Incorrect return values for pow(-1,nan) and tgamma(-inf) (IDFGH-5559)

[dpgeorge]

Environment

• Module or chip used: ESP32-PICO-D4 and ESP32C3
• IDF version: v4.4-dev-2128-gf45b60b28
• Build System: idf.py
• Compiler version: xtensa-esp32-elf-gcc (crosstool-NG esp-2021r1) 8.4.0
• Operating System: Linux

Problem Description

The return value of pow(-1, NAN) and tgamma(-INFINITY) are incorrect (according to standard C function semantics).

Expected Behavior

• pow(-1, NAN) should return NAN; see https://pubs.opengroup.org/onlinepubs/9699919799/functions/pow.html
• tgamma(-INFINITY) should return NAN; see https://pubs.opengroup.org/onlinepubs/9699919799/functions/tgamma.html

Actual Behavior

• pow(-1, NAN) returns -1.000
• tgamma(-INFINITY) returns INFINITY

Steps to reproduce / code

Create a simple program (eg based on hello_world) and add these lines:

    printf("pow(-1, nan) = %f\n", pow(-1, NAN));
    printf("tgamma(-inf) = %f\n", tgamma(-INFINITY));

Then compile, flash, run, and inspect output.

[0xjakob]

@dpgeorge I think I found the issue with the pow() function. It seems to be a toolchain bug in newlib which has been fixed in newer versions than the one IDF uses. I'll check tgamma() as well. We'll probably provide a quick fix for IDF versions and/or include the fix in the newer IDF toolchain releases.

[0xjakob]

It's been fixed, but only very recently: https://sourceware.org/git/?p=newlib-cygwin.git;a=commit;h=bb25dd1b0f39f343b764fd0db861cb9a30441407

[zikalino]

@dpgeorge we should have a fix for this issue very soon, as we will be upgrading toolchains to 2022r1.

[0xjakob]

@dpgeorge Note that the fix for tgamma() is only available on v5.0 (will be released soon). The other versions are stuck with an older toolchain version which brings an older newlib version that can't easily be fixed. In contrast, pow() has been fixed on v4.4 as well.

[dpgeorge]

OK, thanks for fixing!