ART: Change rrr add and sub for ARM64

OpRegRegImm will fall back to loading a constant into a register
and then doing the operation with three registers. That is, for
example, the case when we allocate large stack frames. However,
the currently chosen operations are add/sub shifted, which does
*not* allow to specify SP (x31 will be interpreted as xzr). Switch
to add/sub extended. There won't be a practical difference, as we
do not call with anything other than 0 shift.

Change-Id: I2b78df9f044d2963e3e890777c855b339952f9f4

3 files changed, 23 insertions, 5 deletions

diff --git a/compiler/dex/quick/arm64/arm64_lir.h b/compiler/dex/quick/arm64/arm64_lir.h
index c1ce03d..3a6012e 100644
--- a/compiler/dex/quick/arm64/arm64_lir.h
+++ b/compiler/dex/quick/arm64/arm64_lir.h
@@ -219,7 +219,8 @@ enum ArmOpcode {
   kA64First = 0,
   kA64Adc3rrr = kA64First,  // adc [00011010000] rm[20-16] [000000] rn[9-5] rd[4-0].
   kA64Add4RRdT,      // add [s001000100] imm_12[21-10] rn[9-5] rd[4-0].
-  kA64Add4rrro,      // add [00001011000] rm[20-16] option[15-13] imm_3[12-10] rn[9-5] rd[4-0].
+  kA64Add4rrro,      // add [00001011000] rm[20-16] imm_6[15-10] rn[9-5] rd[4-0].
+  kA64Add4rrre,      // add [00001011001] rm[20-16] option[15-13] imm_3[12-10] rn[9-5] rd[4-0].
   kA64Adr2xd,        // adr [0] immlo[30-29] [10000] immhi[23-5] rd[4-0].
   kA64And3Rrl,       // and [00010010] N[22] imm_r[21-16] imm_s[15-10] rn[9-5] rd[4-0].
   kA64And4rrro,      // and [00001010] shift[23-22] [N=0] rm[20-16] imm_6[15-10] rn[9-5] rd[4-0].
@@ -328,7 +329,8 @@ enum ArmOpcode {
   kA64Stxr3wrX,      // stxr[11001000000] rs[20-16] [011111] rn[9-5] rt[4-0].
   kA64Stlxr3wrX,     // stlxr[11001000000] rs[20-16] [111111] rn[9-5] rt[4-0].
   kA64Sub4RRdT,      // sub [s101000100] imm_12[21-10] rn[9-5] rd[4-0].
-  kA64Sub4rrro,      // sub [s1001011000] rm[20-16] option[15-13] imm_3[12-10] rn[9-5] rd[4-0].
+  kA64Sub4rrro,      // sub [s1001011000] rm[20-16] imm_6[15-10] rn[9-5] rd[4-0].
+  kA64Sub4rrre,      // sub [s1001011001] rm[20-16] option[15-13] imm_3[12-10] rn[9-5] rd[4-0].
   kA64Subs3rRd,      // subs[s111000100] imm_12[21-10] rn[9-5] rd[4-0].
   kA64Tst3rro,       // tst alias of "ands rzr, arg1, arg2, arg3".
   kA64Ubfm4rrdd,     // ubfm[s10100110] N[22] imm_r[21-16] imm_s[15-10] rn[9-5] rd[4-0].
diff --git a/compiler/dex/quick/arm64/assemble_arm64.cc b/compiler/dex/quick/arm64/assemble_arm64.cc
index c5bd005..fe3bd6a 100644
--- a/compiler/dex/quick/arm64/assemble_arm64.cc
+++ b/compiler/dex/quick/arm64/assemble_arm64.cc
@@ -115,6 +115,10 @@ const ArmEncodingMap Arm64Mir2Lir::EncodingMap[kA64Last] = {
                  kFmtRegR, 4, 0, kFmtRegR, 9, 5, kFmtRegR, 20, 16,
                  kFmtShift, -1, -1, IS_QUAD_OP | REG_DEF0_USE1,
                  "add", "!0r, !1r, !2r!3o", kFixupNone),
+    ENCODING_MAP(WIDE(kA64Add4rrre), SF_VARIANTS(0x0b200000),
+                 kFmtRegROrSp, 4, 0, kFmtRegROrSp, 9, 5, kFmtRegR, 20, 16,
+                 kFmtExtend, -1, -1, IS_QUAD_OP | REG_DEF0_USE12,
+                 "add", "!0r, !1r, !2r!3o", kFixupNone),
     // Note: adr is binary, but declared as tertiary. The third argument is used while doing the
     //   fixups and contains information to identify the adr label.
     ENCODING_MAP(kA64Adr2xd, NO_VARIANTS(0x10000000),
@@ -558,6 +562,10 @@ const ArmEncodingMap Arm64Mir2Lir::EncodingMap[kA64Last] = {
                  kFmtRegR, 4, 0, kFmtRegR, 9, 5, kFmtRegR, 20, 16,
                  kFmtShift, -1, -1, IS_QUAD_OP | REG_DEF0_USE12,
                  "sub", "!0r, !1r, !2r!3o", kFixupNone),
+    ENCODING_MAP(WIDE(kA64Sub4rrre), SF_VARIANTS(0x4b200000),
+                 kFmtRegROrSp, 4, 0, kFmtRegROrSp, 9, 5, kFmtRegR, 20, 16,
+                 kFmtExtend, -1, -1, IS_QUAD_OP | REG_DEF0_USE12,
+                 "sub", "!0r, !1r, !2r!3o", kFixupNone),
     ENCODING_MAP(WIDE(kA64Subs3rRd), SF_VARIANTS(0x71000000),
                  kFmtRegR, 4, 0, kFmtRegROrSp, 9, 5, kFmtBitBlt, 21, 10,
                  kFmtUnused, -1, -1, IS_TERTIARY_OP | REG_DEF0_USE1 | SETS_CCODES,
diff --git a/compiler/dex/quick/arm64/utility_arm64.cc b/compiler/dex/quick/arm64/utility_arm64.cc
index 2254b8b..9b32a46 100644
--- a/compiler/dex/quick/arm64/utility_arm64.cc
+++ b/compiler/dex/quick/arm64/utility_arm64.cc
@@ -660,6 +660,7 @@ LIR* Arm64Mir2Lir::OpRegRegImm64(OpKind op, RegStorage r_dest, RegStorage r_src1
   int32_t log_imm = -1;
   bool is_wide = r_dest.Is64Bit();
   ArmOpcode wide = (is_wide) ? WIDE(0) : UNWIDE(0);
+  int info = 0;
 
