arm: arm backend code gen fixes

diff --git a/orc-test/orctest.c b/orc-test/orctest.c
index 2a1373a..cdcea6b 100644 (file)
--- a/orc-test/orctest.c
+++ b/orc-test/orctest.c
@@ -174,7 +174,7 @@ orc_test_gcc_compile_neon (OrcProgram *p)
   sprintf(dump_filename, "%s-dump.bin", base);
   sprintf(dump_dis_filename, "%s-dump.dis", base);
 
-  target = orc_target_get_by_name ("neon");
+  target = orc_target_get_by_name ("arm");
   flags = orc_target_get_default_flags (target);
   flags |= ORC_TARGET_NEON_CLEAN_COMPILE;
 

diff --git a/orc/orcarm.c b/orc/orcarm.c
index 6bfea2b..19968a1 100644 (file)
--- a/orc/orcarm.c
+++ b/orc/orcarm.c
@@ -389,19 +389,19 @@ orc_arm_emit_cmp_imm (OrcCompiler *compiler, int src1, int value)
 /*    1
  *  1 0 9 8 7 6 5 4 3 2 1 0
  * +-+-+-+-+-+-+-+-+-+-+-+-+
- * |    Si   | St  |0| Rm  |
+ * |    Si   | St|0|   Rm  |
  * +-+-+-+-+-+-+-+-+-+-+-+-+
  */
-#define arm_so_rsi(Si,St,Rm)   ((((Si)&31)<<7)|(((St)&7)<<5)|((Rm)&7))
+#define arm_so_rsi(Si,St,Rm)   ((((Si)&31)<<7)|(((St)&3)<<5)|((Rm)&15))
 #define arm_so_rrx(Rm)         arm_so_rsi(0,ORC_ARM_ROR,Rm)
 #define arm_so_r(Rm)           ((Rm)&15)
 /*    1
  *  1 0 9 8 7 6 5 4 3 2 1 0
  * +-+-+-+-+-+-+-+-+-+-+-+-+
- * |  Rs   |0| St  |1| Rm  |
+ * |  Rs   |0| St|1|   Rm  |
  * +-+-+-+-+-+-+-+-+-+-+-+-+
  */
-#define arm_so_rsr(Rs,St,Rm)   (0x008|(((Rs)&15)<<8)|(((St)&7)<<5)|((Rm)&7))
+#define arm_so_rsr(Rs,St,Rm)   (0x010|(((Rs)&15)<<8)|(((St)&3)<<5)|((Rm)&15))
 
 /* data processing instructions */
 /*    3   2 2 2 2 2     2 2 1     1 1     1   1
@@ -523,7 +523,7 @@ orc_arm_emit_dp (OrcCompiler *p, int type, OrcArmCond cond, OrcArmDP opcode,
  * | cond  |      mode     |   Rn  |  Rd   |0 0 0 0|  op   |  Rm   |
  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
  */
-#define arm_code_par(cond,mode,Rn,Rd,op,Rm) (((cond)<<28)|((mode)<<20)|((Rn)<<16)|((Rd)<<12)|((op)<<4)|(Rm))
+#define arm_code_par(cond,mode,Rn,Rd,op,Rm) (((cond)<<28)|((mode)<<20)|(((Rn)&0xf)<<16)|(((Rd)&0xf)<<12)|((op)<<4)|((Rm)&0xf)|0xf00)
 
 void
 orc_arm_emit_par (OrcCompiler *p, int op, int mode, OrcArmCond cond,
@@ -559,7 +559,7 @@ orc_arm_emit_par (OrcCompiler *p, int op, int mode, OrcArmCond cond,
  * | cond  |0 1 1 0 0 0 0 0|   Rn  |  Rd   |rot|0 0|0 1 1 1|  Rm   |
  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
  */
-#define arm_code_xt(op,cond,Rn,Rd,r8,Rm) (op|((cond)<<28)|((Rn)<<16)|((Rd)<<12)|(((r8)&0x18)<<7)|(Rm))
+#define arm_code_xt(op,cond,Rn,Rd,r8,Rm) (op|((cond)<<28)|(((Rn)&0xf)<<16)|(((Rd)&0xf)<<12)|((((r8)&0xf)&0x18)<<7)|((Rm)&0xf))
 
 void
 orc_arm_emit_xt (OrcCompiler *p, int op, OrcArmCond cond,
@@ -599,7 +599,7 @@ orc_arm_emit_xt (OrcCompiler *p, int op, OrcArmCond cond,
   orc_arm_emit (p, code);
 }
 
