5 files changed, 193 insertions, 19 deletions

diff --git a/src/lj_ccall.c b/src/lj_ccall.c
index c3eb25f6..71331f39 100644
--- a/src/lj_ccall.c
+++ b/src/lj_ccall.c
@@ -168,6 +168,8 @@
 #elif LJ_TARGET_ARM
 /* -- ARM calling conventions --------------------------------------------- */
 
+#if LJ_ABI_SOFTFP
+
 #define CCALL_HANDLE_STRUCTRET \
   /* Return structs of size <= 4 in a GPR. */ \
   cc->retref = !(sz <= 4); \
@@ -186,13 +188,70 @@
 #define CCALL_HANDLE_COMPLEXARG \
   /* Pass complex by value in 2 or 4 GPRs. */
 
-/* ARM has a softfp ABI. */
+#define CCALL_HANDLE_REGARG_FP1
+#define CCALL_HANDLE_REGARG_FP2
+
+#else
+
+#define CCALL_HANDLE_STRUCTRET \
+  cc->retref = !ccall_classify_struct(cts, ctr, ct); \
+  if (cc->retref) cc->gpr[ngpr++] = (GPRArg)dp;
+
+#define CCALL_HANDLE_STRUCTRET2 \
+  if (ccall_classify_struct(cts, ctr, ct) > 1) sp = (uint8_t *)&cc->fpr[0]; \
+  memcpy(dp, sp, ctr->size);
+
+#define CCALL_HANDLE_COMPLEXRET \
+  if (!(ct->info & CTF_VARARG)) cc->retref = 0;  /* Return complex in FPRs. */
+
+#define CCALL_HANDLE_COMPLEXRET2 \
+  if (!(ct->info & CTF_VARARG)) memcpy(dp, &cc->fpr[0], ctr->size);
+
+#define CCALL_HANDLE_STRUCTARG \
+  isfp = (ccall_classify_struct(cts, d, ct) > 1);
+  /* Pass all structs by value in registers and/or on the stack. */
+
+#define CCALL_HANDLE_COMPLEXARG \
+  isfp = 1;  /* Pass complex by value in FPRs or on stack. */
+
+#define CCALL_HANDLE_REGARG_FP1 \
+  if (isfp && !(ct->info & CTF_VARARG)) { \
+    if ((d->info & CTF_ALIGN) > CTALIGN_PTR) { \
+      if (nfpr + (n >> 1) <= CCALL_NARG_FPR) { \
+        dp = &cc->fpr[nfpr]; \
+        nfpr += (n >> 1); \
+        goto done; \
+      } \
+    } else { \
+      if (sz > 1 && fprodd != nfpr) fprodd = 0; \
+      if (fprodd) { \
+        if (2*nfpr+n <= 2*CCALL_NARG_FPR+1) { \
+          dp = (void *)&cc->fpr[fprodd-1].f[1]; \
+          nfpr += (n >> 1); \
+          if ((n & 1)) fprodd = 0; else fprodd = nfpr-1; \
+          goto done; \
+        } \
+      } else { \
+        if (2*nfpr+n <= 2*CCALL_NARG_FPR) { \
+          dp = (void *)&cc->fpr[nfpr]; \
+          nfpr += (n >> 1); \
+          if ((n & 1)) fprodd = ++nfpr; else fprodd = 0; \
+          goto done; \
+        } \
+      } \
+    } \
+    fprodd = 0;  /* No reordering after the first FP value is on stack. */ \
+  } else {
+
+#define CCALL_HANDLE_REGARG_FP2 }
+
+#endif
+
 #define CCALL_HANDLE_REGARG \
+  CCALL_HANDLE_REGARG_FP1 \
   if ((d->info & CTF_ALIGN) > CTALIGN_PTR) { \
     if (ngpr < maxgpr) \
       ngpr = (ngpr + 1u) & ~1u;  /* Align to regpair. */ \
-    else \
-      nsp = (nsp + 1u) & ~1u;  /* Align argument on stack. */ \
   } \
   if (ngpr < maxgpr) { \
     dp = &cc->gpr[ngpr]; \
@@ -204,7 +263,10 @@
       ngpr += n; \
     } \
     goto done; \
-  }
+  } CCALL_HANDLE_REGARG_FP2
+
+#define CCALL_HANDLE_RET \
+  if ((ct->info & CTF_VARARG)) sp = (uint8_t *)&cc->gpr[0];
 
