finish jit_allocai implementation

2006-11-04  Paolo Bonzini  <bonzini@gnu.org>

	* lightning/ppc/core.h: Implement jit_allocai, define JIT_FP to be R1.
	* lightning/ppc/funcs.h: Store frame size into _jitl.  Store R1 before
	the STMW, so that the offset is unchanged when we patch the STMW.
	* lightning/i386/core.h: Define JIT_FP to be EBP.
	* lightning/i386/core-32.h: Implement jit_allocai, put LEAVE in the
	epilog if jit_allocai was used.
	* lightning/i386/core-64.h: Implement jit_allocai, put LEAVE in the
	epilog if jit_allocai was used.

git-archimport-id: bonzini@gnu.org--2004b/lightning--stable--1.2--patch-36

ChangeLog

@@ -1,3 +1,20 @@
+2006-11-04  Paolo Bonzini  <bonzini@gnu.org>
+
+	* lightning/ppc/core.h: Implement jit_allocai, define JIT_FP to be R1.
+	* lightning/ppc/funcs.h: Store frame size into _jitl.  Store R1 before
+	the STMW, so that the offset is unchanged when we patch the STMW.
+	* lightning/i386/core.h: Define JIT_FP to be EBP.
+	* lightning/i386/core-32.h: Implement jit_allocai, put LEAVE in the
+	epilog if jit_allocai was used.
+	* lightning/i386/core-64.h: Implement jit_allocai, put LEAVE in the
+	epilog if jit_allocai was used.
+
+2006-11-04  Ludovic Courtes  <ludo@chbouib.org>
+
+	* lightning/sparc/core.h: Implement jit_allocai.
+	* tests/allocai.c: New.
+	* tests/Makefile.am: Point to new tests.
+
 2006-11-03  Paolo Bonzini  <bonzini@gnu.org>
 
 	* lightning/ppc/core.h: Fix jit_bms using BNE rather than BGT.

NEWS

@@ -9,7 +9,6 @@ o   Support for stack-allocated variables.  Because of this,
     backends defining JIT_FP should now rename it to JIT_AP.
     JIT_FP is now a user-visible register used in ldxi/ldxr
     to access stack-allocated variables.
-    [a promise for now, not yet implemented!]
 
 
 ---

lightning/i386/core-32.h

@@ -41,21 +41,43 @@
 struct jit_local_state {
   int	framesize;
   int	argssize;
+  int	alloca_offset;
+  int	alloca_slack;
 };
 
 #define jit_base_prolog() (PUSHLr(_EBP), MOVLrr(_ESP, _EBP), PUSHLr(_EBX), PUSHLr(_ESI), PUSHLr(_EDI))
-#define jit_prolog(n) (_jitl.framesize = 8, jit_base_prolog())
+#define jit_prolog(n) (_jitl.framesize = 8, _jitl.alloca_offset = -12, jit_base_prolog())
 
-/* The += allows for stack pollution */
+/* Used internally.  SLACK is used by the Darwin ABI which keeps the stack
+   aligned to 16-bytes.  */
+
+#define jit_allocai_internal(amount, slack)				  \
+  (((amount) < _jitl.alloca_slack					  \
+    ? 0									  \
+    : (_jitl.alloca_slack += (amount) + (slack),			  \
+       ((amount) + (slack) == sizeof (int)				  \
+        ? PUSHLr(_EAX)							  \
+        : SUBLir((amount) + (slack), _ESP)))),				  \
+   _jitl.alloca_slack -= (amount),					  \
+   _jitl.alloca_offset -= (amount))
+   
+/* The += in argssize allows for stack pollution */
 
 #ifdef __APPLE__
-  /* Stack must stay 16-byte aligned: */
+/* Stack must stay 16-byte aligned: */
 # define jit_prepare_i(ni)	(((ni & 0x3) \
                                   ? SUBLir(4 * ((((ni) + 3) & ~(0x3)) - (ni)), JIT_SP) \
                                   : (void)0), \
                                  _jitl.argssize += (((ni) + 3) & ~(0x3)))
+
+#define jit_allocai(n)						\
+  jit_allocai_internal ((n), (_jitl.alloca_slack - (n)) & 15)
+
 #else
 # define jit_prepare_i(ni)	(_jitl.argssize += (ni))
+
+#define jit_allocai(n)						\
+  jit_allocai_internal ((n), 0)
 #endif
 