-#define arm_code_pkh(op,cond,Rn,Rd,sh,Rm) (op|((cond)<<28)|((Rn)<<16)|((Rd)<<12)|((sh)<<7)|(Rm))
+#define arm_code_pkh(op,cond,Rn,Rd,sh,Rm) (op|((cond)<<28)|(((Rn)&0xf)<<16)|(((Rd)&0xf)<<12)|((sh)<<7)|((Rm)&0xf))
 void
 orc_arm_emit_pkh (OrcCompiler *p, int op, OrcArmCond cond,
     int Rd, int Rn, int Rm, int sh)
@@ -657,7 +657,11 @@ orc_arm_emit_sat (OrcCompiler *p, int op, OrcArmCond cond,
   if (op < 2) {
     code = arm_code_sat (sat_opcodes[op], cond, sat, Rd, sh, asr, Rm);
   } else {
-    code = arm_code_par (cond, par_mode[op], sat, Rd, par_op[op], Rm);
+    if (op == 3) {
+      code = arm_code_par (cond, par_mode[op], sat, Rd, par_op[op], Rm);
+    } else {
+      code = arm_code_par (cond, par_mode[op], sat - 1, Rd, par_op[op], Rm);
+    }
   }
   ORC_ASM_CODE(p,"  %s%s %s, #%d, %s%s\n",
       sat_insn_names[op], orc_arm_cond_name(cond),
@@ -674,7 +678,7 @@ orc_arm_emit_rv (OrcCompiler *p, int op, OrcArmCond cond,
     int Rd, int Rm)
 {
   orc_uint32 code;
-  static const orc_uint32 rv_opcodes[] = { 0x06b00030, 0x06e000b0 };
+  static const orc_uint32 rv_opcodes[] = { 0x06b00030, 0x06bf0fb0 };
   static const char *rv_insn_names[] = { "rev", "rev16" };
 
   code = arm_code_rv (rv_opcodes[op], cond, Rd, Rm);

diff --git a/orc/orcprogram-arm.c b/orc/orcprogram-arm.c
index 4573630..069612b 100644 (file)
--- a/orc/orcprogram-arm.c
+++ b/orc/orcprogram-arm.c
@@ -115,7 +115,6 @@ orc_compiler_orc_arm_init (OrcCompiler *compiler)
   for(i=ORC_GP_REG_BASE;i<ORC_GP_REG_BASE+9;i++){
     compiler->valid_regs[i] = 1;
   }
-  compiler->valid_regs[ORC_ARM_A1] = 0;
   //compiler->valid_regs[ORC_ARM_SB] = 0;
   compiler->valid_regs[ORC_ARM_IP] = 0;
   compiler->valid_regs[ORC_ARM_SP] = 0;
@@ -129,6 +128,12 @@ orc_compiler_orc_arm_init (OrcCompiler *compiler)
     compiler->alloc_regs[i] = 0;
     compiler->used_regs[i] = 0;
   }
+  compiler->exec_reg = ORC_ARM_A1;
+  compiler->valid_regs[compiler->exec_reg] = 0;
+  compiler->gp_tmpreg = ORC_ARM_A2;
+  compiler->valid_regs[compiler->gp_tmpreg] = 0;
+  compiler->tmpreg = ORC_ARM_A3;
+  compiler->valid_regs[compiler->tmpreg] = 0;
 
   compiler->loop_shift = 0;
 }

diff --git a/orc/orcrules-arm.c b/orc/orcrules-arm.c
index 8774f3b..4d2fcee 100644 (file)
--- a/orc/orcrules-arm.c
+++ b/orc/orcrules-arm.c
@@ -35,18 +35,48 @@ arm_rule_ ## opcode (OrcCompiler *p, void *user, OrcInstruction *insn) \
 }
 
