I have a problem in our 3.0-based backend that I'm not sure how it should be handled.
In summary, I don't know if
* The regalloc verifier is too picky?
* The register allocator is at fault for spilling the full register when
only parts of it are defined?
* Something else?
The architecture I'm compiling for has registers with three disjoint
parts: g, h and l. In one function we receive a value that resides in
the g and h part of a register.
# *** IR Dump Before Linear Scan Register Allocator ***:
# Machine code for function accumconv:
Function Live Ins: %a0_gh in %vreg0, %a1_gh in %vreg1
BB#0: derived from LLVM BB %0
Live Ins: %a0_gh %a1_gh
%vreg1<def> = COPY %a1_gh; aNgh_0_7:%vreg1
%vreg56<def> = mv_any16 0; aNl_0_7:%vreg56
%vreg57<def> = REG_SEQUENCE %vreg1, hi24, %vreg56, lo16;
aN40_0_7:%vreg57 aNgh_0_7:%vreg1 aNl_0_7:%vreg56
So the in-argument in a1_gh is saved in vreg1, and used later in a
REG_SEQUENCE instruction to write a full register, vreg57.
After the REG_SEQUENCE has been eliminated we instead get
16 %vreg57:hi24<def> = COPY %a1_gh<kill>; aN40_0_7:%vreg57
832 %vreg57:lo16<def> = mv_any16 0; aN40_0_7:%vreg57
So both REG_SEQUENCE and the temporary vreg1 have been removed, and
a1_gh is written directly at the right place in vreg57 in the function
entry and the l part of vreg57 is written much later.
During register allocation a new temporary vreg59 seems to be introduced for some part of vreg57 and we get
16L %vreg59:hi24<def,undef> = COPY %a1_gh<kill>, %vreg59<imp-def>;
Due to high register pressure vreg59 is spilled
**** Local spiller rewriting MBB '':
%vreg59:hi24<def,undef> = COPY %a1_gh<kill>, %vreg59<imp-def>;
Store: Store40FI %a1_40<kill>,<fi#0>;
And now the verifier complains:
*** Bad machine code: Using an undefined physical register ***
- function: accumconv
- basic block: 0x97baae0 (BB#0)
- instruction: Store40FI %a1_40<kill>,<fi#0>;
- operand 0: %a1_40<kill>
LLVM ERROR: Found 1 machine code errors.
Since the g and h parts of a1 are live in to the function I suppose that the issue here is that we have a use of the full a1 register in the spill instruction Store40FI, while the l part of it is undefined.