MIPS64: Add min/max MSA instructions
Added min_s.df, max_s.df, min_u.df, max_u.df, fmin.df and fmax.df MSA
instructions in assembler, disassembler and tests.
These instructions are needed for min/max support in ART Vectorizer.
Test: mma test-art-host-gtest
Change-Id: I4e8dd18ca501ac09d938a49388e4a43116660ec9
diff --git a/compiler/utils/mips64/assembler_mips64.cc b/compiler/utils/mips64/assembler_mips64.cc
index b8b800a..24900a7 100644
--- a/compiler/utils/mips64/assembler_mips64.cc
+++ b/compiler/utils/mips64/assembler_mips64.cc
@@ -1456,6 +1456,86 @@
EmitMsa3R(0x7, 0x3, wt, ws, wd, 0x10);
}
+void Mips64Assembler::Max_sB(VectorRegister wd, VectorRegister ws, VectorRegister wt) {
+ CHECK(HasMsa());
+ EmitMsa3R(0x2, 0x0, wt, ws, wd, 0xe);
+}
+
+void Mips64Assembler::Max_sH(VectorRegister wd, VectorRegister ws, VectorRegister wt) {
+ CHECK(HasMsa());
+ EmitMsa3R(0x2, 0x1, wt, ws, wd, 0xe);
+}
+
+void Mips64Assembler::Max_sW(VectorRegister wd, VectorRegister ws, VectorRegister wt) {
+ CHECK(HasMsa());
+ EmitMsa3R(0x2, 0x2, wt, ws, wd, 0xe);
+}
+
+void Mips64Assembler::Max_sD(VectorRegister wd, VectorRegister ws, VectorRegister wt) {
+ CHECK(HasMsa());
+ EmitMsa3R(0x2, 0x3, wt, ws, wd, 0xe);
+}
+
+void Mips64Assembler::Max_uB(VectorRegister wd, VectorRegister ws, VectorRegister wt) {
+ CHECK(HasMsa());
+ EmitMsa3R(0x3, 0x0, wt, ws, wd, 0xe);
+}
+
+void Mips64Assembler::Max_uH(VectorRegister wd, VectorRegister ws, VectorRegister wt) {
+ CHECK(HasMsa());
+ EmitMsa3R(0x3, 0x1, wt, ws, wd, 0xe);
+}
+
+void Mips64Assembler::Max_uW(VectorRegister wd, VectorRegister ws, VectorRegister wt) {
+ CHECK(HasMsa());
+ EmitMsa3R(0x3, 0x2, wt, ws, wd, 0xe);
+}
+
+void Mips64Assembler::Max_uD(VectorRegister wd, VectorRegister ws, VectorRegister wt) {
+ CHECK(HasMsa());
+ EmitMsa3R(0x3, 0x3, wt, ws, wd, 0xe);
+}
+
+void Mips64Assembler::Min_sB(VectorRegister wd, VectorRegister ws, VectorRegister wt) {
+ CHECK(HasMsa());
+ EmitMsa3R(0x4, 0x0, wt, ws, wd, 0xe);
+}
+
+void Mips64Assembler::Min_sH(VectorRegister wd, VectorRegister ws, VectorRegister wt) {
+ CHECK(HasMsa());
+ EmitMsa3R(0x4, 0x1, wt, ws, wd, 0xe);
+}
+
+void Mips64Assembler::Min_sW(VectorRegister wd, VectorRegister ws, VectorRegister wt) {
+ CHECK(HasMsa());
+ EmitMsa3R(0x4, 0x2, wt, ws, wd, 0xe);
+}
+
+void Mips64Assembler::Min_sD(VectorRegister wd, VectorRegister ws, VectorRegister wt) {
+ CHECK(HasMsa());
+ EmitMsa3R(0x4, 0x3, wt, ws, wd, 0xe);
+}
+
+void Mips64Assembler::Min_uB(VectorRegister wd, VectorRegister ws, VectorRegister wt) {
+ CHECK(HasMsa());
+ EmitMsa3R(0x5, 0x0, wt, ws, wd, 0xe);
+}
+
+void Mips64Assembler::Min_uH(VectorRegister wd, VectorRegister ws, VectorRegister wt) {
+ CHECK(HasMsa());
+ EmitMsa3R(0x5, 0x1, wt, ws, wd, 0xe);
+}
+
+void Mips64Assembler::Min_uW(VectorRegister wd, VectorRegister ws, VectorRegister wt) {
