[PATCH] D21622: RFC: AMDGPU/R600: Add implicitarg.ptr intrinsic

[Jan Vesely via llvm-commits]

jvesely added a comment.

Hi,

this is still an RFC patch, because although it works (passes global_offset piglit with suitably adapted libclc).
It fails with dynamic offsets:

  kernel void local_id(volatile global int *out, int j) {
          __attribute__((address_space(7))) uint * ptr =
                  __builtin_r600_implicitarg_ptr();
          out[0] = ptr[j];
  }

only works if j == 0 (otherwise returns 0 or garbage). Note that changing the last line to

  out[0] = ptr[j+1]

works OK.
The generated assembly looks OK:

  	ALU 1, @8, KC0[CB0:0-32], KC1[]
  	TEX 0 @6
  	ALU 1, @10, KC0[CB0:0-32], KC1[]
  	MEM_RAT_CACHELESS STORE_RAW T0.X, T1.X, 1
  	CF_END
  	PAD
  	Fetch clause starting at 6:
  	VTX_READ_32 T0.X, T0.X, 44
  	ALU clause starting at 8:
  	  LSHL * T0.X, KC0[2].Z, literal.x,  
  	2(2.802597e-45), 0(0.000000e+00)
  	ALU clause starting at 10:
  	  LSHR * T1.X, KC0[2].Y, literal.x,  
  	2(2.802597e-45), 0(0.000000e+00)

is there any interaction between constant and vertex buffer that needs to be handled? (i didn't find anything in the specs) or errata referring to this situation?


Repository:
  rL LLVM

http://reviews.llvm.org/D21622