 #define jit_pusharg_i(rs)	PUSHLr(rs)
@@ -74,7 +96,7 @@ struct jit_local_state {
 
 #define jit_patch_long_at(jump_pc,v)  (*_PSL((jump_pc) - sizeof(long)) = _jit_SL((jit_insn *)(v) - (jump_pc)))
 #define jit_patch_at(jump_pc,v)  jit_patch_long_at(jump_pc, v)
-#define jit_ret() (POPLr(_EDI), POPLr(_ESI), POPLr(_EBX), POPLr(_EBP), RET_())
+#define jit_ret()		(POPLr(_EDI), POPLr(_ESI), POPLr(_EBX), (_jitl.alloca_offset < -12 ? LEAVE_() : POPLr(_EBP)), RET_())
 
 #endif /* __lightning_core_h */
 

lightning/i386/core-64.h

@@ -40,8 +40,23 @@ struct jit_local_state {
   int   long_jumps;
   int   nextarg_geti;
   int	argssize;
+  int   alloca_offset;
+  int   alloca_slack;
 };
 
+
+/* Keep the stack 16-byte aligned, the SSE hardware prefers it this way.  */
+#define jit_allocai_internal(amount, slack)                           \
+  (((amount) < _jitl.alloca_slack                                     \
+    ? 0                                                               \
+    : (_jitl.alloca_slack += (amount) + (slack),                      \
+      SUBQir((amount) + (slack), _ESP))),                             \
+   _jitl.alloca_slack -= (amount),                                    \
+   _jitl.alloca_offset -= (amount))
+
+#define jit_allocai(n)                                                \
+  jit_allocai_internal ((n), (_jitl.alloca_slack - (n)) & 15)
+
 /* 3-parameter operation */
 #define jit_qopr_(d, s1, s2, op1d, op2d)				\
 	( (s2 == d) ? op1d :						\
@@ -95,7 +110,7 @@ struct jit_local_state {
 #define jit_popr_l(rs)		POPQr(rs)
 
 #define jit_base_prolog() (PUSHQr(_EBP), MOVQrr(_ESP, _EBP), PUSHQr(_EBX), PUSHQr(_R12), PUSHQr(_R13))
-#define jit_prolog(n) (_jitl.nextarg_geti = 0, jit_base_prolog())
+#define jit_prolog(n) (_jitl.nextarg_geti = 0, _jitl.alloca_offset = -24, jit_base_prolog())
 
 /* Stack isn't used for arguments: */
 #define jit_prepare_i(ni)	(_jitl.argssize = 0)
@@ -154,7 +169,7 @@ static int jit_arg_reg_order[] = { _EDI, _ESI, _EDX, _ECX };
 #define jit_patch_long_at(jump_pc,v)  (*_PSL((jump_pc) - sizeof(long)) = _jit_SL((jit_insn *)(v)))
 #define jit_patch_short_at(jump_pc,v)  (*_PSI((jump_pc) - sizeof(int)) = _jit_SI((jit_insn *)(v) - (jump_pc)))
 #define jit_patch_at(jump_pc,v) (_jitl.long_jumps ? jit_patch_long_at((jump_pc)-3, v) : jit_patch_short_at(jump_pc, v))
-#define jit_ret() (POPQr(_R13), POPQr(_R12), POPQr(_EBX), POPQr(_EBP), RET_())
+#define jit_ret() (POPQr(_R13), POPQr(_R12), POPQr(_EBX), (_jitl.alloca_offset < -24 ? LEAVE_() : POPQr(_EBP)), RET_())
 
 #define _jit_ldi_l(d, is)		MOVQmr((is), 0,    0,    0,  (d))
 #define jit_ldr_l(d, rs)		MOVQmr(0,    (rs), 0,    0,  (d))

lightning/i386/core-i386.h

@@ -35,6 +35,7 @@
 #define __lightning_core_i386_h
 
 #define JIT_AP			_EBP
+#define JIT_FP			_EBP
 #define JIT_SP			_ESP
 #define JIT_RET			_EAX
 

lightning/ppc/core.h

@@ -42,9 +42,27 @@ struct jit_local_state {
    int	nextarg_geti;  /* Next r20-r25 reg. to be read */
    int	nextarg_getd;  /* The FP args are picked up from FPR1 -> FPR10 */
    int  nbArgs;        /* Number of arguments for the prolog */
+
+   int  frame_size, slack;
+   jit_insn *stwu;
 };
 
