Add ARM64EC runtime test coverage

[StephanTLavavej]

Followup to #5815.

• Runtime test coverage for ARM64EC.
  • Like the ARM64 runtime test coverage, I'm using 10 shards, which I've discovered is our current maximum (I am unsure how to increase it further).
• Compile with -fuse-ld=link, link with /machine:arm64ec.
  • Filed for followup: Investigate ARM64EC runtime test failures #5825
• Add ARM64EC workarounds to P0811R3_midpoint_lerp.
  • Updated for followup: ARM64/ARM64EC runtime assertions in P0811R3_midpoint_lerp #5685
• Fix Dev09_158457_tr1_mem_fn_calling_conventions for ARM64EC.
  • The expected output didn't consider ARM64EC, although the product code and the test code correctly avoided emitting __vectorcall for ARM64EC:

    STL/stl/inc/type_traits

    Lines 396 to 397 in defc75b

    	#if ((defined(_M_IX86) && _M_IX86_FP >= 2) || (defined(_M_X64) && !defined(_M_ARM64EC))) && !defined(_M_CEE)
    	#define _EMIT_VECTORCALL(FUNC, OPT1, OPT2, OPT3) FUNC(__vectorcall, OPT1, OPT2, OPT3)

    STL/tests/std/tests/Dev09_158457_tr1_mem_fn_calling_conventions/test.cpp

    Lines 67 to 68 in defc75b

    	#if (defined(_M_IX86) || (defined(_M_X64) && !defined(_M_ARM64EC))) && !defined(_M_CEE)
    	int __vectorcall g() {

    STL/tests/std/tests/Dev09_158457_tr1_mem_fn_calling_conventions/test.cpp

    Lines 118 to 119 in defc75b

    	#if (defined(_M_IX86) || (defined(_M_X64) && !defined(_M_ARM64EC))) && !defined(_M_CEE)
    	int __vectorcall z(int i) {

• Work around <stacktrace> bug.
  • Filed for followup: <stacktrace>: ARM64EC isn't working properly #5830
• Simplify P0881R7_stacktrace to a single thread.
  • I spent a lot of time being confused why the output was duplicated and sometimes interleaved. The internal implementation of <stacktrace> takes a lock, which is why this test was spinning up a second thread, but I think this isn't worth it. If we really wanted to stress the lock, it would have to look very different, and we haven't had issues with this (nor is the code rapidly churning). With @AlexGuteniev's blessing I'm changing this to be single-threaded.
• Add /OPT:REF,NOICF to universal_prefix.lst, fixing P0896R4_P1614R2_comparisons.
  • After printf-instrumenting this test, I found that the issue was attempting to verify that &f1 and &f2 (the addresses of empty functions) were non-equal, and these functions were getting ICFed for ARM64EC:

    STL/tests/std/tests/P0896R4_P1614R2_comparisons/test.cpp

    Lines 397 to 398 in defc75b

    	void f1() {}
    	void f2() {}

    STL/tests/std/tests/P0896R4_P1614R2_comparisons/test.cpp

    Line 479 in defc75b

    test_equality_comparable(&f1, &f2, strong_ordering::less); // This means "not equal"

  • This affects all tests (std/tr1/libcxx) for all architectures. I'm actually surprised we haven't encountered this before, but the linker docs clearly say "By default, /OPT:ICF is enabled by the linker unless /OPT:NOICF or /DEBUG is specified."
  • /OPT:REF,NOICF is what we use for debug builds of the product itself. There's no reason to emit test executables that are more bloated than necessary (REF should be very cheap in terms of CPU time, and possibly even saving I/O time). We never want ICF for test executables, as it impacts correctness in this corner case.
• Simplify names to 'Build ARCH' and 'Test ARCH'.
  • This is a cosmetic improvement. All of the Build stages just build, and all of the Test stages actually execute tests (the fact that they also build is implied, although they don't build quite as thoroughly - no /analyze for the STL's own sources and no benchmarks).
• Distinguish the Cross builds.
  • For clarity, since native is the default expectation. Will be helpful if we ever encounter errors that are specific to a host arch.
• Add an early ARM64-native build.
  • This is partially to make toolset updates easier for me, but also to detect any critical problems with the ARM64 pool that would doom the large test runs. I've structured the dependencies carefully to avoid making the critical path even longer (i.e. the total latency of all checks).
• Style: Explicitly mark Code Format as having no dependencies, even though it's first.
  • This was the only order-dependent one, and I think it's simpler to be explicit.