+ CHECK(HasMsa());
+ EmitMsa3R(0x5, 0x2, wt, ws, wd, 0xe);
+}
+
+void Mips64Assembler::Min_uD(VectorRegister wd, VectorRegister ws, VectorRegister wt) {
+ CHECK(HasMsa());
+ EmitMsa3R(0x5, 0x3, wt, ws, wd, 0xe);
+}
+
void Mips64Assembler::FaddW(VectorRegister wd, VectorRegister ws, VectorRegister wt) {
CHECK(HasMsa());
EmitMsa3R(0x0, 0x0, wt, ws, wd, 0x1b);
@@ -1496,6 +1576,26 @@
EmitMsa3R(0x1, 0x3, wt, ws, wd, 0x1b);
}
+void Mips64Assembler::FmaxW(VectorRegister wd, VectorRegister ws, VectorRegister wt) {
+ CHECK(HasMsa());
+ EmitMsa3R(0x7, 0x0, wt, ws, wd, 0x1b);
+}
+
+void Mips64Assembler::FmaxD(VectorRegister wd, VectorRegister ws, VectorRegister wt) {
+ CHECK(HasMsa());
+ EmitMsa3R(0x7, 0x1, wt, ws, wd, 0x1b);
+}
+
+void Mips64Assembler::FminW(VectorRegister wd, VectorRegister ws, VectorRegister wt) {
+ CHECK(HasMsa());
+ EmitMsa3R(0x6, 0x0, wt, ws, wd, 0x1b);
+}
+
+void Mips64Assembler::FminD(VectorRegister wd, VectorRegister ws, VectorRegister wt) {
+ CHECK(HasMsa());
+ EmitMsa3R(0x6, 0x1, wt, ws, wd, 0x1b);
+}
+
void Mips64Assembler::Ffint_sW(VectorRegister wd, VectorRegister ws) {
CHECK(HasMsa());
EmitMsa2RF(0x19e, 0x0, ws, wd, 0x1e);
diff --git a/compiler/utils/mips64/assembler_mips64.h b/compiler/utils/mips64/assembler_mips64.h
index 9b40645..b212958 100644
--- a/compiler/utils/mips64/assembler_mips64.h
+++ b/compiler/utils/mips64/assembler_mips64.h
@@ -704,6 +704,22 @@
void Aver_uH(VectorRegister wd, VectorRegister ws, VectorRegister wt);
void Aver_uW(VectorRegister wd, VectorRegister ws, VectorRegister wt);
void Aver_uD(VectorRegister wd, VectorRegister ws, VectorRegister wt);
+ void Max_sB(VectorRegister wd, VectorRegister ws, VectorRegister wt);
+ void Max_sH(VectorRegister wd, VectorRegister ws, VectorRegister wt);
+ void Max_sW(VectorRegister wd, VectorRegister ws, VectorRegister wt);
+ void Max_sD(VectorRegister wd, VectorRegister ws, VectorRegister wt);
+ void Max_uB(VectorRegister wd, VectorRegister ws, VectorRegister wt);
+ void Max_uH(VectorRegister wd, VectorRegister ws, VectorRegister wt);
+ void Max_uW(VectorRegister wd, VectorRegister ws, VectorRegister wt);
+ void Max_uD(VectorRegister wd, VectorRegister ws, VectorRegister wt);
+ void Min_sB(VectorRegister wd, VectorRegister ws, VectorRegister wt);
+ void Min_sH(VectorRegister wd, VectorRegister ws, VectorRegister wt);
+ void Min_sW(VectorRegister wd, VectorRegister ws, VectorRegister wt);
+ void Min_sD(VectorRegister wd, VectorRegister ws, VectorRegister wt);
+ void Min_uB(VectorRegister wd, VectorRegister ws, VectorRegister wt);
+ void Min_uH(VectorRegister wd, VectorRegister ws, VectorRegister wt);
+ void Min_uW(VectorRegister wd, VectorRegister ws, VectorRegister wt);
+ void Min_uD(VectorRegister wd, VectorRegister ws, VectorRegister wt);
void FaddW(VectorRegister wd, VectorRegister ws, VectorRegister wt);