+/* Patch a `stwu' instruction (with immediate operand) so that it decreases
+   r1 by AMOUNT.  AMOUNT should already be rounded so that %sp remains quadword
+   aligned.  */
+#define jit_patch_stwu(amount)                               \
+  (*(_jitl.stwu) &= ~_MASK (16),                               \
+   *(_jitl.stwu) |= _s16 ((amount)))
+
+#define jit_allocai(n)							  \
+   (_jitl.frame_size += (n),						  \
+    ((n) <= _jitl.slack							  \
+     ? 0 : jit_patch_stwu (-((_jitl.frame_size + 15) & ~15))),		  \
+    _jitl.slack = ((_jitl.frame_size + 15) & ~15) - _jitl.frame_size,	  \
+    _jitl.frame_size - (n))
+
 #define JIT_SP			1
+#define JIT_FP			1
 #define JIT_RET			3
 #define JIT_R_NUM		3
 #define JIT_V_NUM		7
@@ -52,9 +70,6 @@ struct jit_local_state {
 #define JIT_V(i)		(31-(i))
 #define JIT_AUX			JIT_V(JIT_V_NUM)  /* for 32-bit operands & shift counts */
 
-#define jit_pfx_start()   (_jit.jitl.trampolines)
-#define jit_pfx_end()     (_jit.jitl.free)
-
 /* If possible, use the `small' instruction (rd, rs, imm)
  * else load imm into r26 and use the `big' instruction (rd, rs, r26)
  */

lightning/ppc/funcs.h

@@ -91,13 +91,9 @@ static void
 _jit_epilog(jit_state *jit)
 {
   int n = _jitl.nbArgs;
-  int frame_size, ofs;
   int first_saved_reg = JIT_AUX - n;
   int num_saved_regs = 32 - first_saved_reg;
-
+  int frame_size = (_jitl.frame_size + 15) & ~15;
-  frame_size = 24 + 32 + num_saved_regs * 4;	/* r24..r31 + args		   */
-  frame_size += 15;			/* the stack must be quad-word     */
-  frame_size &= ~15;			/* aligned			   */
 
 #ifdef __APPLE__
   LWZrm(0, frame_size + 8, 1);		/* lwz   r0, x+8(r1)  (ret.addr.)  */
@@ -106,19 +102,12 @@ _jit_epilog(jit_state *jit)
 #endif
   MTLRr(0);				/* mtspr LR, r0			   */
 
-  ofs = frame_size - num_saved_regs * 4;
+  LMWrm(first_saved_reg, 24 + 32, 1);	/* lmw   rI, ofs(r1)		   */
-  LMWrm(first_saved_reg, ofs, 1);	/* lmw   rI, ofs(r1)		   */
   ADDIrri(1, 1, frame_size);		/* addi  r1, r1, x		   */
   BLR();				/* blr				   */
 }
 
 /* Emit a prolog for a function.
-   Upon entrance to the trampoline:
-     - LR      = address where the real code for the function lies
-     - R3-R8   = parameters
-   Upon finishing the trampoline:
-     - R0      = return address for the function
-     - R25-R20 = parameters (order is reversed, 1st argument is R25)
   
    The +32 in frame_size computation is to accound for the parameter area of
    a function frame. 
@@ -126,7 +115,7 @@ _jit_epilog(jit_state *jit)
    On PPC the frame must have space to host the arguments of any callee.
    However, as it currently stands, the argument to jit_trampoline (n) is
    the number of arguments of the caller we generate. Therefore, the
-   callee can overwrite a part of the stack (saved register area when it
+   callee can overwrite a part of the stack (saved register area) when it
    flushes its own parameter on the stack. The addition of a constant 
    offset = 32 is enough to hold eight 4 bytes arguments.  This is less
    than perfect but is a reasonable work around for now. 
@@ -134,8 +123,8 @@ _jit_epilog(jit_state *jit)
 static void
 _jit_prolog(jit_state *jit, int n)
 {
-  int frame_size;
+  int orig_frame_size, frame_size;
-  int ofs, i;
+  int i;
   int first_saved_reg = JIT_AUX - n;
   int num_saved_regs = 32 - first_saved_reg;
 
