[Bug rtl-optimization/52657] New: [4.7/4.8 regression] error on asm during GMP build

[ebotcazou at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52657

             Bug #: 52657
           Summary: [4.7/4.8 regression] error on asm during GMP build
    Classification: Unclassified
           Product: gcc
           Version: 4.7.0
            Status: UNCONFIRMED
          Keywords: ice-on-valid-code
          Severity: normal
          Priority: P3
         Component: rtl-optimization
        AssignedTo: unassig...@gcc.gnu.org
        ReportedBy: ebotca...@gcc.gnu.org
            Target: ia64-*-linux


Created attachment 26944
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=26944
Reduced testcase

GMP cannot be compiled anymore on IA-64/Linux, the error being:

gcc -std=gnu99 -DHAVE_CONFIG_H -I. -I. -I.. -D__GMP_WITHIN_GMP -I..
-DOPERATION_redc_2 -O2 -pedantic -c 
redc_2.c -o redc_2.o
redc_2.c: In function '__gmpn_redc_2':
redc_2.c:87:7: error: 'asm' operand requires impossible reload
redc_2.c:87:7: error: 'asm' operand requires impossible reload
redc_2.c:87:7: error: 'asm' operand requires impossible reload
redc_2.c:87:7: error: 'asm' operand requires impossible reload
make[2]: *** [redc_2.lo] Error 1

The error is issued by this very old piece of code in reload_as_needed:

          /* If this was an ASM, make sure that all the reload insns
         we have generated are valid.  If not, give an error
         and delete them.  */
          if (asm_noperands (PATTERN (insn)) >= 0)
        for (p = NEXT_INSN (prev); p != next; p = NEXT_INSN (p))
          if (p != insn && INSN_P (p)
              && GET_CODE (PATTERN (p)) != USE
              && (recog_memoized (p) < 0
              || (extract_insn (p), ! constrain_operands (1))))
            {
              error_for_asm (insn,
                     "%<asm%> operand requires "
                     "impossible reload");
              delete_insn (p);
            }

Now it seems to overlook the case where the reload insns contain pseudos that 
didn't get hard regs and thus have equivalent memory references.  They will be 
replaced by these memory references only _after_ reload_as_needed has run.

Commenting out the piece code generates correct assembly AFAICS and the GMP 
testsuite is clean.  But it seems quite astonishing that this issue has never 
surfaced until now (the code was added by Richard Kenner almost 20 years ago).