vararg related bug in custom lowering of FORMAL_ARGUMENTS

[llvmbot]

Bugzilla Link	3470
Version	trunk
OS	Linux
Reporter	LLVM Bugzilla Contributor

Extended Description

The following LLVM IR snippet:

target datalayout = "E-p:32:32:128-i1:8:128-i8:8:128-i16:16:128-i32:32:128-i64:32:128-f32:32:128-f64:64:128-v64:64:64-v128:128:128-a0:0:128-s0:128:128"
target triple = "spu"

define void @​a(...) {
entry:
call void @​b()
ret void
}

Is translated to this assembly code:

    .file   "vararg.bc"
    .text


    .align  3
    .global a
    .type   a, @​function

a:
stqd $lr, 16($sp)
stqd $sp, -32($sp)
ai $sp, $sp, -32
stqd $79, 1280($sp)
stqd $78, 1264($sp)
stqd $77, 1248($sp)
....
stqd $4, 80($sp)
stqd $3, 64($sp)
brasl $lr, b
lqd $lr, 48($sp)
ai $sp, $sp, 32
bi $lr
.size a,.-a

Since function 'a' accepts a variable number of arguments, all argument registers are spilled to the runtime stack.

However, the size of the stack frame is wrong, the space for the spilled registers is not taken into account. This results in a corrupted runtime stack.

There's another, less critical issue here:
$3-$79 are spilled here, but according to the SPU ABI only $3-$74 are used for argument passing, $75-$79 are scratch registers which don't need to be spilled to memory.

[llvmbot]

Clearly the stack adjustment is wrong. Will investigate and fix.

However, I'm going to agree to disagree with the number of registers spilled -- see section 2.2.1 of the v1.4 ABI spec. The table states $3-$79 are volatile and therefore should be spilled.

R3 - R79 volatile First seventy-seven quadwords of a function's argument list and
its return value.

[llvmbot]

Interesting, it looks like the scratch registers were added after 1.4. In the current SPU ABI 1.8 specification it says the following:

R3 – R74 Volatile First 72 quadwords of a function’s argument list and its return value.
R75 – R79 Volatile Scratch Registers.

I just checked what kind of code GCC produces and it also does not spill $75-$79.

[llvmbot]

Looks like I need to update my copy of the ABI. Not that difficult to stop LLVM from spilling 75-79.

Probably a good idea to spill $2, the environment pointer, just in case. The LLVM backend doesn't use it simply to preserve its semantics as an environment pointer (someday, someone will port a Common Lisp or Scheme implementation where this will be important.)

[llvmbot]

The change to eliminate R75 – R79 was done at the time I and others added SPU overlay capability.

[arsenm]

SPU backend was removed