This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
[PowerPC] Avoid ICE on DFmode subreg
- From: Alan Modra <amodra at bigpond dot net dot au>
- To: gcc-patches at gcc dot gnu dot org
- Cc: David Edelsohn <dje at watson dot ibm dot com>
- Date: Wed, 12 Apr 2006 10:19:57 +0930
- Subject: [PowerPC] Avoid ICE on DFmode subreg
This change should fix the AIX testcase failures caused by weird
subregs, by simply not generating them in the first place.
This testcase (-m32 -O2 -mcall-aixdesc -mxl-compat on linux)
extern int g();
void f (void) { g (0, 0.0, 0.0, 0.0, 0.0); }
ICEs since Dale's reg alloc changes when trying to handle the last
double arg stored partially in memory and partially in a gpr as well as
in a fpr.
Bootstrapped and regression tested powerpc-linux and powerpc64-linux.
OK mainline?
* config/rs6000/rs6000.c (rs6000_mixed_function_arg): Update
magic NULL_RTX comment.
(function_arg): Store entire fp arg to mem if any part should go
on stack.
(rs6000_arg_partial_bytes): Adjust for above change.
Index: gcc/config/rs6000/rs6000.c
===================================================================
--- gcc/config/rs6000/rs6000.c (revision 112849)
+++ gcc/config/rs6000/rs6000.c (working copy)
@@ -5058,17 +5072,13 @@ rs6000_mixed_function_arg (enum machine_
if (align_words + n_units > GP_ARG_NUM_REG)
/* Not all of the arg fits in gprs. Say that it goes in memory too,
using a magic NULL_RTX component.
- FIXME: This is not strictly correct. Only some of the arg
- belongs in memory, not all of it. However, there isn't any way
- to do this currently, apart from building rtx descriptions for
- the pieces of memory we want stored. Due to bugs in the generic
- code we can't use the normal function_arg_partial_nregs scheme
- with the PARALLEL arg description we emit here.
- In any case, the code to store the whole arg to memory is often
- more efficient than code to store pieces, and we know that space
- is available in the right place for the whole arg. */
- /* FIXME: This should be fixed since the conversion to
- TARGET_ARG_PARTIAL_BYTES. */
+ This is not strictly correct. Only some of the arg belongs in
+ memory, not all of it. However, the normal scheme using
+ function_arg_partial_nregs can result in unusual subregs, eg.
+ (subreg:SI (reg:DF) 4), which are not handled well. The code to
+ store the whole arg to memory is often more efficient than code
+ to store pieces, and we know that space is available in the right
+ place for the whole arg. */
rvec[k++] = gen_rtx_EXPR_LIST (VOIDmode, NULL_RTX, const0_rtx);
i = 0;
@@ -5310,9 +5320,8 @@ function_arg (CUMULATIVE_ARGS *cum, enum
include the portion actually in registers here. */
enum machine_mode rmode = TARGET_32BIT ? SImode : DImode;
rtx off;
- int i=0;
- if (align_words + n_words > GP_ARG_NUM_REG
- && (TARGET_32BIT && TARGET_POWERPC64))
+ int i = 0;
+ if (align_words + n_words > GP_ARG_NUM_REG)
/* Not all of the arg fits in gprs. Say that it
goes in memory too, using a magic NULL_RTX
component. Also see comment in
@@ -5391,18 +5400,20 @@ rs6000_arg_partial_bytes (CUMULATIVE_ARG
align_words = rs6000_parm_start (mode, type, cum->words);
- if (USE_FP_FOR_ARG_P (cum, mode, type)
+ if (USE_FP_FOR_ARG_P (cum, mode, type))
+ {
/* If we are passing this arg in the fixed parameter save area
(gprs or memory) as well as fprs, then this function should
- return the number of bytes passed in the parameter save area
- rather than bytes passed in fprs. */
- && !(type
- && (cum->nargs_prototype <= 0
- || (DEFAULT_ABI == ABI_AIX
- && TARGET_XL_COMPAT
- && align_words >= GP_ARG_NUM_REG))))
- {
- if (cum->fregno + ((GET_MODE_SIZE (mode) + 7) >> 3) > FP_ARG_MAX_REG + 1)
+ return the number of partial bytes passed in the parameter
+ save area rather than partial bytes passed in fprs. */
+ if (type
+ && (cum->nargs_prototype <= 0
+ || (DEFAULT_ABI == ABI_AIX
+ && TARGET_XL_COMPAT
+ && align_words >= GP_ARG_NUM_REG)))
+ return 0;
+ else if (cum->fregno + ((GET_MODE_SIZE (mode) + 7) >> 3)
+ > FP_ARG_MAX_REG + 1)
ret = (FP_ARG_MAX_REG + 1 - cum->fregno) * 8;
else if (cum->nargs_prototype >= 0)
return 0;
--
Alan Modra
IBM OzLabs - Linux Technology Centre