void FaddD(VectorRegister wd, VectorRegister ws, VectorRegister wt);
@@ -713,6 +729,10 @@
void FmulD(VectorRegister wd, VectorRegister ws, VectorRegister wt);
void FdivW(VectorRegister wd, VectorRegister ws, VectorRegister wt);
void FdivD(VectorRegister wd, VectorRegister ws, VectorRegister wt);
+ void FmaxW(VectorRegister wd, VectorRegister ws, VectorRegister wt);
+ void FmaxD(VectorRegister wd, VectorRegister ws, VectorRegister wt);
+ void FminW(VectorRegister wd, VectorRegister ws, VectorRegister wt);
+ void FminD(VectorRegister wd, VectorRegister ws, VectorRegister wt);
void Ffint_sW(VectorRegister wd, VectorRegister ws);
void Ffint_sD(VectorRegister wd, VectorRegister ws);
diff --git a/compiler/utils/mips64/assembler_mips64_test.cc b/compiler/utils/mips64/assembler_mips64_test.cc
index fbebe0c..bdf9598 100644
--- a/compiler/utils/mips64/assembler_mips64_test.cc
+++ b/compiler/utils/mips64/assembler_mips64_test.cc
@@ -2998,6 +2998,86 @@
"aver_u.d");
}
+TEST_F(AssemblerMIPS64Test, Max_sB) {
+ DriverStr(RepeatVVV(&mips64::Mips64Assembler::Max_sB, "max_s.b ${reg1}, ${reg2}, ${reg3}"),
+ "max_s.b");
+}
+
+TEST_F(AssemblerMIPS64Test, Max_sH) {
+ DriverStr(RepeatVVV(&mips64::Mips64Assembler::Max_sH, "max_s.h ${reg1}, ${reg2}, ${reg3}"),
+ "max_s.h");
+}
+
+TEST_F(AssemblerMIPS64Test, Max_sW) {
+ DriverStr(RepeatVVV(&mips64::Mips64Assembler::Max_sW, "max_s.w ${reg1}, ${reg2}, ${reg3}"),
+ "max_s.w");
+}
+
+TEST_F(AssemblerMIPS64Test, Max_sD) {
+ DriverStr(RepeatVVV(&mips64::Mips64Assembler::Max_sD, "max_s.d ${reg1}, ${reg2}, ${reg3}"),
+ "max_s.d");
+}
+
+TEST_F(AssemblerMIPS64Test, Max_uB) {
+ DriverStr(RepeatVVV(&mips64::Mips64Assembler::Max_uB, "max_u.b ${reg1}, ${reg2}, ${reg3}"),
+ "max_u.b");
+}
+
+TEST_F(AssemblerMIPS64Test, Max_uH) {
+ DriverStr(RepeatVVV(&mips64::Mips64Assembler::Max_uH, "max_u.h ${reg1}, ${reg2}, ${reg3}"),
+ "max_u.h");
+}
+
+TEST_F(AssemblerMIPS64Test, Max_uW) {
+ DriverStr(RepeatVVV(&mips64::Mips64Assembler::Max_uW, "max_u.w ${reg1}, ${reg2}, ${reg3}"),
+ "max_u.w");
+}
+
+TEST_F(AssemblerMIPS64Test, Max_uD) {
+ DriverStr(RepeatVVV(&mips64::Mips64Assembler::Max_uD, "max_u.d ${reg1}, ${reg2}, ${reg3}"),
+ "max_u.d");
+}
+
+TEST_F(AssemblerMIPS64Test, Min_sB) {
+ DriverStr(RepeatVVV(&mips64::Mips64Assembler::Min_sB, "min_s.b ${reg1}, ${reg2}, ${reg3}"),
+ "min_s.b");
+}
+
+TEST_F(AssemblerMIPS64Test, Min_sH) {
+ DriverStr(RepeatVVV(&mips64::Mips64Assembler::Min_sH, "min_s.h ${reg1}, ${reg2}, ${reg3}"),
+ "min_s.h");
+}
+
+TEST_F(AssemblerMIPS64Test, Min_sW) {
+ DriverStr(RepeatVVV(&mips64::Mips64Assembler::Min_sW, "min_s.w ${reg1}, ${reg2}, ${reg3}"),
+ "min_s.w");
+}
+
+TEST_F(AssemblerMIPS64Test, Min_sD) {
+ DriverStr(RepeatVVV(&mips64::Mips64Assembler::Min_sD, "min_s.d ${reg1}, ${reg2}, ${reg3}"),
+ "min_s.d");
+}
+
+TEST_F(AssemblerMIPS64Test, Min_uB) {
