1module &module:1:0:$full:$large:$default; 2 3/* A test for vector operands. */ 4 5/* { dg-do compile } */ 6/* { dg-options "-fdump-tree-original" } */ 7 8prog kernel &Kernel(kernarg_u64 %input_ptr, kernarg_u64 %output_ptr) 9{ 10 ld_kernarg_u64 $d0, [%input_ptr]; 11 ld_v2_global_f32 ($s0, $s1), [$d0]; 12 ld_v3_global_f32 ($s2, $s3, $s4), [$d0 + 8]; 13 ld_v4_global_f32 ($s5, $s6, $s7, $s8), [$d0 + 20]; 14 15 add_f32 $s9, $s0, $s1; 16 combine_v2_b64_b32 $d2, ($s1, $s0); 17 combine_v2_b64_b32 $d3, ($s2, $s3); 18 19 add_pp_f32x2 $d4, $d2, $d3; 20 21 expand_v2_b32_b64 ($s0, $s3), $d4; 22 23 ld_kernarg_u64 $d1, [%output_ptr]; 24 st_v2_global_f32 ($s0, $s1), [$d1]; 25 st_v3_global_f32 ($s2, $s3, $s4), [$d1 + 8]; 26 st_v4_global_f32 ($s5, $s6, $s7, $s8), [$d1 + 20]; 27 28 ret; 29}; 30 31/* The v2 load is done via casting to a vector datatype ptr. */ 32/* { dg-final { scan-tree-dump " = MEM\\\[\\\(vector\\\(2\\\) <float:32> \\\*\\\)" "original"} } */ 33 34/* The v3 load is scalarized (at the moment) due to gcc requiring 2's exponent wide vectors. */ 35/* { dg-final { scan-tree-dump "s0 = .*BIT_FIELD_REF <mem_read.\[0-9\]+, 32, 0>\\\)?;\[\n ]+s1 = .*BIT_FIELD_REF <mem_read.\[0-9\]+, 32, 32>\\\)?;" "original"} } */ 36 37/* The v4 load is done via casting to a vector datatype ptr. */ 38/* { dg-final { scan-tree-dump " = MEM\\\[\\\(vector\\\(4\\\) <float:32> \\\*\\\)" "original"} } */ 39 40/* The combines are generated to vector constructors. */ 41/* { dg-final { scan-tree-dump "{.*s1\\\)?, .*s0\\\)?}" "original"} } */ 42/* { dg-final { scan-tree-dump "{.*s2\\\)?, .*s3\\\)?}" "original"} } */ 43 44/* Expands to BIT_FIELD_REFs. */ 45/* { dg-final { scan-tree-dump "s0 = \(VIEW_CONVERT_EXPR.*\\\(\)?BIT_FIELD_REF <d4, 32, 0>\\\)?;" "original"} } */ 46/* { dg-final { scan-tree-dump "s3 = \(VIEW_CONVERT_EXPR.*\\\(\)?BIT_FIELD_REF <d4, 32, 32>\\\)?;" "original"} } */ 47 48/* The v1 store is done via casting to a vector datatype ptr and constructing a vector from the inputs. */ 49/* { dg-final { scan-tree-dump "MEM\\\[\\\(vector\\\(2\\\) <float:32> \\\*\\\)\\\(<float:32> \\\*\\\) d1\\\] = " "original"} } */ 50 51/* The v3 store is scalarized (at the moment) due to gcc requiring 2's exponent wide vectors. */ 52/* { dg-final { scan-tree-dump "\\\*\\\(<float:32> \\\*\\\) \\\(\\\(sizetype\\\) d1 \\\+ 8\\\) \\\+ 0 = VIEW_CONVERT_EXPR<<float:32>>\\\(s2\\\);" "original"} } */ 53/* { dg-final { scan-tree-dump "\\\*\\\(<float:32> \\\*\\\) \\\(\\\(sizetype\\\) d1 \\\+ 8\\\) \\\+ 4 = VIEW_CONVERT_EXPR<<float:32>>\\\(s3\\\);" "original"} } */ 54/* { dg-final { scan-tree-dump "\\\*\\\(<float:32> \\\*\\\) \\\(\\\(sizetype\\\) d1 \\\+ 8\\\) \\\+ 8 = VIEW_CONVERT_EXPR<<float:32>>\\\(s4\\\);" "original"} } */ 55 56/* The v4 store is done via casting to a vector datatype and constructing a vector from the inputs. */ 57/* { dg-final { scan-tree-dump "MEM\\\[\\\(vector\\\(4\\\) <float:32> \\\*\\\)\\\(<float:32> \\\*\\\) \\\(\\\(sizetype\\\) d1 \\\+ 20\\\)\\\] = {VIEW_CONVERT_EXPR<<float:32>>\\\(s5\\\), VIEW_CONVERT_EXPR<<float:32>>\\\(s6\\\), VIEW_CONVERT_EXPR<<float:32>>\\\(s7\\\), VIEW_CONVERT_EXPR<<float:32>>\\\(s8\\\)};" "original"} } */ 58