 #elif LJ_TARGET_PPC
 /* -- PPC calling conventions --------------------------------------------- */
@@ -453,6 +515,49 @@ static void ccall_struct_ret(CCallState *cc, int *rcl, uint8_t *dp, CTSize sz)
 }
 #endif
 
+/* -- ARM hard-float ABI struct classification ---------------------------- */
+
+#if LJ_TARGET_ARM && !LJ_ABI_SOFTFP
+
+/* Classify a struct based on its fields. */
+static unsigned int ccall_classify_struct(CTState *cts, CType *ct, CType *ctf)
+{
+  CTSize sz = ct->size;
+  unsigned int r = 0, n = 0, isu = (ct->info & CTF_UNION);
+  if ((ctf->info & CTF_VARARG)) goto noth;
+  while (ct->sib) {
+    ct = ctype_get(cts, ct->sib);
+    if (ctype_isfield(ct->info)) {
+      CType *sct = ctype_rawchild(cts, ct);
+      if (ctype_isfp(sct->info)) {
+        r |= sct->size;
+        if (!isu) n++; else if (n == 0) n = 1;
+      } else if (ctype_iscomplex(sct->info)) {
+        r |= (sct->size >> 1);
+        if (!isu) n += 2; else if (n < 2) n = 2;
+      } else {
+        goto noth;
+      }
+    } else if (ctype_isbitfield(ct->info)) {
+      goto noth;
+    } else if (ctype_isxattrib(ct->info, CTA_SUBTYPE)) {
+      CType *sct = ctype_child(cts, ct);
+      if (sct->size > 0) {
+        unsigned int s = ccall_classify_struct(cts, sct, ctf);
+        if (s <= 1) goto noth;
+        r |= (s & 255);
+        if (!isu) n += (s >> 8); else if (n < (s >>8)) n = (s >> 8);
+      }
+    }
+  }
+  if ((r == 4 || r == 8) && n <= 4)
+    return r + (n << 8);
+noth:  /* Not a homogeneous float/double aggregate. */
+  return (sz <= 4);  /* Return structs of size <= 4 in a GPR. */
+}
+
+#endif
+
 /* -- Common C call handling ---------------------------------------------- */
 
 /* Infer the destination CTypeID for a vararg argument. */
@@ -494,6 +599,9 @@ static int ccall_set_args(lua_State *L, CTState *cts, CType *ct,
   MSize maxgpr, ngpr = 0, nsp = 0, narg;
 #if CCALL_NARG_FPR
   MSize nfpr = 0;
+#if LJ_TARGET_ARM
+  MSize fprodd = 0;
+#endif
 #endif
 
   /* Clear unused regs to get some determinism in case of misdeclaration. */
diff --git a/src/lj_ccall.h b/src/lj_ccall.h
index 5985c4a9..62f963e1 100644
--- a/src/lj_ccall.h
+++ b/src/lj_ccall.h
@@ -51,12 +51,21 @@ typedef intptr_t GPRArg;
 #elif LJ_TARGET_ARM
 
 #define CCALL_NARG_GPR          4
-#define CCALL_NARG_FPR          0
 #define CCALL_NRET_GPR          2       /* For softfp double. */
+#if LJ_ABI_SOFTFP
+#define CCALL_NARG_FPR          0
 #define CCALL_NRET_FPR          0
+#else
+#define CCALL_NARG_FPR          8
+#define CCALL_NRET_FPR          4
+#endif
 #define CCALL_SPS_FREE          0
 
 typedef intptr_t GPRArg;
+typedef union FPRArg {
+  double d;
+  float f[2];
+} FPRArg;
 
 #elif LJ_TARGET_PPC
 
@@ -122,7 +131,7 @@ LJ_STATIC_ASSERT(CCALL_NUM_FPR <= CCALL_MAX_FPR);
 
