[AMDGPU][MC] Fix for Bug 28211 + LIT tests

- corrected DS_GWS_* opcodes (see VI_Shader_Programming#16.pdf for detailed description)
  - address operand is not used
  - several opcodes have data operand
  - all opcodes have offset modifier
- DS_AND_SRC2_B32: corrected typo in mnemo
- DS_WRAP_RTN_F32 replaced with DS_WRAP_RTN_B32
- added CI/VI opcodes:
  - DS_CONDXCHG32_RTN_B64
  - DS_GWS_SEMA_RELEASE_ALL
- added VI opcodes:
  - DS_CONSUME
  - DS_APPEND
  - DS_ORDERED_COUNT

Differential Revision: https://reviews.llvm.org/D31707

llvm-svn: 299767

diff --git a/llvm/lib/Target/AMDGPU/DSInstructions.td b/llvm/lib/Target/AMDGPU/DSInstructions.td
index 468d67f..8dad3e8 100644 (file)
--- a/llvm/lib/Target/AMDGPU/DSInstructions.td
+++ b/llvm/lib/Target/AMDGPU/DSInstructions.td
@@ -203,20 +203,28 @@ class DS_1A <string opName> : DS_Pseudo<opName,
   let has_data1 = 0;
 }
 
-class DS_1A_GDS <string opName> : DS_Pseudo<opName,
-  (outs),
-  (ins VGPR_32:$addr),
-  "$addr gds"> {
+class DS_GWS <string opName, dag ins, string asmOps>
+: DS_Pseudo<opName, (outs), ins, asmOps> {
 
-  let has_vdst    = 0;
-  let has_data0   = 0;
-  let has_data1   = 0;
-  let has_offset  = 0;
-  let has_offset0 = 0;
-  let has_offset1 = 0;
+  let has_vdst  = 0;
+  let has_addr  = 0;
+  let has_data0 = 0;
+  let has_data1 = 0;
+
+  let has_gds   = 0;
+  let gdsValue  = 1;
+  let AsmMatchConverter = "cvtDSGds";
+}
+
+class DS_GWS_0D <string opName>
+: DS_GWS<opName,
+  (ins offset:$offset, gds:$gds), "$offset gds">;
+
+class DS_GWS_1D <string opName>
+: DS_GWS<opName,
+  (ins VGPR_32:$data0, offset:$offset, gds:$gds), "$data0$offset gds"> {
 
-  let has_gds     = 0;
-  let gdsValue    = 1;
+  let has_data0 = 1;
 }
 
 class DS_VOID <string opName> : DS_Pseudo<opName,
@@ -390,11 +398,11 @@ def DS_WRXCHG2_RTN_B64     : DS_1A2D_RET<"ds_wrxchg2_rtn_b64", VReg_128, VReg_64
 def DS_WRXCHG2ST64_RTN_B64 : DS_1A2D_RET<"ds_wrxchg2st64_rtn_b64", VReg_128, VReg_64>,
                              AtomicNoRet<"ds_wrxchg2st64_b64", 1>;
 
-def DS_GWS_INIT       : DS_1A_GDS<"ds_gws_init">;
-def DS_GWS_SEMA_V     : DS_1A_GDS<"ds_gws_sema_v">;
-def DS_GWS_SEMA_BR    : DS_1A_GDS<"ds_gws_sema_br">;
-def DS_GWS_SEMA_P     : DS_1A_GDS<"ds_gws_sema_p">;
-def DS_GWS_BARRIER    : DS_1A_GDS<"ds_gws_barrier">;
+def DS_GWS_INIT       : DS_GWS_1D<"ds_gws_init">;
+def DS_GWS_SEMA_V     : DS_GWS_0D<"ds_gws_sema_v">;
+def DS_GWS_SEMA_BR    : DS_GWS_1D<"ds_gws_sema_br">;
+def DS_GWS_SEMA_P     : DS_GWS_0D<"ds_gws_sema_p">;
+def DS_GWS_BARRIER    : DS_GWS_1D<"ds_gws_barrier">;
 
 def DS_ADD_SRC2_U32   : DS_1A<"ds_add_src2_u32">;
 def DS_SUB_SRC2_U32   : DS_1A<"ds_sub_src2_u32">;
@@ -405,7 +413,7 @@ def DS_MIN_SRC2_I32   : DS_1A<"ds_min_src2_i32">;
 def DS_MAX_SRC2_I32   : DS_1A<"ds_max_src2_i32">;
 def DS_MIN_SRC2_U32   : DS_1A<"ds_min_src2_u32">;
 def DS_MAX_SRC2_U32   : DS_1A<"ds_max_src2_u32">;
-def DS_AND_SRC2_B32   : DS_1A<"ds_and_src_b32">;
+def DS_AND_SRC2_B32   : DS_1A<"ds_and_src2_b32">;
 def DS_OR_SRC2_B32    : DS_1A<"ds_or_src2_b32">;
 def DS_XOR_SRC2_B32   : DS_1A<"ds_xor_src2_b32">;
 def DS_MIN_SRC2_F32   : DS_1A<"ds_min_src2_f32">;
@@ -448,25 +456,22 @@ def DS_READ2_B64     : DS_1A_Off8_RET<"ds_read2_b64", VReg_128>;
 def DS_READ2ST64_B64 : DS_1A_Off8_RET<"ds_read2st64_b64", VReg_128>;
 }
 
-let SubtargetPredicate = isSICI in {
 def DS_CONSUME       : DS_0A_RET<"ds_consume">;
 def DS_APPEND        : DS_0A_RET<"ds_append">;
 def DS_ORDERED_COUNT : DS_1A_RET_GDS<"ds_ordered_count">;
-}
 
