Type propagation for partial registers needs to be improved.

Hello,

What is the feature you'd like to have?
Binary ninja cannot detect/parse doubles in HLIL/PseudoC view if the binary is an ARM binary

Additional Information:

I want to report a decompilation incapability. I have realized this while solving a challenge from r0. Mirror download link

In this situation we have a 32 bit ARM binary. Anyone can run binary in Raspbian.

This binary has been initializing some doubles. Then uses them. I have realized that Binary Ninja HLIL and Pseudo C views cannot decompile the doubles in exact values:

Disassembly:

00008624  e13f8fe2   adr     r3, data_89b0
00008628  d020c3e1   ldrd    r2, r3, [r3]  {data_89b0}  {0x400921cac083126f}
0000862c  f4214be1   strd    r2, r3, [r11,  #-0x14] {var_18}
00008630  0e3d8fe2   adr     r3, data_89b8

Pseudo C

00008628      int32_t r2;
00008628      int32_t r3;
00008628      r3 = HIGHW(0x400921cac083126f);
00008628      r2 = LOWW(0x400921cac083126f);

Let's cross check this binary in IDA Pro:

Disassembly:

.text:00008624 ADR     R3, dword_89B0
.text:00008628 LDRD    R2, R3, [R3]
.text:0000862C STRD    R2, R3, [R11,#f1]
.text:00008630 ADR     R3, dword_89B8

Pseudo C

  double v20; // [sp+68h] [bp-1Ch]
  double v21; // [sp+70h] [bp-14h]
  int i; // [sp+7Ch] [bp-8h]

  v21 = 3.1415;
  v20 = 6.283;

BTW, I have extracted the pseudo c code and recompiled as x86, then check it through Binary Ninja:

Pseudo C

00001243      double var_58 = ((double)((long double)3.1415000000000002));
00001263      double var_48 = ((double)((long double)9.9000000000000004));
.
.
.
snip code
.
.
.
((long double)((double)((long double)6.2830000000000004)))

Final words

Binary Ninja have problems detecting/parsing doubles if the binary is an ARM binary, but can decompile/detect doubles if the binary is x86 binary

Start by reproducing the reported ARM case from the linked r0 challenge binary and compare the shown LDRD/STRD disassembly with its HLIL and Pseudo C output. Done means the partial-register type propagation represents the initialized values as doubles with their exact values, rather than separate int32_t HIGHW/LOWW values.