Ian Lance Taylor wrote:
Michael Eager <[EMAIL PROTECTED]> writes:
I'm running into a situation where reload is replacing
a pseudo-register in an insn with a memory reference.
The problem is that this is happening in a memory ref.
The initial pattern is something like
(set (reg/v:SI 1) (mem/s:SI (plus:SI
(reg/f:SI 30)
(const_int 4)) ))
After reload, this becomes
(set (reg/v:SI 1) (mem/s:SI (plus:SI
(mem/f/c:SI (plus:SI (reg/f:SI 19) (const_int -2))
(const_int 4))
The reg-equiv for pseudo-reg 30 is the mem ref in the middle.
Any idea about why this might be happening, or, better,
how to get reload to generate a load from reg-equiv to a reg
rather than replace the pseudo-reg with the reg-equiv?
This is a normal occurrence during the course of reload. If this
escapes from reload, it suggests that your GO_IF_LEGITIMATE_ADDRESS
macro is incorrectly accepting a MEM as a memory address, or that your
insn operand predicate is somehow accepting it.
I checked the GO_IF_LEGITIMATE_ADDRESS macro. It's not seeing
any MEMs. The insn was previously recognized before the following
code at the end of reload replaces the pseudo-reg with the reg-equivalent:
1107 /* Now eliminate all pseudo regs by modifying them into
1108 their equivalent memory references.
1109 The REG-rtx's for the pseudos are modified in place,
1110 so all insns that used to refer to them now refer to memory.
1111
1112 For a reg that has a reg_equiv_address, all those insns
1113 were changed by reloading so that no insns refer to it any longer;
1114 but the DECL_RTL of a variable decl may refer to it,
1115 and if so this causes the debugging info to mention the variable. */
1116
1117 for (i = FIRST_PSEUDO_REGISTER; i < max_regno; i++)
1118 {
1119 rtx addr = 0;
1120
1121 if (reg_equiv_mem[i])
1122 addr = XEXP (reg_equiv_mem[i], 0);
1123
1124 if (reg_equiv_address[i])
1125 addr = reg_equiv_address[i];
1126
1127 if (addr)
1128 {
1129 if (reg_renumber[i] < 0)
1130 {
1131 rtx reg = regno_reg_rtx[i];
1132
1133 REG_USERVAR_P (reg) = 0;
1134 PUT_CODE (reg, MEM);
1135 XEXP (reg, 0) = addr;
1136 if (reg_equiv_memory_loc[i])
1137 MEM_COPY_ATTRIBUTES (reg, reg_equiv_memory_loc[i]);
1138 else
1139 {
1140 MEM_IN_STRUCT_P (reg) = MEM_SCALAR_P (reg) = 0;
1141 MEM_ATTRS (reg) = 0;
1142 }
1143 MEM_NOTRAP_P (reg) = 1;
1144 }
1145 else if (reg_equiv_mem[i])
1146 XEXP (reg_equiv_mem[i], 0) = addr;
1147 }
1148 }
reg_equiv_memory_loc[i] is set in alter_reg() and is a valid
memory reference.
There's nothing in this code which checks the insn which is
being modified to verify that it is still a valid memory ref after
the pseudo-reg is replaced by a memory ref. Apparently, this
code assumes that a pseudo-reg can be replaced by a memory ref
without creating an invalid instruction. I don't find anything
which enforces this assumption.
Perhaps there's some place earlier in reload where this should
have been handled, but don't find it.
I'm building gcc-4.3.2, but the reload code looks the same in
the top of the tree. Any help would be appreciated.
--
Michael Eager [EMAIL PROTECTED]
1960 Park Blvd., Palo Alto, CA 94306 650-325-8077
