "psubb %{reg2}, %{reg1}"), "psubb");
+}
+
+TEST_F(AssemblerX86Test, PAddW) {
+ DriverStr(RepeatFF(&x86::X86Assembler::paddw, "paddw %{reg2}, %{reg1}"), "paddw");
+}
+
+TEST_F(AssemblerX86Test, PSubW) {
+ DriverStr(RepeatFF(&x86::X86Assembler::psubw, "psubw %{reg2}, %{reg1}"), "psubw");
+}
+
+TEST_F(AssemblerX86Test, PMullW) {
+ DriverStr(RepeatFF(&x86::X86Assembler::pmullw, "pmullw %{reg2}, %{reg1}"), "pmullw");
+}
+
TEST_F(AssemblerX86Test, PAddD) {
DriverStr(RepeatFF(&x86::X86Assembler::paddd, "paddd %{reg2}, %{reg1}"), "paddd");
}
@@ -533,6 +541,14 @@
DriverStr(RepeatFF(&x86::X86Assembler::pmulld, "pmulld %{reg2}, %{reg1}"), "pmulld");
}
+TEST_F(AssemblerX86Test, PAddQ) {
+ DriverStr(RepeatFF(&x86::X86Assembler::paddq, "paddq %{reg2}, %{reg1}"), "paddq");
+}
+
+TEST_F(AssemblerX86Test, PSubQ) {
+ DriverStr(RepeatFF(&x86::X86Assembler::psubq, "psubq %{reg2}, %{reg1}"), "psubq");
+}
+
TEST_F(AssemblerX86Test, XorPD) {
DriverStr(RepeatFF(&x86::X86Assembler::xorpd, "xorpd %{reg2}, %{reg1}"), "xorpd");
}
@@ -581,6 +597,67 @@
DriverStr(RepeatFFI(&x86::X86Assembler::pshufd, 1, "pshufd ${imm}, %{reg2}, %{reg1}"), "pshufd");
}
+TEST_F(AssemblerX86Test, Punpcklbw) {
+ DriverStr(RepeatFF(&x86::X86Assembler::punpcklbw, "punpcklbw %{reg2}, %{reg1}"), "punpcklbw");
+}
+
+TEST_F(AssemblerX86Test, Punpcklwd) {
+ DriverStr(RepeatFF(&x86::X86Assembler::punpcklwd, "punpcklwd %{reg2}, %{reg1}"), "punpcklwd");
+}
+
+TEST_F(AssemblerX86Test, Punpckldq) {
+ DriverStr(RepeatFF(&x86::X86Assembler::punpckldq, "punpckldq %{reg2}, %{reg1}"), "punpckldq");
+}
+
+TEST_F(AssemblerX86Test, Punpcklqdq) {
+ DriverStr(RepeatFF(&x86::X86Assembler::punpcklqdq, "punpcklqdq %{reg2}, %{reg1}"), "punpcklqdq");
+}
+
+TEST_F(AssemblerX86Test, psllw) {
+ GetAssembler()->psllw(x86::XMM0, CreateImmediate(16));
+ DriverStr("psllw $0x10, %xmm0\n", "psllwi");
+}
+
+TEST_F(AssemblerX86Test, pslld) {
+ GetAssembler()->pslld(x86::XMM0, CreateImmediate(16));
+ DriverStr("pslld $0x10, %xmm0\n", "pslldi");
+}
+
+TEST_F(AssemblerX86Test, psllq) {
+ GetAssembler()->psllq(x86::XMM0, CreateImmediate(16));
+ DriverStr("psllq $0x10, %xmm0\n", "psllqi");
+}
+
+TEST_F(AssemblerX86Test, psraw) {
+ GetAssembler()->psraw(x86::XMM0, CreateImmediate(16));
+ DriverStr("psraw $0x10, %xmm0\n", "psrawi");
+}
+
+TEST_F(AssemblerX86Test, psrad) {
+ GetAssembler()->psrad(x86::XMM0, CreateImmediate(16));
+ DriverStr("psrad $0x10, %xmm0\n", "psradi");
+}
+
+TEST_F(AssemblerX86Test, psrlw) {
+ GetAssembler()->psrlw(x86::XMM0, CreateImmediate(16));
+ DriverStr("psrlw $0x10, %xmm0\n", "psrlwi");
+}
+
+TEST_F(AssemblerX86Test, psrld) {
+ GetAssembler()->psrld(x86::XMM0, CreateImmediate(16));
+ DriverStr("psrld $0x10, %xmm0\n", "psrldi");
+}
+
+TEST_F(AssemblerX86Test, psrlq) {
+ GetAssembler()->psrlq(x86::XMM0, CreateImmediate(16));
+ DriverStr("psrlq $0x10, %xmm0\n", "psrlqi");
+}
+
+TEST_F(AssemblerX86Test, psrldq) {
+ GetAssembler()->psrldq(x86::XMM0, CreateImmediate(16));
+ DriverStr("psrldq $0x10, %xmm0\n", "psrldqi");
+}
+
/////////////////
// Near labels //
/////////////////