codegenDecl on functions overrides all attributes

[mratsim]

codegenDecl pragma on functions will override all function attributes.

Example:

proc foo() =
  echo 1

{.pragma: hot, codegenDecl: "__attribute__((hot)) $# $#$#".}

proc bar() {.hot.} =
  echo 1

proc baz() {.inline.} =
  echo 1

proc oops() {.inline, hot.} =
  echo 1

foo()
bar()
baz()
oops()

This produces (in release emode)

N_LIB_PRIVATE N_NIMCALL(void, foo_wznx9adj3TlVt3o5axJ9aYOg)(void) {
	echoBinSafe(TM_YoGu5Gbh4YxJ9c0z7IstHKg_2, 1);
}

__attribute__((hot)) void bar_wznx9adj3TlVt3o5axJ9aYOg_2(void) {
	echoBinSafe(TM_YoGu5Gbh4YxJ9c0z7IstHKg_2, 1);
}

static N_INLINE(void, baz_Ij9bzunIMj7NU3a7ws3YxBQattr_func)(void) {
	echoBinSafe(TM_YoGu5Gbh4YxJ9c0z7IstHKg_2, 1);
}

__attribute__((hot)) void oops_Ij9bzunIMj7NU3a7ws3YxBQ_2attr_func(void) {
	echoBinSafe(TM_YoGu5Gbh4YxJ9c0z7IstHKg_2, 1);
}

In the first case N_LIB_PRIVATE N_NIMCALL has been replaced and in the second case static N_INLINE has been replaced.

For a bit more context, I'm writing a domain specific compiler in Nim for machine learning and image processing and I need to implement one function per SIMD supported (nothing, SSE, AVX, AVX512, ...). Unfortunately, ideally those should be in separate C files to avoid the compiler using incompatible features in older sets which seems complicated. Alternatively I can use GCC/Clang only attributes like __attribute__((__target__("avx2"))).

[mratsim]

I checked how this is implemented in ccgtypes genProcHeader:

Nim/compiler/ccgtypes.nim

Lines 888 to 912 in 2610b16

	proc genProcHeader(m: BModule, prc: PSym): Rope =
	var
	rettype, params: Rope
	genCLineDir(result, prc.info, m.config)
	# using static is needed for inline procs
	if lfExportLib in prc.loc.flags:
	if isHeaderFile in m.flags:
	result.add "N_LIB_IMPORT "
	else:
	result.add "N_LIB_EXPORT "
	elif prc.typ.callConv == ccInline:
	result.add "static "
	elif {sfImportc, sfExportc} * prc.flags == {}:
	result.add "N_LIB_PRIVATE "
	var check = initIntSet()
	fillLoc(prc.loc, locProc, prc.ast[namePos], mangleName(m, prc), OnUnknown)
	genProcParams(m, prc.typ, rettype, params, check)
	# careful here! don't access ``prc.ast`` as that could reload large parts of
	# the object graph!
	if prc.constraint.isNil:
	addf(result, "$1($2, $3)$4",
	[rope(CallingConvToStr[prc.typ.callConv]), rettype, prc.loc.r,
	params])
	else:
	result = prc.cgDeclFrmt % [rettype, prc.loc.r, params]

I'm not sure there is a way to do what is needed in a backward compatible way.

I've found the current workaround:

proc foo() =
  echo 1

{.pragma: hot, codegenDecl: "__attribute__((hot)) N_LIB_PRIVATE N_NIMCALL($#, $#)$#".}
{.pragma: hot_inline, codegenDecl: "__attribute__((hot)) static N_INLINE($#, $#)$#".}

proc bar() {.hot.} =
  echo 1

proc baz() {.inline.} =
  echo 1

proc oops() {.hot_inline.} =
  echo 1

foo()
bar()
baz()
oops()

N_LIB_PRIVATE N_NIMCALL(void, foo_ZEVWdPzph9aEvWzKnk2jr4w)(void) {
	echoBinSafe(TM_rnFJuQ50wUpy8kftbt2QJA_2, 1);
}

__attribute__((hot)) N_LIB_PRIVATE N_NIMCALL(void, bar_ZEVWdPzph9aEvWzKnk2jr4w_2)(void) {
	echoBinSafe(TM_rnFJuQ50wUpy8kftbt2QJA_2, 1);
}

static N_INLINE(void, baz_txURmaVFiHb6ZhpniaNZhgattr_func3)(void) {
	echoBinSafe(TM_rnFJuQ50wUpy8kftbt2QJA_2, 1);
}

__attribute__((hot)) static N_INLINE(void, oops_ZEVWdPzph9aEvWzKnk2jr4w_3)(void) {
	echoBinSafe(TM_rnFJuQ50wUpy8kftbt2QJA_2, 1);
}

[mratsim]

Removing the low priority tag.

We are considering using LLVM Xray for instrumenting Nimbus.
Fine-grained instrumenting can be controlled with GCC attributes:

void alt_always_instrumented() __attribute__((xray_always_instrument));
void alt_never_instrumented() __attribute__((xray_never_instrument));