Hojun-Cho/ww
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

w6c+wwstage: emit over-cap tuple return via sret callee-side (#10 Fold A)

Browse Source 
A tuple return whose SysV register-return footprint exceeds the caps
(> 4 integer eightbytes or > 2 SSE eightbytes) previously LOUD-STOPPED
at the N_RETURN SEND. Fold A makes the CALLEE emit such a return through
the existing >24B-struct sret skeleton:

  - classifier (cg_sret_retsize / sretretsize) grows a TY_TUPLE arm:
    walk the element footprint over the SAME caps the SEND uses, and
    return the tuple's natural total size (type table) when over-cap,
    else 0. The gp/sse caps are factored to a single shared SSoT
    (TUPLE_GPCAP / TUPLE_SSECAP — cgen.c macros in cstage, cgen.ww defs
    in wwstage) consumed by the classifier AND every emit/receive site
    (the SEND, the destructure guards, the cgcall arg guard) — so
    classify and emit can't disagree in either stage.
  - the SEND replaces the loud-stop with a write-through: cgexpr each
    element, store it through *(@sretarg) at its packed layout offset
    (the t.0/t.1 positional layout), each at its natural width so a
    narrow tail stores MOVL/MOVB not an over-MOVQ (#169); the dest base
    reloads into DX each step since a wide element clobbers AX/BX/CX.
    Then the existing struct-sret epilogue (MOVQ @sretarg->AX; ret).
  - the prologue already wires @sretarg when the classifier is nonzero.

The CALL/receive side is deliberately untouched: the N_MLET/N_MASSIGN
destructure loud-stops stay, so an over-cap tuple return is not yet
usefully callable. The end-to-end round-trip arrives with Fold B (#10-B).

Symmetric cstage (cmd/w6c/cgen.c) + wwstage (cgen.ww / cgenstmt.ww /
cgenutil.ww); combined.ww amalgams regenerated. Test 798 asserts the
callee now COMPILES (no loud-stop) and w6c vs w6c_ww .s byte-identical
across all-wide, str, narrow-tail, and float-over-cap shapes; no runtime
row (uncallable until Fold B). All 236 pass incl. 990-997 byte-id.
This commit is contained in:
Hojun-Cho
2026-06-01 11:18:44 +09:00
parent fb62aa38f1
commit 19e6b68d03
8 changed files with 605 additions and 84 deletions
Download Patch File Download Diff File Expand all files Collapse all files

