Tom1 wrote on Thu, 17 January 2019 14:12

mirek wrote on Thu, 17 January 2019 14:58

...
It starts to look like Win32 does not like the sharing of HDC... (despite of what the documentation says).

What about this:

void RenderCharacterSys(FontGlyphConsumer& sw, double x, double y, int ch, Font fnt)
{
	RLOG(CoWork::GetWorkerIndex() << " RenderCharSys " << fnt << " " << ch << " " << (char)ch);
	static Atomic h;
	h++;
	HFONT hfont = GetWin32Font(fnt, 0);
	VERIFY(hfont);
	if(hfont) {
		HDC hdc = CreateIC("DISPLAY", NULL, NULL, NULL);
		HFONT ohfont = (HFONT) ::SelectObject(hdc, hfont);
		GLYPHMETRICS gm;
		MAT2 m_matrix;
		memset(&m_matrix, 0, sizeof(m_matrix));
		m_matrix.eM11.value = 1;
		m_matrix.eM22.value = 1;
		int gsz = GetGlyphOutlineW(hdc, ch, GGO_NATIVE|GGO_UNHINTED, &gm, 0, NULL, &m_matrix);
		if(gsz < 0) {
			LOGFONT lf;
			int ret = GetObject(hfont, sizeof(lf), &lf);
			RLOG("Failed " << CoWork::GetWorkerIndex() << " RenderCharSys " << fnt << " " << ch << " " << (char)ch
			               << ", real: "  << " " << lf.lfFaceName << " " << lf.lfHeight);
			RDUMP((void *)hdc);
			RDUMP((void *)hfont);
			RDUMP(GetLastErrorMessage());
		}
		VERIFY(gsz >= 0);
		if(gsz < 0)
			return;
		StringBuffer gb(gsz);
		gsz = GetGlyphOutlineW(hdc, ch, GGO_NATIVE|GGO_UNHINTED, &gm, gsz, ~gb, &m_matrix);
		if(gsz < 0) {
			LOGFONT lf;
			int ret = GetObject(hfont, sizeof(lf), &lf);
			RLOG("Failed2 " << CoWork::GetWorkerIndex() << " RenderCharSys " << fnt << " " << ch << " " << (char)ch
			               << ", real: "  << " " << lf.lfFaceName << " " << lf.lfHeight);
			RDUMP((void *)hdc);
			RDUMP((void *)hfont);
			RDUMP(GetLastErrorMessage());
		}
		VERIFY(gsz >= 0);
		if(gsz < 0)
			return;
		RenderCharPath(~gb, gsz, sw, x, y + fnt.GetAscent());
		::SelectObject(hdc, ohfont);
		::DeleteDC(hdc);
	}
	h--;
	VERIFY(h == 0);
}

As far as I can see, not a single bad or missing letter in 8 minutes of systematic intensive scanning of names. No VERIFYs/ASSERTs either. The log is attached.

BR, Tom

Well, so be it, lets call it a fix. I will remove shared information context from the trunk... There will be small performance penalty, but it should be negligible (as there are caches on many levels anyway).

Hopefully my interpretation is right and we are not just hiding some other bug...