 /* -- C call state -------------------------------------------------------- */
 
-typedef struct CCallState {
+typedef LJ_ALIGN(8) struct CCallState {
   void (*func)(void);           /* Pointer to called function. */
   uint32_t spadj;               /* Stack pointer adjustment. */
   uint8_t nsp;                  /* Number of stack slots. */
@@ -135,10 +144,10 @@ typedef struct CCallState {
 #elif LJ_TARGET_PPC
   uint8_t nfpr;                 /* Number of arguments in FPRs. */
 #endif
-#if CCALL_NUM_FPR
 #if LJ_32
   int32_t align1;
 #endif
+#if CCALL_NUM_FPR
   FPRArg fpr[CCALL_NUM_FPR];    /* Arguments/results in FPRs. */
 #endif
   GPRArg gpr[CCALL_NUM_GPR];    /* Arguments/results in GPRs. */
diff --git a/src/lj_ccallback.c b/src/lj_ccallback.c
index a9567bc5..430643ee 100644
--- a/src/lj_ccallback.c
+++ b/src/lj_ccallback.c
@@ -310,22 +310,53 @@ void lj_ccallback_mcode_free(CTState *cts)
 
 #elif LJ_TARGET_ARM
 
+#if LJ_ABI_SOFTFP
+
+#define CALLBACK_HANDLE_REGARG_FP1      UNUSED(isfp);
+#define CALLBACK_HANDLE_REGARG_FP2
+
+#else
+
+#define CALLBACK_HANDLE_REGARG_FP1 \
+  if (isfp) { \
+    if (n == 1) { \
+      if (fprodd) { \
+        sp = &cts->cb.fpr[fprodd-1]; \
+        fprodd = 0; \
+        goto done; \
+      } else if (nfpr + 1 <= CCALL_NARG_FPR) { \
+        sp = &cts->cb.fpr[nfpr++]; \
+        fprodd = nfpr; \
+        goto done; \
+      } \
+    } else { \
+      if (nfpr + 1 <= CCALL_NARG_FPR) { \
+        sp = &cts->cb.fpr[nfpr++]; \
+        goto done; \
+      } \
+    } \
+    fprodd = 0;  /* No reordering after the first FP value is on stack. */ \
+  } else {
+
+#define CALLBACK_HANDLE_REGARG_FP2      }
+
+#endif
+
 #define CALLBACK_HANDLE_REGARG \
-  UNUSED(isfp); \
+  CALLBACK_HANDLE_REGARG_FP1 \
   if (n > 1) ngpr = (ngpr + 1u) & ~1u;  /* Align to regpair. */ \
   if (ngpr + n <= maxgpr) { \
     sp = &cts->cb.gpr[ngpr]; \
     ngpr += n; \
     goto done; \
-  }
+  } CALLBACK_HANDLE_REGARG_FP2
 
 #elif LJ_TARGET_PPC
 
 #define CALLBACK_HANDLE_REGARG \
   if (isfp) { \
     if (nfpr + 1 <= CCALL_NARG_FPR) { \
-      sp = &cts->cb.fpr[nfpr]; \
-      nfpr += 1; \
+      sp = &cts->cb.fpr[nfpr++]; \
       cta = ctype_get(cts, CTID_DOUBLE);  /* FPRs always hold doubles. */ \
       goto done; \
     } \
@@ -382,6 +413,9 @@ static void callback_conv_args(CTState *cts, lua_State *L)
   MSize ngpr = 0, nsp = 0, maxgpr = CCALL_NARG_GPR;
 #if CCALL_NARG_FPR
   MSize nfpr = 0;
+#if LJ_TARGET_ARM
+  MSize fprodd = 0;
+#endif
 #endif
 
   if (slot < cts->cb.sizeid && (id = cts->cb.cbid[slot]) != 0) {
diff --git a/src/lj_ctype.h b/src/lj_ctype.h
index 7953654f..7c3b667c 100644
--- a/src/lj_ctype.h
+++ b/src/lj_ctype.h
@@ -155,7 +155,7 @@ typedef struct CType {
 #define CCALL_MAX_GPR           8
 #define CCALL_MAX_FPR           8
 
