Improve code generation for simple functions

Improve code generation for simple functions in METAL C:

int func(const void* a, const void* b) {
return memcmp(a,b,8);
}

in ILP32 mode the following code is generated
FUNC DS 0F
STM 14,15,12(13)
LR 15,13
L 13,8(,13)
ST 15,4(,13)
@@BGN@2 DS 0H
USING @@AUTO@2,13
* return memcmp(a,b,8);
USING @@PARMD@2,1
L 14,@2a
L 1,@3b
LA 15,0
CLC 0(8,14),0(1)
BRE @2L2
LA 15,1
BRH @2L2
LHI 15,-1
* }
@2L2 DS 0H
DROP
L 13,4(,13)
L 14,12(,13)
BR 14

Similar code (also not preserving GPR 1, inspired by GCC):
USING @@PARMD@2,1
L 15,@a
L 1,@b
CLC 0(8,1),0(15)
IPM 15
SLL 15,2
SRA 15,30
BR 14

in LP64 mode the currently generated code is even less efficient:
FUNC DS 0FD
STMG 14,4,8(13) << 7 registers saved
LGR 15,13
LG 13,136(,13)
STG 15,128(,13)
@@BGN@2 DS 0H
LLILH 4,X'C6F4' << DSA established and not used
OILL 4,X'E2C1' << GPR 4 used to fill in an eyecatcher
ST 4,4(,13) << even though GPR 15 could have been used
USING @@AUTO@2,13
* return memcmp(a,b,8);
USING @@PARMD@2,1
LG 14,@2a
LG 15,@3b
LGHI 0,0
CLC 0(8,14),0(15)
BRE @2L4
LGHI 0,1
BRH @2L4
LGHI 0,-1
@2L4 DS 0H
LGFR 15,0
* }
@2L2 DS 0H
DROP
LG 13,128(,13)
LG 14,8(,13)
LMG 1,4,32(13)

if LP64 linkage actually requires GPR 1 to be preserved, the following code could be generated
USING @@PARMD@2,1
LGR 0,1
LG 15,@a
LG 1,@b
CLC 0(8,1),0(15)
LGR 1,0
IPM 15
SLLG 15,15,34
SRAG 15,15,62
BR 14
if GPR 1 can be changed as well LGRs could be avoided

The inefficiency of the generated code in LP64 mode can be easily seen in the following example
void g() {}
The generated code is
G DS 0FD
STMG 14,4,8(13)
LGR 15,13
LG 13,136(,13)
STG 15,128(,13)
@@BGN@1 DS 0H
LLILH 4,X'C6F4'
OILL 4,X'E2C1'
ST 4,4(,13)
USING @@AUTO@1,13
* }
@1L3 DS 0H
DROP
LG 13,128(,13)
LG 14,8(,13)
LMG 1,4,32(13)
BR 14 <<< only this instruction should have been generated