[Bug optimization/13250] New: [SH] Gcc code for rotation clobbers the register, but gcc continues to use the register as if it was not clobbered

["uwe at netbsd dot org" <gcc-bugzilla at gcc dot gnu dot org>]

gcc-3.3.2 and 3.4 miscompile sha256 digest code for SuperH, minimized test case 
attached. 
 
The sha256 transformation computes a sum that has (amongst others) the following 
subexpressions: 
 
+ (... ^ ((((e)) >> (25)) | (((e)) << (32 - (25))))) 
+ (((e) & (f)) ^ ((~(e)) & (g))) 
 
Code generated with -O is: 
 
	swap.w	r4,r1		! r4 contains "e" 
	shll8	r4		! "e" in r4 gets clobbered 
	shlr8	r1 
	or	r4,r1 
	rotr	r1		! rotate "e" 25 bits 
	xor	r1,r2 
	mov.l	@(8,r14),r1 
	add	r1,r2		! enclosing sum 
	mov	r4,r3		! <-- gcc thinks r4 still has the value of "e" !!! 
	and	r13,r3		! e & f 
	not	r4,r1		! ~e 
	and	r6,r1		! ~e & g 
	xor	r1,r3		! (e & f) ^ (~e & g) 
	add	r3,r2		! add into the sum

-- 
           Summary: [SH] Gcc code for rotation clobbers the register, but
                    gcc continues to use the register as if it was not
                    clobbered
           Product: gcc
           Version: 3.4
            Status: UNCONFIRMED
          Severity: normal
          Priority: P2
         Component: optimization
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: uwe at netbsd dot org
                CC: gcc-bugs at gcc dot gnu dot org
 GCC build triplet: i386-unknown-freebsd4.8
  GCC host triplet: i386-unknown-freebsd4.8
GCC target triplet: shle--netbsdelf


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=13250