@@ -143,20 +132,31 @@ _jit_prolog(jit_state *jit, int n)
   _jitl.nextarg_getd = 1;
   _jitl.nbArgs = n;
 
-  frame_size = 24 + 32 + num_saved_regs * 4;	/* r27..r31 + args		   */
-  frame_size += 15;			/* the stack must be quad-word     */
-  frame_size &= ~15;			/* aligned			   */
-
   MFLRr(0);
+
+#ifdef __APPLE__
+  STWrm(0, 8, 1);			/* stw   r0, 8(r1)	   */
+#else
+  STWrm(0, 4, 1);			/* stw   r0, 4(r1)	   */
+#endif
+
+  /* 0..55 -> frame data
+     56..frame_size -> saved registers
+
+     The STMW instruction is patched by jit_allocai, thus leaving
+     the space for the allocai above the 56 bytes.  jit_allocai is
+     also able to reuse the slack space needed to keep the stack
+     quadword-aligned.  */
+
+  _jitl.frame_size = 24 + 32 + num_saved_regs * 4;	/* r27..r31 + args */
+
+  /* The stack must be quad-word aligned.  */
+  frame_size = (_jitl.frame_size + 15) & ~15;
+  _jitl.slack = frame_size - _jitl.frame_size;
+  _jitl.stwu = _jit.x.pc;
   STWUrm(1, -frame_size, 1);		/* stwu  r1, -x(r1)	   */
 
-  ofs = frame_size - num_saved_regs * 4;
+  STMWrm(first_saved_reg, 24 + 32, 1);		/* stmw  rI, ofs(r1)	   */
-  STMWrm(first_saved_reg, ofs, 1);		/* stmw  rI, ofs(r1)		   */
-#ifdef __APPLE__
-  STWrm(0, frame_size + 8, 1);		/* stw   r0, x+8(r1)		   */
-#else
-  STWrm(0, frame_size + 4, 1);		/* stw   r0, x+4(r1)		   */
-#endif
   for (i = 0; i < n; i++)
     MRrr(JIT_AUX-1-i, 3+i);		/* save parameters below r24	   */
 }