130
cmd/w6c/cgen.c
View File

@@ -159,18 +159,11 @@ type_chase_named(Type *t)
return t;
}
/* cg_sret_retsize — if `rt` is a plain TY_STRUCT > 24B, return its
* natural size (the sret threshold); else 0. Tagged unions, tuples,
* str, and slices route through their existing register-return ABIs
* regardless of size. Task #23. */
static int
cg_sret_retsize(Type *rt)
{
rt = type_chase_named(rt);
if (rt == NULL || rt->kind != TY_STRUCT) return 0;
if ((int)rt->size <= 24) return 0;
return (int)rt->size;
}
/* cg_sret_retsize — sret classifier; defined after the tuple register-
* return helpers (tuple_rseq / tuple_ebytes / fld_isfloat) it consults
* for the over-cap-tuple arm. Forward-declared here for the earlier
* callers (cgcall, fn prologue). Task #23 / #10. */
static int cg_sret_retsize(Type *rt);
static int
node_isfloat(Node *n)
@@ -242,6 +235,14 @@ static const int tuple_rseq[] = { D_AX, D_DX, D_CX, D_R8 };
* return convergence (#171) is a call-site swap, not a redesign. */
static const int tuple_sse_seq[] = { D_X0, D_X1 };
/* #10: the register-return-ABI caps — the SINGLE SSoT shared by the sret
* classifier (cg_sret_retsize over-cap-tuple arm) AND every emit/receive
* site (N_RETURN tuple SEND, N_MLET/N_MASSIGN destructure, cgcall guard).
* Classify and emit MUST agree on these, else a tuple gets classified
* sret by one and in-reg by the other → corruption. */
#define TUPLE_GPCAP ((int)nelem(tuple_rseq))
#define TUPLE_SSECAP ((int)nelem(tuple_sse_seq))
static int
tuple_ebytes(int wide)
{
@@ -293,6 +294,41 @@ fld_isfloat(Type *t, int *isf32)
return 0;
}
/* cg_sret_retsize — sret classification by natural return size:
* - plain TY_STRUCT > 24B → its natural size (the #23 threshold).
* - TY_TUPLE whose SysV register-return footprint exceeds the caps
* (> TUPLE_GPCAP integer eightbytes or > TUPLE_SSECAP float
* eightbytes) → its natural total size, so the callee returns it
* via sret instead of registers (#10). The element footprint walk
* matches the N_RETURN tuple SEND exactly (a float = 1 SSE
* eightbyte, a slice/str its 3-word header, a scalar 1 GP word).
* Everything else (in-cap tuples, tagged unions, str, slices, scalars)
* routes through its register-return ABI → 0. */
static int
cg_sret_retsize(Type *rt)
{
rt = type_chase_named(rt);
if (rt == NULL) return 0;
if (rt->kind == TY_STRUCT)
return (int)rt->size <= 24 ? 0 : (int)rt->size;
if (rt->kind == TY_TUPLE) {
int gptotal = 0, ssecount = 0, f32;
for (Tparam *p = rt->params; p; p = p->next) {
Type *pu = type_chase_named(p->type);
int wide = pu && (pu->kind == TY_SLICE
|| pu->kind == TY_STR);
if (fld_isfloat(p->type, &f32))
ssecount++;
else
gptotal += tuple_ebytes(wide);
}
if (gptotal > TUPLE_GPCAP || ssecount > TUPLE_SSECAP)
return (int)rt->size;
return 0;
}
return 0;
}
/* fld_issigned — true iff a sub-word field/element load needs sign
* extension (i8 → MOVSBQ, i16 → MOVSWQ, i32 → MOVSXD). Follows NAMED
* and ENUM aliases via type_isunsigned, then peels off the unsigned
@@ -5843,10 +5879,7 @@ cgexpr(Cg *c, Node *n, Local *locals)
/* The producing call already satisfied #164's
* return caps; guard anyway (tuple_store indexes
* tuple_rseq[4] / tuple_sse_seq[2]). */
if (gptot > (int)(sizeof tuple_rseq
/ sizeof tuple_rseq[0])
|| sstot > (int)(sizeof tuple_sse_seq
/ sizeof tuple_sse_seq[0]))
if (gptot > TUPLE_GPCAP || sstot > TUPLE_SSECAP)
fatal("tuple arg exceeds return-cursor ABI "
"capacity; see #163/#164");
