Work around missing __mulodi4() on ARMv7 + Clang builds.

__mulodi4() is roughly,

    int64_t __mulodi4(int64_t a, int64_t b, int* overflow) {
        int64_t result = a*b;
        *overflow = did_overflow(result);
        return result;
    }

It's used by Clang's __builtin_smulll_overflow(), which is basically the same
except it swaps the positions of the return value and the out parameter.

DNG SDK in turn uses __builtin_smulll_overflow().

When building ARMv7 binaries with Clang, __mulodi4() turns up as a missing symbol.
That's because it's defined in a separate compiler-rt library.  I have not been
able to link that.  Instead, I've provided our own trivial __builtin_smulll_overflow().

Before this patch both nanobench and DM failed dynamically linking __mulodi4().  After this patch nanobench runs for a while, eventually segfaulting.  DM immediately segfaults.

BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2011073002

Review-Url: https://codereview.chromium.org/2011073002

diff --git a/gyp/dng_sdk.gyp b/gyp/dng_sdk.gyp
index 3d426241e6fc330d57d7655da01839a90d4fb0d5..be0c0f97db1673566d2eb3ef85a88863d093debd 100644 (file)
--- a/gyp/dng_sdk.gyp
+++ b/gyp/dng_sdk.gyp
@@ -143,6 +143,14 @@
     ['skia_os != "linux"', {
       'sources': ['<@(headers)'],
     }],
+    ['skia_arch_type == "arm" and skia_clang_build', {
+      # DNG SDK uses __builtin_smulll_overflow() to detect 64x64 bit multiply overflow.
+      # On ARMv7, Clang implements this with __mulodi4() in libclang_rt.
+      # I can't quite figure out how to link that here, so instead here's a shim for
+      # __builtin_smulll_overflow() that multiplies normally assuming no overflow.
+      # Tracked in b/29412086.
+      'defines': [ '__builtin_smulll_overflow(x,y,p)=(*(p)=(x)*(y), false)' ],
+    }],
   ],
   'dependencies': [
     'libjpeg-turbo-selector.gyp:libjpeg-turbo-selector',