-typedef LJ_ALIGN(8) union FPRCBArg { double d; float f; } FPRCBArg;
+typedef LJ_ALIGN(8) union FPRCBArg { double d; float f[2]; } FPRCBArg;
 
 /* C callback state. Defined here, to avoid dragging in lj_ccall.h. */
 
diff --git a/src/vm_arm.dasc b/src/vm_arm.dasc
index 26f97aa3..355a53e6 100644
--- a/src/vm_arm.dasc
+++ b/src/vm_arm.dasc
@@ -2414,8 +2414,19 @@ static void build_subroutines(BuildCtx *ctx)
   |.type CTSTATE, CTState, PC
   |  ldr CTSTATE, GL:r12->ctype_state
   |   add DISPATCH, r12, #GG_G2DISP
-  |  strd CARG12, CTSTATE->cb.gpr[0]
+  |.if FPU
+  |  str r4, SAVE_R4
+  |  add r4, sp, CFRAME_SPACE+4+8*8
+  |  vstmdb r4!, {d8-d15}
+  |.endif
+  |.if HFABI
+  |  add r12, CTSTATE, #offsetof(CTState, cb.fpr[8])
+  |.endif
   |  strd CARG34, CTSTATE->cb.gpr[2]
+  |  strd CARG12, CTSTATE->cb.gpr[0]
+  |.if HFABI
+  |  vstmdb r12!, {d0-d7}
+  |.endif
   |  ldr CARG4, [sp]
   |   add CARG3, sp, #CFRAME_SIZE
   |    mov CARG1, CTSTATE
@@ -2448,6 +2459,9 @@ static void build_subroutines(BuildCtx *ctx)
   |  mov CARG2, RA
   |  bl extern lj_ccallback_leave       // (CTState *cts, TValue *o)
   |  ldrd CARG12, CTSTATE->cb.gpr[0]
+  |.if HFABI
+  |  vldr d0, CTSTATE->cb.fpr[0]
+  |.endif
   |  b ->vm_leave_unw
   |.endif
   |
@@ -2460,9 +2474,15 @@ static void build_subroutines(BuildCtx *ctx)
   |  ldr CARG1, CCSTATE:CARG1->spadj
   |   ldrb CARG2, CCSTATE->nsp
   |    add CARG3, CCSTATE, #offsetof(CCallState, stack)
+  |.if HFABI
+  |  add RB, CCSTATE, #offsetof(CCallState, fpr[0])
+  |.endif
   |  mov r11, sp
   |  sub sp, sp, CARG1                  // Readjust stack.
   |   subs CARG2, CARG2, #1
+  |.if HFABI
+  |  vldm RB, {d0-d7}
+  |.endif
   |    ldr RB, CCSTATE->func
   |   bmi >2
   |1:  // Copy stack slots.
@@ -2471,14 +2491,17 @@ static void build_subroutines(BuildCtx *ctx)
   |  subs CARG2, CARG2, #1
   |  bpl <1
   |2:
-  |  ldr CARG1, CCSTATE->gpr[0]
-  |  ldr CARG2, CCSTATE->gpr[1]
-  |  ldr CARG3, CCSTATE->gpr[2]
-  |  ldr CARG4, CCSTATE->gpr[3]
+  |  ldrd CARG12, CCSTATE->gpr[0]
+  |  ldrd CARG34, CCSTATE->gpr[2]
   |  blx RB
   |  mov sp, r11
-  |  str CRET1, CCSTATE->gpr[0]
-  |  str CRET2, CCSTATE->gpr[1]
+  |.if HFABI
+  |  add r12, CCSTATE, #offsetof(CCallState, fpr[4])
+  |.endif
+  |  strd CRET1, CCSTATE->gpr[0]
+  |.if HFABI
+  |  vstmdb r12!, {d0-d3}
+  |.endif
   |  pop {CCSTATE, r5, r11, pc}
   |.endif
   |// Note: vm_ffi_call must be the last function in this object file!