if (cg_tupargscr == 0) {
@@ -8687,9 +8720,7 @@ cgstmt(Cg *c, Node *n, Local **locals, int *frame)
* Both rows are loud-stopped at their cap (rule-7, never a
* silent collide): INTEGER 4, SSE 2. The SAME class split
* drives the receive sites. */
int gpcap = (int)(sizeof tuple_rseq / sizeof tuple_rseq[0]);
int ssecap = (int)(sizeof tuple_sse_seq
/ sizeof tuple_sse_seq[0]);
int ssecap = TUPLE_SSECAP;
int gptotal = 0, ssecount = 0, f32;
for (Node *e = n->lhs->list; e; e = e->next) {
if (fld_isfloat(e->type, &f32))
@@ -8698,14 +8729,49 @@ cgstmt(Cg *c, Node *n, Local **locals, int *frame)
gptotal += tuple_ebytes(node_isstr(e)
|| node_isslice(e));
}
if (gptotal > gpcap)
fatal("tuple return exceeds integer register-return "
"ABI capacity (%d eightbytes: AX,DX,CX,R8); "
"see return-ABI #10", gpcap);
if (ssecount > ssecap)
fatal("tuple return exceeds SSE register-return ABI "
"capacity (%d eightbytes: X0,X1); "
"see return-ABI #10", ssecap);
if (gptotal > TUPLE_GPCAP || ssecount > ssecap) {
/* #10 Fold A: over-cap tuple returns via sret. The
* prologue wired @sretarg (cg_sret_retsize agrees on
* the caps — the shared SSoT), holding the caller-
* prealloc dest. Store each element through
* *(@sretarg) at its packed layout offset (running
* sum of element sizes — the t.0/t.1 positional
* layout, N_DOT TY_TUPLE arm), each at its natural
* width so a narrow tail doesn't over-MOVQ (#169);
* the dest base is reloaded into DX each step since a
* wide element's cgexpr clobbers AX/BX/CX. Then reuse
* the struct-sret epilogue. The CALL/receive side
* stays loud-stopped (#10 Fold B). */
int foff = 0;
for (Node *e = n->lhs->list; e; e = e->next) {
int isflt = fld_isfloat(e->type, &f32);
int wide = node_isstr(e) || node_isslice(e);
int esz = e->type ? (int)e->type->size : 8;
cgexpr(c, e, *locals);
ins2(c, A_MOVQ,
amem(D_BP, cg_sret_arg_off), areg(D_DX));
if (isflt)
ins2(c, f32 ? A_MOVSS : A_MOVSD,
areg(D_X0), amem(D_DX, foff));
else if (wide) {
ins2(c, A_MOVQ, areg(D_AX),
amem(D_DX, foff + 0));
ins2(c, A_MOVQ, areg(D_BX),
amem(D_DX, foff + 8));
ins2(c, A_MOVQ, areg(D_CX),
amem(D_DX, foff + 16));
} else
ins2(c, fldstoreop(e->type, esz),
areg(D_AX), amem(D_DX, foff));
foff += esz;
}
ins2(c, A_MOVQ, amem(D_BP, cg_sret_arg_off),
areg(D_AX));
ins2(c, A_MOVQ, areg(D_BP), areg(D_SP));
ins1(c, A_POPQ, areg(D_BP));
ins0(c, A_RET);
break;
}
int fscr = 0;
if (ssecount > 0) {
if (cg_tupfscr != 0)
@@ -8936,9 +9002,8 @@ cgstmt(Cg *c, Node *n, Local **locals, int *frame)
* INTEGER cursor into an 8B slot. Both rows loud-stop at their
* cap. */
cgexpr(c, n->rhs, *locals);
int gpcap = (int)(sizeof tuple_rseq / sizeof tuple_rseq[0]);
int ssecap = (int)(sizeof tuple_sse_seq
/ sizeof tuple_sse_seq[0]);
int gpcap = TUPLE_GPCAP;
int ssecap = TUPLE_SSECAP;
int gptotal = 0, ssetotal = 0, lf32;
for (Node *l = n->list; l; l = l->next) {
Type *t = l->type;
@@ -8995,9 +9060,8 @@ cgstmt(Cg *c, Node *n, Local **locals, int *frame)
Type *rt = n->rhs ? n->rhs->type : NULL;
Type *ru = (rt && rt->kind == TY_NAMED) ? rt->under : rt;
Tparam *tp0 = (ru && ru->kind == TY_TUPLE) ? ru->params : NULL;
int gpcap = (int)(sizeof tuple_rseq / sizeof tuple_rseq[0]);
int ssecap = (int)(sizeof tuple_sse_seq
/ sizeof tuple_sse_seq[0]);
int gpcap = TUPLE_GPCAP;
int ssecap = TUPLE_SSECAP;
int gptotal = 0, ssetotal = 0, mf32;
for (Tparam *tp = tp0; tp; tp = tp->next) {
Type *u = (tp->type && tp->type->kind == TY_NAMED)