 /* multiplies */
-#define orc_arm_smulxy(cond,x,y,Rd,Rm,Rs) (0x01600080|((cond)<<28)|((Rd)<<16)|((Rs)<<8)|((y)<<6)|((x)<<5)|(Rm))
-#define orc_arm_emit_smulbb(p,cond,Rd,Rm,Rs) orc_arm_emit (p, orc_arm_smulxy (cond,0,0,Rd,Rm,Rs))
-#define orc_arm_emit_smulbt(p,cond,Rd,Rm,Rs) orc_arm_emit (p, orc_arm_smulxy (cond,0,1,Rd,Rm,Rs))
-#define orc_arm_emit_smultb(p,cond,Rd,Rm,Rs) orc_arm_emit (p, orc_arm_smulxy (cond,1,0,Rd,Rm,Rs))
-#define orc_arm_emit_smultt(p,cond,Rd,Rm,Rs) orc_arm_emit (p, orc_arm_smulxy (cond,1,1,Rd,Rm,Rs))
-
-#define orc_arm_mul(cond,S,Rd,Rm,Rs) (0x00000090|((cond)<<28)|((S)<<20)|((Rd)<<16)|((Rs)<<8)|(Rm))
-#define orc_arm_emit_mul(p,cond,S,Rd,Rm,Rs) orc_arm_emit (p, orc_arm_mul (cond,S,Rd,Rm,Rs))
-
-#define orc_arm_mull(op,cond,S,RdH,RdL,Rm,Rs) (op|((cond)<<28)|((S)<<20)|((RdH)<<16)|((RdL)<<12)|((Rs)<<8)|(Rm))
-#define orc_arm_emit_smull(p,cond,S,RdL,RdH,Rm,Rs) orc_arm_emit(p,orc_arm_mull (0x00c00090,cond,S,RdH,RdL,Rm,Rs))
-#define orc_arm_emit_umull(p,cond,S,RdL,RdH,Rm,Rs) orc_arm_emit(p,orc_arm_mull (0x00800090,cond,S,RdH,RdL,Rm,Rs))
+#define orc_arm_smulxy(cond,x,y,Rd,Rm,Rs) (0x01600080|((cond)<<28)|(((Rd)&15)<<16)|(((Rs)&15)<<8)|((y)<<6)|((x)<<5)|((Rm)&15))
+#define orc_arm_emit_smulbb(p,cond,Rd,Rm,Rs) do { \
+  ORC_ASM_CODE (p, "smulbb %s, %s, %s\n", orc_arm_reg_name (Rd), \
+      orc_arm_reg_name (Rm), orc_arm_reg_name(Rs)); \
+  orc_arm_emit (p, orc_arm_smulxy (cond,0,0,Rd,Rm,Rs)); \
+} while (0)
+#define orc_arm_emit_smulbt(p,cond,Rd,Rm,Rs) do { \
+  ORC_ASM_CODE (p, "smulbt %s, %s, %s\n", orc_arm_reg_name (Rd), \
+      orc_arm_reg_name (Rm), orc_arm_reg_name(Rs)); \
+  orc_arm_emit (p, orc_arm_smulxy (cond,0,1,Rd,Rm,Rs)); \
+} while (0)
+#define orc_arm_emit_smultb(p,cond,Rd,Rm,Rs) do { \
+  ORC_ASM_CODE (p, "smultb %s, %s, %s\n", orc_arm_reg_name (Rd), \
+      orc_arm_reg_name (Rm), orc_arm_reg_name(Rs)); \
+  orc_arm_emit (p, orc_arm_smulxy (cond,1,0,Rd,Rm,Rs)); \
+} while (0)
+#define orc_arm_emit_smultt(p,cond,Rd,Rm,Rs) do { \
+  ORC_ASM_CODE (p, "smultt %s, %s, %s\n", orc_arm_reg_name (Rd), \
+      orc_arm_reg_name (Rm), orc_arm_reg_name(Rs)); \
+  orc_arm_emit (p, orc_arm_smulxy (cond,1,1,Rd,Rm,Rs)); \
+} while (0)
+
+#define orc_arm_mul(cond,S,Rd,Rm,Rs) (0x00000090|((cond)<<28)|((S)<<20)|(((Rd)&15)<<16)|(((Rs)&15)<<8)|((Rm)&15))
+#define orc_arm_emit_mul(p,cond,S,Rd,Rm,Rs) do { \
+  ORC_ASM_CODE (p, "mul %s, %s, %s\n", orc_arm_reg_name (Rd), \
+      orc_arm_reg_name (Rm), orc_arm_reg_name(Rs)); \
+  orc_arm_emit (p, orc_arm_mul (cond,S,Rd,Rm,Rs)); \
+} while (0)
+
+#define orc_arm_mull(op,cond,S,RdH,RdL,Rm,Rs) (op|((cond)<<28)|((S)<<20)|(((RdH)&15)<<16)|(((RdL)&15)<<12)|(((Rs)&15)<<8)|((Rm)&15))
+#define orc_arm_emit_smull(p,cond,S,RdL,RdH,Rm,Rs) do { \
+  ORC_ASM_CODE (p, "mull %s, %s, %s, %s\n", orc_arm_reg_name (RdH), \
+      orc_arm_reg_name (RdL), \
+      orc_arm_reg_name (Rm), orc_arm_reg_name(Rs)); \
+  orc_arm_emit(p,orc_arm_mull (0x00c00090,cond,S,RdH,RdL,Rm,Rs)); \
+} while (0)
+#define orc_arm_emit_umull(p,cond,S,RdL,RdH,Rm,Rs) do { \
+  ORC_ASM_CODE (p, "umull %s, %s, %s, %s\n", orc_arm_reg_name (RdH), \
+      orc_arm_reg_name (RdL), \
+      orc_arm_reg_name (Rm), orc_arm_reg_name(Rs)); \
+  orc_arm_emit(p,orc_arm_mull (0x00800090,cond,S,RdH,RdL,Rm,Rs)); \
+} while (0)
 
 void
 orc_arm_loadw (OrcCompiler *compiler, int dest, int src1, int offset)
@@ -517,7 +547,7 @@ arm_rule_shrsX (OrcCompiler *p, void *user, OrcInstruction *insn)
   int src1 = ORC_SRC_ARG (p, insn, 0);
   int dest = ORC_DEST_ARG (p, insn, 0);
   int mask = p->tmpreg;
-  int tmp = ORC_ARM_IP;
+  int tmp = ORC_ARM_V8;
   int src2type = ORC_SRC_TYPE (p, insn, 1);
   int size = ORC_PTR_TO_INT(user);
   int loop = 4 / size; /* number of items in one register */