[LLVMdev] Problems with live intervals and spilling when having sub registers?
patrik.h.hagglund at ericsson.com
Fri Oct 21 03:27:45 PDT 2011
I'm having some trouble understanding if the live intervals calculated
for one of my testcases are correct or not.
I have the following instructions:
272L %vreg67:lo<def> = mv_any16 65535; R:%vreg67
288L %vreg64:hi<def> = mv_any16 16383; R:%vreg64
304L %vreg64:lo<def> = COPY %vreg67:lo; R:%vreg64,%vreg67
320L %vreg6<def> = COPY %vreg64<kill>; R:%vreg6,%vreg64
So, %vreg64, which has two (disjunct) sub registers, gets its two parts
written at 288 and 304, and then the whole %vreg64 is copied to %vreg6
at 320. And for this I get the following intervals for %vreg64:
%vreg64 = [288d,304d:1)[304d,320d:0) 0 at 304d 1 at 288d
Later %vreg6 and %vreg64 are merged. The register allocator spills
%vreg6 (fi#14) and the resulting code is
%r0l<def> = mv_any16 65535
Store32FI %r0<kill>, <fi#14>; mem:ST4[FixedStack14] %r0h<def> = mv_any16
16383 %r1<def> = Load32FI <fi#14>; mem:LD4[FixedStack14]
So the whole %r0 register is spilled to the stack after only one half is
written, and then what is supposed to go into the other half is never
used. Instead the previously written value is reloaded from the stack.
I don't know what the live intervals above should look like in the case
when the involved register has sub registers, but it looks suspicious to
me. And if the intervals are indeed as intended, then I suppose the
handling of them in the register allocator have some flaws.
I'm running on version 2.9 using the linear scan allocator.
More information about the llvm-dev