py/emitnative: Improve Viper register-indexed code for Arm.

This commit lets the Viper code generator use optimised code sequences
for register-indexed load and store operations when generating Arm code.

The existing code defaulted to generic multi-operations code sequences
for Arm code on most cases.  Now optimised implementations are provided
for register-indexed loads and stores of all data sizes, taking at most
two machine opcodes for each operation.

Signed-off-by: Alessandro Gatti <a.gatti@frob.it>

py/asmarm.c

@@ -343,6 +343,12 @@ void asm_arm_ldrh_reg_reg(asm_arm_t *as, uint rd, uint rn) {
     emit_al(as, 0x1d000b0 | (rn << 16) | (rd << 12));
 }
 
+void asm_arm_ldrh_reg_reg_reg(asm_arm_t *as, uint rd, uint rm, uint rn) {
+    // ldrh doesn't support scaled register index
+    emit_al(as, 0x1a00080 | (ASM_ARM_REG_R8 << 12) | rn); // mov r8, rn, lsl #1
+    emit_al(as, 0x19000b0 | (rm << 16) | (rd << 12) | ASM_ARM_REG_R8); // ldrh rd, [rm, r8];
+}
+
 void asm_arm_ldrh_reg_reg_offset(asm_arm_t *as, uint rd, uint rn, uint byte_offset) {
     if (byte_offset < 0x100) {
         // ldrh rd, [rn, #off]
@@ -360,6 +366,16 @@ void asm_arm_ldrb_reg_reg(asm_arm_t *as, uint rd, uint rn) {
     emit_al(as, 0x5d00000 | (rn << 16) | (rd << 12));
 }
 
+void asm_arm_ldrb_reg_reg_reg(asm_arm_t *as, uint rd, uint rm, uint rn) {
+    // ldrb rd, [rm, rn]
+    emit_al(as, 0x7d00000 | (rm << 16) | (rd << 12) | rn);
+}
+
+void asm_arm_ldr_reg_reg_reg(asm_arm_t *as, uint rd, uint rm, uint rn) {
+    // ldr rd, [rm, rn, lsl #2]
+    emit_al(as, 0x7900100 | (rm << 16) | (rd << 12) | rn);
+}
+
 void asm_arm_str_reg_reg(asm_arm_t *as, uint rd, uint rm, uint byte_offset) {
     // str rd, [rm, #off]
     emit_al(as, 0x5800000 | (rm << 16) | (rd << 12) | byte_offset);

py/asmarm.h

@@ -116,6 +116,12 @@ void asm_arm_ldrb_reg_reg(asm_arm_t *as, uint rd, uint rn);
 void asm_arm_str_reg_reg(asm_arm_t *as, uint rd, uint rm, uint byte_offset);
 void asm_arm_strh_reg_reg(asm_arm_t *as, uint rd, uint rm);
 void asm_arm_strb_reg_reg(asm_arm_t *as, uint rd, uint rm);
+
+// load from array
+void asm_arm_ldr_reg_reg_reg(asm_arm_t *as, uint rd, uint rm, uint rn);
+void asm_arm_ldrh_reg_reg_reg(asm_arm_t *as, uint rd, uint rm, uint rn);
+void asm_arm_ldrb_reg_reg_reg(asm_arm_t *as, uint rd, uint rm, uint rn);
+
 // store to array
 void asm_arm_str_reg_reg_reg(asm_arm_t *as, uint rd, uint rm, uint rn);
 void asm_arm_strh_reg_reg_reg(asm_arm_t *as, uint rd, uint rm, uint rn);

py/emitnative.c

@@ -1638,6 +1638,10 @@ static void emit_native_load_subscr(emit_t *emit) {
             switch (vtype_base) {
                 case VTYPE_PTR8: {
                     // pointer to 8-bit memory
+                    #if N_ARM
+                    asm_arm_ldrb_reg_reg_reg(emit->as, REG_RET, REG_ARG_1, reg_index);
+                    break;
+                    #endif
                     // TODO optimise to use thumb ldrb r1, [r2, r3]
                     ASM_ADD_REG_REG(emit->as, REG_ARG_1, reg_index); // add index to base
                     ASM_LOAD8_REG_REG(emit->as, REG_RET, REG_ARG_1); // store value to (base+index)
@@ -1645,7 +1649,10 @@ static void emit_native_load_subscr(emit_t *emit) {
                 }
                 case VTYPE_PTR16: {
                     // pointer to 16-bit memory
-                    #if N_XTENSA || N_XTENSAWIN
+                    #if N_ARM
+                    asm_arm_ldrh_reg_reg_reg(emit->as, REG_RET, REG_ARG_1, reg_index);
+                    break;
+                    #elif N_XTENSA || N_XTENSAWIN
                     asm_xtensa_op_addx2(emit->as, REG_ARG_1, reg_index, REG_ARG_1);
                     asm_xtensa_op_l16ui(emit->as, REG_RET, REG_ARG_1, 0);
                     break;
@@ -1657,7 +1664,10 @@ static void emit_native_load_subscr(emit_t *emit) {
                 }
                 case VTYPE_PTR32: {
                     // pointer to word-size memory
-                    #if N_RV32
+                    #if N_ARM
+                    asm_arm_ldr_reg_reg_reg(emit->as, REG_RET, REG_ARG_1, reg_index);
+                    break;
+                    #elif N_RV32
                     asm_rv32_opcode_slli(emit->as, REG_TEMP2, reg_index, 2);
                     asm_rv32_opcode_cadd(emit->as, REG_ARG_1, REG_TEMP2);
                     asm_rv32_opcode_lw(emit->as, REG_RET, REG_ARG_1, 0);