kit

commit 1b065cf837d4d5833314562416492fa11d90b786
parent b4d06b0e6b23ba1c8e9beb2984cd60d3a2b9b0ba
Author: Ryan Sepassi <rsepassi@gmail.com>
Date:   Wed, 27 May 2026 12:24:31 -0700

aa64: pass/return long double (fp128) in full 128-bit q register

aa_part_scalar_type() classified any FP ABI part larger than 4 bytes as
F64, so a 16-byte fp128 (long double on aarch64-linux) was moved through
a 64-bit d register on the return path: the backend emitted `ldur d0`
instead of `ldur q0`, truncating the result to its low 64 bits. For most
values that low half is zero, so long double arithmetic returned 0.

Map size>8 FP parts to F128 so the 128-bit q-register move path is taken.
This affects both user code and the self-compiled runtime (fp_tf/fp_ti),
whose __addtf3/__extenddftf2/etc. carried the same truncation.

Also make the rt object/archive rules depend on $(BIN) as a regular (not
order-only) prerequisite: cfree compiles and archives its own runtime, so
a codegen change in the compiler must rebuild the rt. The archive lists
$(RT_OBJS) explicitly rather than $^ so the now-regular cfree binary
prereq is not itself archived.

Diffstat:
M
rt/Makefile
 | 
14
+++++++++-----
M
src/arch/aa64/native.c
 | 
8
+++++---

2 files changed, 14 insertions(+), 8 deletions(-)
diff --git a/rt/Makefile b/rt/Makefile
@@ -242,20 +242,24 @@ RT_OBJS_$(1) := $$(patsubst rt/lib/%,$$(RT_BUILD_DIR)/$(1)/%.o,$$(RT_SRCS_$(1)))
 
 rt-$(1): $$(RT_BUILD_DIR)/$(1)/libcfree_rt.a
 
-$$(RT_BUILD_DIR)/$(1)/libcfree_rt.a: $$(RT_OBJS_$(1)) | $$(BIN)
+# Regular (not order-only) dep on $(BIN): cfree compiles and archives its own
+# runtime, so a codegen or `ar` change in the compiler must rebuild the rt.
+# The archive lists $(RT_OBJS) explicitly rather than $^ so the cfree binary
+# (now a regular prereq) is not itself archived.
+$$(RT_BUILD_DIR)/$(1)/libcfree_rt.a: $$(RT_OBJS_$(1)) $$(BIN)
 	@mkdir -p $$(dir $$@)
 	@rm -f $$@
-	$$(RT_AR) rcs $$@ $$^
+	$$(RT_AR) rcs $$@ $$(RT_OBJS_$(1))
 
-$$(RT_BUILD_DIR)/$(1)/%.s.o: rt/lib/%.s | $$(BIN)
+$$(RT_BUILD_DIR)/$(1)/%.s.o: rt/lib/%.s $$(BIN)
 	@mkdir -p $$(dir $$@)
 	$$(RT_AS) $$(RT_ASFLAGS_$(1)) $$(RT_AS_COMPILE_FLAGS) $$< -o $$@
 
-$$(RT_BUILD_DIR)/$(1)/%.S.o: rt/lib/%.S | $$(BIN)
+$$(RT_BUILD_DIR)/$(1)/%.S.o: rt/lib/%.S $$(BIN)
 	@mkdir -p $$(dir $$@)
 	$$(RT_AS) $$(RT_ASFLAGS_$(1)) $$(RT_AS_COMPILE_FLAGS) $$< -o $$@
 
-$$(RT_BUILD_DIR)/$(1)/%.o: rt/lib/% | $$(BIN)
+$$(RT_BUILD_DIR)/$(1)/%.o: rt/lib/% $$(BIN)
 	@mkdir -p $$(dir $$@)
 	$$(RT_CC) $$(RT_CFLAGS_$(1)) -c $$< -o $$@
 
diff --git a/src/arch/aa64/native.c b/src/arch/aa64/native.c
@@ -1800,9 +1800,11 @@ static u32 aa_part_stack_align(const ABIArgPart* part) {
  * args/results are split into parts; each part must move at its own width, not
  * the (possibly >8-byte) aggregate width. */
 static CfreeCgTypeId aa_part_scalar_type(const ABIArgPart* part) {
-  if (part->cls == ABI_CLASS_FP)
-    return part->size <= 4u ? builtin_id(CFREE_CG_BUILTIN_F32)
-                            : builtin_id(CFREE_CG_BUILTIN_F64);
+  if (part->cls == ABI_CLASS_FP) {
+    if (part->size <= 4u) return builtin_id(CFREE_CG_BUILTIN_F32);
+    if (part->size <= 8u) return builtin_id(CFREE_CG_BUILTIN_F64);
+    return builtin_id(CFREE_CG_BUILTIN_F128);
+  }
   switch (part->size) {
     case 1u:
       return builtin_id(CFREE_CG_BUILTIN_I8);