   switch (op) {
     case kOpLsl: {
@@ -692,7 +693,14 @@ LIR* Arm64Mir2Lir::OpRegRegImm64(OpKind op, RegStorage r_dest, RegStorage r_src1
         return NewLIR4(opcode | wide, r_dest.GetReg(), r_src1.GetReg(), abs_value >> 12, 1);
       } else {
         log_imm = -1;
-        alt_opcode = (neg) ? kA64Add4rrro : kA64Sub4rrro;
+        alt_opcode = (neg) ? kA64Add4rrre : kA64Sub4rrre;
+        // To make it correct, we need:
+        // 23..21 = 001  = extend
+        // 15..13 = 01x  = LSL/UXTW/X  /  x defines wide or not
+        // 12..10 = 000  = no shift (in case of SP)
+        // => info =   00101x000
+        // =>      =0x 0 5  (0/8)
+        info = (is_wide ? 8 : 0) | 0x50;
       }
       break;
     // case kOpRsub:
@@ -734,7 +742,7 @@ LIR* Arm64Mir2Lir::OpRegRegImm64(OpKind op, RegStorage r_dest, RegStorage r_src1
   if (log_imm >= 0) {
     return NewLIR3(opcode | wide, r_dest.GetReg(), r_src1.GetReg(), log_imm);
   } else {
-    RegStorage r_scratch = AllocTemp();
+    RegStorage r_scratch;
     if (IS_WIDE(wide)) {
       r_scratch = AllocTempWide();
       LoadConstantWide(r_scratch, value);
@@ -743,7 +751,7 @@ LIR* Arm64Mir2Lir::OpRegRegImm64(OpKind op, RegStorage r_dest, RegStorage r_src1
       LoadConstant(r_scratch, value);
     }
     if (EncodingMap[alt_opcode].flags & IS_QUAD_OP)
-      res = NewLIR4(alt_opcode | wide, r_dest.GetReg(), r_src1.GetReg(), r_scratch.GetReg(), 0);
+      res = NewLIR4(alt_opcode | wide, r_dest.GetReg(), r_src1.GetReg(), r_scratch.GetReg(), info);
     else
       res = NewLIR3(alt_opcode | wide, r_dest.GetReg(), r_src1.GetReg(), r_scratch.GetReg());
     FreeTemp(r_scratch);