[RFC cond-optab] Never try conditional stores in wider modes if pattern fails

[Paolo Bonzini <bonzini at gnu dot org>]

Hello, this patch comes from trying to convert Alpha to use cbranch and
cstore.  It fixes assembly language differences for -mlong-double-128.

It goes like this.  I'm testing this simple function

   void branch_lt_flt (void) { if (flt < 5) f (g ()); }

with -mlong-double-128.  For some reason I've not yet analyzed, the jump
is gimplified to "if (flt < 5.0e+0 != 0)" which requires a cstore.  The
target is !TARGET_FIX, so it is not possible to emit a DFmode cstore.

My patched GCC finds no cstoresf4 pattern.  Then it goes on to find a
cstoredf4 pattern, which is there but fails, and then finds a cstoretf4
pattern, which succeeds because the TFmode comparison libcall has a
DImode result.

Current GCC first looks for a cmpsf pattern and does not find it, then
for a cmpdf pattern and finds it.  Then, since the slt pattern fails,
GCC does *not* go on with a cmptf pattern.

This sort of makes sense in this case, but I wonder if it should always
be like this.

Paolo

2009-03-24  Paolo Bonzini  <bonzini@gnu.org>

	* expmed.c (emit_store_flag): Don't try cstore optab further
	if one existing pattern fails.

Index: gcc/expmed.c
===================================================================
--- gcc/expmed.c        (branch cond-optab)
+++ gcc/expmed.c        (working copy)
@@ -5316,6 +5316,7 @@ emit_store_flag (rtx target, enum rtx_co
            }

          delete_insns_since (last);
+         break;
        }
     }