Fix trampoline on PPC (#38980)

* [Make] Normalize ppc64le to powerpc64le

* [CLI] Fix trampoline on PowerPC

1. Add global entry
2. Don't store `r2` on parent frame
3. Conservative load from toc

* remove jl_compile_extern_c from exported funcs

Author: vchuravy

Make.inc

@@ -795,6 +795,11 @@ XC_HOST := $(ARCH)$(shell echo $(BUILD_MACHINE) | sed "s/[^-]*\(.*\)$$/\1/")
 endif
 endif
 
+# Normalize ppc64le to powerpc64le
+ifeq ($(ARCH), ppc64le)
+override ARCH := powerpc64le
+endif
+
 ifeq ($(ARCH),mingw32)
 $(error "the mingw32 compiler you are using fails the openblas testsuite. please see the README.windows document for a replacement")
 else ifeq (cygwin, $(shell $(CC) -dumpmachine | cut -d\- -f3))

cli/trampolines/trampolines_powerpc64le.S

@@ -0,0 +1,25 @@
+#include "../../src/jl_exported_funcs.inc"
+
+// Notes:
+// bctr: branch to CTR without LR update (tail-call)
+// localentry: On PPC functions have a localentry that assumes r2 contains
+//             the TOC pointer, and a global entry point that sets r2.
+// See 64-Bit ELF V2 ABI Specification: Power Architecture v1.4
+
+#define XX(name) \
+.global name; \
+.type name##, @function; \
+.cfi_startproc; \
+name##: ; \
+    addis 2, 12, .TOC.-name##@ha; \
+    addi 2, 2, .TOC.-name##@l; \
+    .localentry name##,.-name##; \
+    addis 9,2,name##_addr@toc@ha; \
+    ld 12,name##_addr@toc@l(9); \
+    mtctr 12; \
+    bctr; \
+.cfi_endproc; \
+.size name##,.-name##; \
+
+JL_EXPORTED_FUNCS(XX)
+#undef XX

cli/trampolines/trampolines_ppc64le.S

@@ -111,7 +111,6 @@
     XX(jl_clock_now) \
     XX(jl_close_uv) \
     XX(jl_code_for_staged) \
-    XX(jl_compile_extern_c) \
     XX(jl_compile_hint) \
     XX(jl_compress_argnames) \
     XX(jl_compress_ir) \