X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=ghc%2Frts%2FStgCRun.c;h=fb7179f75eff1e43d2f239c5c5bee0c5f3081b54;hb=49bff3215bf3fe9ada24dac2cf80f97db4e597dd;hp=0ef550c7d30941e9359af03f34e84aec4b2f09aa;hpb=cb5ccf0a7588f9d67835d99a8135b02f3f5ba884;p=ghc-hetmet.git

diff --git a/ghc/rts/StgCRun.c b/ghc/rts/StgCRun.c
index 0ef550c..fb7179f 100644
--- a/ghc/rts/StgCRun.c
+++ b/ghc/rts/StgCRun.c
@@ -1,5 +1,5 @@
 /* -----------------------------------------------------------------------------
- * $Id: StgCRun.c,v 1.34 2002/06/03 13:08:41 matthewc Exp $
+ * $Id: StgCRun.c,v 1.39 2003/06/09 13:17:41 matthewc Exp $
  *
  * (c) The GHC Team, 1998-2000
  *
@@ -84,7 +84,7 @@ register double fake_f9 __asm__("$f9");
 extern StgThreadReturnCode StgRun(StgFunPtr f, StgRegTable *basereg STG_UNUSED)
 {
    while (f) {
-      IF_DEBUG(evaluator,
+      IF_DEBUG(interpreter,
 	       fprintf(stderr,"Jumping to ");
 	       printPtr((P_)f); fflush(stdout);
 	       fprintf(stderr,"\n");
@@ -465,25 +465,25 @@ StgRun(StgFunPtr f, StgRegTable *basereg)
 
 extern StgThreadReturnCode StgRun(StgFunPtr f, StgRegTable *basereg);
 
-void StgRunIsImplementedInAssembler(void)
+static void StgRunIsImplementedInAssembler(void)
 {
 	__asm__ volatile (
 		"\n.globl _StgRun\n"
 		"_StgRun:\n"
 		"\tmflr r0\n"
 		"\tbl saveFP # f14\n"
-		"\tstmw r14,-216(r1)\n"
-		"\tstwu r1,-8480(r1)\n"
+		"\tstmw r13,-220(r1)\n"
+		"\tstwu r1,-%0(r1)\n"
 		"\tmtctr r3\n"
 		"\tmr r12,r3\n"
 		"\tbctr\n"
 		".globl _StgReturn\n"
 		"_StgReturn:\n"
 		"\tmr r3,r14\n"
-		"\tla r1,8480(r1)\n"
-		"\tlmw r14,-216(r1)\n"
+		"\tla r1,%0(r1)\n"
+		"\tlmw r13,-220(r1)\n"
 		"\tb restFP # f14\n"
-	);
+	: : "i"(RESERVED_C_STACK_BYTES+288 /*stack frame size*/));
 }
 
 #endif
@@ -491,7 +491,14 @@ void StgRunIsImplementedInAssembler(void)
 /* -----------------------------------------------------------------------------
    IA64 architecture
 
-   Again, in assembler - so we can fiddle with the register stack.
+   Again, in assembler - so we can fiddle with the register stack, and because
+   gcc doesn't handle asm-clobbered callee-saves correctly.
+
+   loc0  - loc15: preserved locals
+   loc16 - loc28: STG registers
+           loc29: saved ar.pfs
+           loc30: saved b0
+           loc31: saved gp (gcc 3.3 uses this slot)
    -------------------------------------------------------------------------- */
 
 #ifdef ia64_TARGET_ARCH
@@ -500,26 +507,49 @@ void StgRunIsImplementedInAssembler(void)
 #undef RESERVED_C_STACK_BYTES
 #define RESERVED_C_STACK_BYTES 1024
 
-void StgRunIsImplementedInAssembler(void)
+#if ((__GNUC__ == 3) && (__GNUC_MINOR__ >= 3)) || (__GNUC__ > 3)
+/* gcc 3.3+: leave an extra slot for gp saves */
+#define LOCALS 32
+#else
+#define LOCALS 31
+#endif
+
+static void StgRunIsImplementedInAssembler(void)
 {
     __asm__ volatile(
 		".global StgRun\n"
 		"StgRun:\n"
-		"\talloc r55 = ar.pfs, 0, 24, 8, 0\n"	/* setup register frame */
-		"\tmov r54 = b0\n"			/* save return address */
-		"\tadds sp = -%0, sp\n"			/* setup stack */
-		"\tld8 r16=[r32],8 ;;\n"	/* branch to f using descriptor */
-		"\tld8 r1=[r32]\n"
-		"\tmov b6=r16\n"
-		"\tbr.few b6 ;;\n"
+		"\talloc loc29 = ar.pfs, 0, %1, 8, 0\n"	/* setup register frame */
+		"\tld8 r18 = [r32],8\n"			/* get procedure address */
+		"\tadds sp = -%0, sp ;;\n"		/* setup stack */
+		"\tld8 gp = [r32]\n"			/* get procedure GP */
+		"\tadds r16 = %0-(6*16), sp\n"
+		"\tadds r17 = %0-(5*16), sp ;;\n"
+		"\tstf.spill [r16] = f16,32\n"		/* spill callee-saved fp regs */
+		"\tstf.spill [r17] = f17,32\n"
+		"\tmov b6 = r18 ;;\n"			/* set target address */
+		"\tstf.spill [r16] = f18,32\n"
+		"\tstf.spill [r17] = f19,32\n"
+		"\tmov loc30 = b0 ;;\n"			/* save return address */
+		"\tstf.spill [r16] = f20,32\n"
+		"\tstf.spill [r17] = f21,32\n"
+		"\tbr.few b6 ;;\n"			/* branch to function */
 		".global StgReturn\n"
 		"StgReturn:\n"
-		"\tmov r8 = 0\n"		/* return value in r8 */
-		"\tmov ar.pfs = r55\n"		/* restore register frame */
-		"\tmov b0 = r54\n"		/* restore return address */
+		"\tmov r8 = loc16\n"		/* return value in r8 */
+		"\tadds r16 = %0-(6*16), sp\n"
+	    	"\tadds r17 = %0-(5*16), sp ;;\n"
+		"\tldf.fill f16 = [r16],32\n"	/* start restoring fp regs */
+		"\tldf.fill f17 = [r17],32\n"
+		"\tmov ar.pfs = loc29 ;;\n"	/* restore register frame */
+		"\tldf.fill f18 = [r16],32\n"
+		"\tldf.fill f19 = [r17],32\n"
+		"\tmov b0 = loc30 ;;\n"		/* restore return address */
+		"\tldf.fill f20 = [r16],32\n"
+		"\tldf.fill f21 = [r17],32\n"
 		"\tadds sp = %0, sp\n"		/* restore stack */
 		"\tbr.ret.sptk.many b0 ;;\n"	/* return */
-	: : "i"(RESERVED_C_STACK_BYTES));
+	: : "i"(RESERVED_C_STACK_BYTES + 6*16), "i"(LOCALS));
 }
 
 #endif