Build: Switch from frontend-PGO to IR-PGO

[kinke]

I.e., from LDC's own AST-based implementation (-fprofile-instr-…) to LLVM-builtin IR-based PGO (-fprofile-…). With the downside of having to enable optimizations for the instrumented compiler (=> longer build time).

[github-actions]

DMD perf check

Metric	Base	PR	Δ
compile hello.d (instr)	213.0 M	211.6 M	-0.690%
compile hello.d -O -release (instr)	231.1 M	229.1 M	-0.827%
compile Phobos (instr)	5,081.2 M	4,902.5 M	-3.519%
compile Phobos codegen (instr)	1,442.3 M	1,429.4 M	-0.896%
compile vibe.d (instr)	15,019.2 M	14,494.2 M	-3.495%
dmd binary size (stripped)	6.91 MB	8.13 MB	+17.76%
peak RSS (compile hello.d)	43.29 MB	44.26 MB	+2.26%

Breakdown — compile hello.d

Phase (wall, self time)	Base	PR	Δ
sema1	10.2 ms	9.3 ms	-8.84%
sema_other	14.2 ms	13.6 ms	-4.47%
sema3	5.9 ms	5.3 ms	-10.43%
parse	31.9 ms	31.4 ms	-1.58%
codegen	2.0 ms	1.9 ms	-3.92%

Breakdown — compile Phobos

-178.8 M instructions: frontend -165.9 M (-4.56%), codegen -12.9 M (-0.90%)

Phase (wall, self time)	Base	PR	Δ
sema3	607 ms	533 ms	-12.21%
sema1	224 ms	203 ms	-9.17%
codegen	401 ms	386 ms	-3.69%
sema_other	206 ms	194 ms	-5.87%
parse	110 ms	106 ms	-3.85%
ctfe	14.2 ms	11.5 ms	-18.81%
sema2	1.5 ms	0.8 ms	-44.64%
inline	4.3 ms	4.4 ms	+2.99%

All measurements

Metric	Base	PR	Δ
compile hello.d (instr)	213.0 M	211.6 M	-0.690%
compile hello.d -O -release (instr)	231.1 M	229.1 M	-0.827%
compile Phobos (instr)	5,081.2 M	4,902.5 M	-3.519%
compile Phobos codegen (instr)	1,442.3 M	1,429.4 M	-0.896%
compile vibe.d (instr)	15,019.2 M	14,494.2 M	-3.495%
dmd binary size (stripped)	6.91 MB	8.13 MB	+17.76%
hello binary size (stripped)	0.72 MB	0.72 MB	0.00%
peak RSS (compile hello.d)	43.29 MB	44.26 MB	+2.26%
peak RSS (compile Phobos)	620.2 MB	616.1 MB	-0.65%
peak RSS (compile vibe.d)	1916 MB	1908 MB	-0.43%
compile dmd itself (wall)	12.1 s	12.1 s	+0.08%
compile hello.d (wall)	64.2 ms	61.5 ms	-4.26%
compile Phobos (wall)	1,568 ms	1,439 ms	-8.23%

_{2c9ff5f vs merge-base 278dce2 · about these metrics}

[kinke]

Oh wow, pretty significant change in metrics! Note that due to optimizing the instrumented compiler, the runtime of the perf CI job (to gather the performance metrics) would be prolonged by ~2.5 minutes. At least according to https://github.com/dlang/dmd/actions/runs/35171311430/job/105043379050?pr=23870, where the non-optimized build for master took 13 secs, and the optimized one 1:27 mins.

FYI @abulgit @0xEAB (testing this on Windows would be nice :])