tests/Makefile.in

@@ -41,7 +41,7 @@ check_PROGRAMS = fibit$(EXEEXT) incr$(EXEEXT) printf$(EXEEXT) \
 	printf2$(EXEEXT) rpn$(EXEEXT) fib$(EXEEXT) fibdelay$(EXEEXT) \
 	add$(EXEEXT) bp$(EXEEXT) testfp$(EXEEXT) funcfp$(EXEEXT) \
 	rpnfp$(EXEEXT) modi$(EXEEXT) ldxi$(EXEEXT) divi$(EXEEXT) \
-	movi$(EXEEXT) ret$(EXEEXT)
+	movi$(EXEEXT) ret$(EXEEXT) allocai$(EXEEXT)
 subdir = tests
 DIST_COMMON = $(srcdir)/Makefile.am $(srcdir)/Makefile.in
 ACLOCAL_M4 = $(top_srcdir)/aclocal.m4
@@ -56,6 +56,11 @@ add_SOURCES = add.c
 add_OBJECTS = add.$(OBJEXT)
 add_LDADD = $(LDADD)
 @DISASS_TRUE@add_DEPENDENCIES = $(top_builddir)/opcode/libdisass.a
+allocai_SOURCES = allocai.c
+allocai_OBJECTS = allocai.$(OBJEXT)
+allocai_LDADD = $(LDADD)
+@DISASS_TRUE@allocai_DEPENDENCIES =  \
+@DISASS_TRUE@	$(top_builddir)/opcode/libdisass.a
 bp_SOURCES = bp.c
 bp_OBJECTS = bp.$(OBJEXT)
 bp_LDADD = $(LDADD)
@@ -129,12 +134,12 @@ COMPILE = $(CC) $(DEFS) $(DEFAULT_INCLUDES) $(INCLUDES) $(AM_CPPFLAGS) \
 	$(CPPFLAGS) $(AM_CFLAGS) $(CFLAGS)
 CCLD = $(CC)
 LINK = $(CCLD) $(AM_CFLAGS) $(CFLAGS) $(AM_LDFLAGS) $(LDFLAGS) -o $@
-SOURCES = add.c bp.c divi.c fib.c fibdelay.c fibit.c funcfp.c incr.c \
+SOURCES = add.c allocai.c bp.c divi.c fib.c fibdelay.c fibit.c \
-	ldxi.c modi.c movi.c printf.c printf2.c ret.c rpn.c rpnfp.c \
+	funcfp.c incr.c ldxi.c modi.c movi.c printf.c printf2.c ret.c \
-	testfp.c
+	rpn.c rpnfp.c testfp.c
-DIST_SOURCES = add.c bp.c divi.c fib.c fibdelay.c fibit.c funcfp.c \
+DIST_SOURCES = add.c allocai.c bp.c divi.c fib.c fibdelay.c fibit.c \
-	incr.c ldxi.c modi.c movi.c printf.c printf2.c ret.c rpn.c \
+	funcfp.c incr.c ldxi.c modi.c movi.c printf.c printf2.c ret.c \
-	rpnfp.c testfp.c
+	rpn.c rpnfp.c testfp.c
 DATA = $(noinst_DATA)
 ETAGS = etags
 CTAGS = ctags
@@ -242,12 +247,13 @@ target_vendor = @target_vendor@
 AM_CPPFLAGS = -I$(top_builddir) -I$(top_srcdir) -I$(top_srcdir)/lightning/$(cpu)
 noinst_DATA = fibit.ok incr.ok printf.ok printf2.ok rpn.ok	\
 	fib.ok fibdelay.ok testfp.ok funcfp.ok rpnfp.ok add.ok	\
-	bp.ok modi.ok ldxi.ok divi.ok movi.ok ret.ok
+	bp.ok modi.ok ldxi.ok divi.ok movi.ok ret.ok		\
+	allocai.ok
 
 EXTRA_DIST = $(noinst_DATA) run-test
 @DISASS_TRUE@LDADD = $(top_builddir)/opcode/libdisass.a
 @REGRESSION_TESTING_TRUE@TESTS = fib fibit fibdelay incr printf printf2 rpn add bp	\
-@REGRESSION_TESTING_TRUE@	testfp funcfp rpnfp modi ldxi divi movi ret
+@REGRESSION_TESTING_TRUE@	testfp funcfp rpnfp modi ldxi divi movi ret allocai
 
 @REGRESSION_TESTING_TRUE@TESTS_ENVIRONMENT = $(srcdir)/run-test
 all: all-am
@@ -289,6 +295,9 @@ clean-checkPROGRAMS:
 add$(EXEEXT): $(add_OBJECTS) $(add_DEPENDENCIES) 
 	@rm -f add$(EXEEXT)
 	$(LINK) $(add_LDFLAGS) $(add_OBJECTS) $(add_LDADD) $(LIBS)
+allocai$(EXEEXT): $(allocai_OBJECTS) $(allocai_DEPENDENCIES) 
+	@rm -f allocai$(EXEEXT)
+	$(LINK) $(allocai_LDFLAGS) $(allocai_OBJECTS) $(allocai_LDADD) $(LIBS)
 bp$(EXEEXT): $(bp_OBJECTS) $(bp_DEPENDENCIES) 
 	@rm -f bp$(EXEEXT)
 	$(LINK) $(bp_LDFLAGS) $(bp_OBJECTS) $(bp_LDADD) $(LIBS)
@@ -345,6 +354,7 @@ distclean-compile:
 	-rm -f *.tab.c
 
 @AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/add.Po@am__quote@
+@AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/allocai.Po@am__quote@
 @AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/bp.Po@am__quote@
 @AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/divi.Po@am__quote@
 @AMDEP_TRUE@@am__include@ @am__quote@./$(DEPDIR)/fib.Po@am__quote@