Renderscript runtime error "failed to read script metadata" on Android 4.1.x phones

Question

I have three Renderscript functions (in three separate RS files) that used to work great on Adroid 4.1.x phones. However, after I upgraded to ADK 18, they started having runtime errors, crashing on the same phones. The errors are like this:

V/RenderScript( 3644): rsContextCreate dev=0x5a2f0fc0
V/RenderScript( 3644): 0x40051010 Launching thread(s), CPUs 0
V/ScriptC ( 3644): Create script for resource = levels
E/bcc     ( 3644): CPU is krait2
E/bcc     ( 3644): Cache dependency levels sha1 mismatch:
E/bcc     ( 3644):   given:  4144ab455aa71df31932ff5baf15583cf1775d72
E/bcc     ( 3644):   cached: d2ab2c2b568810d8ca0c39d730cb2494cae89106
V/ScriptC ( 3644): Create script for resource = sepia
E/bcc     ( 3644): Cache dependency sepia sha1 mismatch:
E/bcc     ( 3644):   given:  6b15396c900dd23c700b8c13e7d5019fb72fcb0e
E/bcc     ( 3644):   cached: 8d011a42147f8d3895db549e0b3fe4a43ed49fe2
V/ScriptC ( 3644): Create script for resource = flip
E/bcinfo  ( 3644): Could not parse bitcode file
E/bcinfo  ( 3644): Invalid SWITCH record
E/RenderScript( 3644): bcinfo: failed to read script metadata
W/dalvikvm( 3644): threadid=1: thread exiting with uncaught exception (group=0x410f4498)
E/AndroidRuntime( 3644): FATAL EXCEPTION: main

It is worth noting that on 4.2.x and newer devices, the Renderscript functions still work.

In the "project.properties" file, I tried sdk.buildtools = 17.0.0, 18.0.1, 18.1.0, 18.1.1, and 19.0.0. None of them worked on 4.1.x phones, though the error message changed slightly with 19.0.0. I googled the error message "failed to read script metadata". Apparently it was added by Stephen Hines. I don't really understand Stephen's changes, so I'd like to ask the following questions:

1. Why were the changes made?
2. What is the most likely reason that my Renderscript functions stopped working on 4.1.x devices?
3. How to fix the crashes?

The flip.rs that's causing the crash is listed below. Suggestions and comments are welcome. Thanks!

#pragma version(1)
#pragma rs java_package_name(com.xxxx.yyyy.zzzz)
#include "rs_time.rsh"

rs_script flipScript;
rs_allocation gIn;
rs_allocation gOut;
int width;
int height;
int direction = 0;

void root(const uchar4 *v_in, uchar4 *v_out, const void *usrData, uint32_t x, uint32_t y) {
    if(direction == 0) { // flip horizontally
        const uchar4 *element = rsGetElementAt(gIn, width - x, y);
        float4 color = rsUnpackColor8888(*element);
        float4 output = {color.r, color.g, color.b};
        *v_out = rsPackColorTo8888(output);
    }
    else if(direction == 1) { // flip vertically
        const uchar4 *element = rsGetElementAt(gIn, x, height - y);
        float4 color = rsUnpackColor8888(*element);
        float4 output = {color.r, color.g, color.b};
        *v_out = rsPackColorTo8888(output);
    }
    else if(direction == 2) { // rotate left
        const uchar4 *element = rsGetElementAt(gIn, width - y, x);
        float4 color = rsUnpackColor8888(*element);
        float4 output = {color.r, color.g, color.b};
        *v_out = rsPackColorTo8888(output);
    }
    else if(direction == 3) { // rotate right
        const uchar4 *element = rsGetElementAt(gIn, y, height - x);
        float4 color = rsUnpackColor8888(*element);
        float4 output = {color.r, color.g, color.b};
        *v_out = rsPackColorTo8888(output);
    }
}

void flip(int testIdx) {
    int64_t  t0, t1;
    int64_t  t;
    t0 = rsUptimeNanos();
    rsForEach(flipScript, gIn, gOut);
    t1 = rsUptimeNanos();
    t = t1 - t0;
    rsDebug("  flip: timer on RS side: ", t);
    timeNanoSec[testIdx] = (float)t;
}

Answer 1

What target API are you compiling for? Can you send me an apk or at least the 3 failing .bc files?

This looks like a bitcode encoding issue for only v16 (JellyBean) devices. It has nothing to do with the changelist you mentioned, but actually has to do with coalescing of ranged values in a switch statement. It is a bug in the compiler frontend, llvm-rs-cc. Libbcc (the on-device compiler) is rightfully rejecting the bad bitcode that a v16 device can't read (namely, the coalesced case ranges). I will file a bug to fix this internally and hopefully get it released in a future SDK update. In the meantime, I suggest you make use of the RenderScript support library, which would bypass this particular bug (since it won't generate v16 bitcode).