+ DriverStr(RepeatVVV(&mips64::Mips64Assembler::Min_uB, "min_u.b ${reg1}, ${reg2}, ${reg3}"),
+ "min_u.b");
+}
+
+TEST_F(AssemblerMIPS64Test, Min_uH) {
+ DriverStr(RepeatVVV(&mips64::Mips64Assembler::Min_uH, "min_u.h ${reg1}, ${reg2}, ${reg3}"),
+ "min_u.h");
+}
+
+TEST_F(AssemblerMIPS64Test, Min_uW) {
+ DriverStr(RepeatVVV(&mips64::Mips64Assembler::Min_uW, "min_u.w ${reg1}, ${reg2}, ${reg3}"),
+ "min_u.w");
+}
+
+TEST_F(AssemblerMIPS64Test, Min_uD) {
+ DriverStr(RepeatVVV(&mips64::Mips64Assembler::Min_uD, "min_u.d ${reg1}, ${reg2}, ${reg3}"),
+ "min_u.d");
+}
+
TEST_F(AssemblerMIPS64Test, FaddW) {
DriverStr(RepeatVVV(&mips64::Mips64Assembler::FaddW, "fadd.w ${reg1}, ${reg2}, ${reg3}"),
"fadd.w");
@@ -3038,6 +3118,26 @@
"fdiv.d");
}
+TEST_F(AssemblerMIPS64Test, FmaxW) {
+ DriverStr(RepeatVVV(&mips64::Mips64Assembler::FmaxW, "fmax.w ${reg1}, ${reg2}, ${reg3}"),
+ "fmax.w");
+}
+
+TEST_F(AssemblerMIPS64Test, FmaxD) {
+ DriverStr(RepeatVVV(&mips64::Mips64Assembler::FmaxD, "fmax.d ${reg1}, ${reg2}, ${reg3}"),
+ "fmax.d");
+}
+
+TEST_F(AssemblerMIPS64Test, FminW) {
+ DriverStr(RepeatVVV(&mips64::Mips64Assembler::FminW, "fmin.w ${reg1}, ${reg2}, ${reg3}"),
+ "fmin.w");
+}
+
+TEST_F(AssemblerMIPS64Test, FminD) {
+ DriverStr(RepeatVVV(&mips64::Mips64Assembler::FminD, "fmin.d ${reg1}, ${reg2}, ${reg3}"),
+ "fmin.d");
+}
+
TEST_F(AssemblerMIPS64Test, Ffint_sW) {
DriverStr(RepeatVV(&mips64::Mips64Assembler::Ffint_sW, "ffint_s.w ${reg1}, ${reg2}"),
"ffint_s.w");
diff --git a/disassembler/disassembler_mips.cc b/disassembler/disassembler_mips.cc
index 8894cc9..91203cb 100644
--- a/disassembler/disassembler_mips.cc
+++ b/disassembler/disassembler_mips.cc
@@ -438,10 +438,16 @@
{ kMsaMask | (0x7 << 23), kMsa | (0x5 << 23) | 0x10, "ave_u", "Vkmn" },
{ kMsaMask | (0x7 << 23), kMsa | (0x6 << 23) | 0x10, "aver_s", "Vkmn" },
{ kMsaMask | (0x7 << 23), kMsa | (0x7 << 23) | 0x10, "aver_u", "Vkmn" },
+ { kMsaMask | (0x7 << 23), kMsa | (0x2 << 23) | 0xe, "max_s", "Vkmn" },
+ { kMsaMask | (0x7 << 23), kMsa | (0x3 << 23) | 0xe, "max_u", "Vkmn" },
+ { kMsaMask | (0x7 << 23), kMsa | (0x4 << 23) | 0xe, "min_s", "Vkmn" },
+ { kMsaMask | (0x7 << 23), kMsa | (0x5 << 23) | 0xe, "min_u", "Vkmn" },
{ kMsaMask | (0xf << 22), kMsa | (0x0 << 22) | 0x1b, "fadd", "Ukmn" },
{ kMsaMask | (0xf << 22), kMsa | (0x1 << 22) | 0x1b, "fsub", "Ukmn" },
{ kMsaMask | (0xf << 22), kMsa | (0x2 << 22) | 0x1b, "fmul", "Ukmn" },
{ kMsaMask | (0xf << 22), kMsa | (0x3 << 22) | 0x1b, "fdiv", "Ukmn" },
+ { kMsaMask | (0xf << 22), kMsa | (0xe << 22) | 0x1b, "fmax", "Ukmn" },
+ { kMsaMask | (0xf << 22), kMsa | (0xc << 22) | 0x1b, "fmin", "Ukmn" },
{ kMsaMask | (0x1ff << 17), kMsa | (0x19e << 17) | 0x1e, "ffint_s", "ukm" },
{ kMsaMask | (0x1ff << 17), kMsa | (0x19c << 17) | 0x1e, "ftint_s", "ukm" },
{ kMsaMask | (0x7 << 23), kMsa | (0x0 << 23) | 0xd, "sll", "Vkmn" },