 //===----------------------------------------------------------------------===//
 // Instruction definitions for CI and newer.
 //===----------------------------------------------------------------------===//
-// Remaining instructions:
-// DS_GWS_SEMA_RELEASE_ALL
-// DS_WRAP_RTN_B32
-// DS_CNDXCHG32_RTN_B64
-// DS_CONDXCHG32_RTN_B128
 
 let SubtargetPredicate = isCIVI in {
 
-def DS_WRAP_RTN_F32 : DS_1A1D_RET <"ds_wrap_rtn_f32">,
-                      AtomicNoRet<"ds_wrap_f32", 1>;
+def DS_WRAP_RTN_B32 : DS_1A2D_RET<"ds_wrap_rtn_b32">, AtomicNoRet<"", 1>;
+
+def DS_CONDXCHG32_RTN_B64 : DS_1A1D_RET<"ds_condxchg32_rtn_b64", VReg_64>,
+                            AtomicNoRet<"", 1>;
+
+def DS_GWS_SEMA_RELEASE_ALL : DS_GWS_0D<"ds_gws_sema_release_all">;
 
 let mayStore = 0 in {
 def DS_READ_B96 : DS_1A_RET<"ds_read_b96", VReg_96>;
@@ -678,8 +683,10 @@ def DS_CMPST_RTN_F32_si   : DS_Real_si<0x31, DS_CMPST_RTN_F32>;
 def DS_MIN_RTN_F32_si     : DS_Real_si<0x32, DS_MIN_RTN_F32>;
 def DS_MAX_RTN_F32_si     : DS_Real_si<0x33, DS_MAX_RTN_F32>;
 
-// FIXME: this instruction is actually CI/VI
-def DS_WRAP_RTN_F32_si    : DS_Real_si<0x34, DS_WRAP_RTN_F32>;
+// These instruction are CI/VI only
+def DS_WRAP_RTN_B32_si    : DS_Real_si<0x34, DS_WRAP_RTN_B32>;
+def DS_CONDXCHG32_RTN_B64_si   : DS_Real_si<0x7e, DS_CONDXCHG32_RTN_B64>;
+def DS_GWS_SEMA_RELEASE_ALL_si : DS_Real_si<0x18, DS_GWS_SEMA_RELEASE_ALL>;
 
 def DS_SWIZZLE_B32_si     : DS_Real_si<0x35, DS_SWIZZLE_B32>;
 def DS_READ_B32_si        : DS_Real_si<0x36, DS_READ_B32>;
@@ -820,11 +827,11 @@ def DS_MIN_F32_vi         : DS_Real_vi<0x12, DS_MIN_F32>;
 def DS_MAX_F32_vi         : DS_Real_vi<0x13, DS_MAX_F32>;
 def DS_NOP_vi             : DS_Real_vi<0x14, DS_NOP>;
 def DS_ADD_F32_vi         : DS_Real_vi<0x15, DS_ADD_F32>;
-def DS_GWS_INIT_vi        : DS_Real_vi<0x19, DS_GWS_INIT>;
-def DS_GWS_SEMA_V_vi      : DS_Real_vi<0x1a, DS_GWS_SEMA_V>;
-def DS_GWS_SEMA_BR_vi     : DS_Real_vi<0x1b, DS_GWS_SEMA_BR>;
-def DS_GWS_SEMA_P_vi      : DS_Real_vi<0x1c, DS_GWS_SEMA_P>;
-def DS_GWS_BARRIER_vi     : DS_Real_vi<0x1d, DS_GWS_BARRIER>;
+def DS_GWS_INIT_vi        : DS_Real_vi<0x99, DS_GWS_INIT>;
+def DS_GWS_SEMA_V_vi      : DS_Real_vi<0x9a, DS_GWS_SEMA_V>;
+def DS_GWS_SEMA_BR_vi     : DS_Real_vi<0x9b, DS_GWS_SEMA_BR>;
+def DS_GWS_SEMA_P_vi      : DS_Real_vi<0x9c, DS_GWS_SEMA_P>;
+def DS_GWS_BARRIER_vi     : DS_Real_vi<0x9d, DS_GWS_BARRIER>;
 def DS_WRITE_B8_vi        : DS_Real_vi<0x1e, DS_WRITE_B8>;
 def DS_WRITE_B16_vi       : DS_Real_vi<0x1f, DS_WRITE_B16>;
 def DS_ADD_RTN_U32_vi     : DS_Real_vi<0x20, DS_ADD_RTN_U32>;
@@ -847,7 +854,7 @@ def DS_CMPST_RTN_B32_vi   : DS_Real_vi<0x30, DS_CMPST_RTN_B32>;
 def DS_CMPST_RTN_F32_vi   : DS_Real_vi<0x31, DS_CMPST_RTN_F32>;
 def DS_MIN_RTN_F32_vi     : DS_Real_vi<0x32, DS_MIN_RTN_F32>;
 def DS_MAX_RTN_F32_vi     : DS_Real_vi<0x33, DS_MAX_RTN_F32>;
-def DS_WRAP_RTN_F32_vi    : DS_Real_vi<0x34, DS_WRAP_RTN_F32>;
+def DS_WRAP_RTN_B32_vi    : DS_Real_vi<0x34, DS_WRAP_RTN_B32>;
 def DS_ADD_RTN_F32_vi     : DS_Real_vi<0x35, DS_ADD_RTN_F32>;
 def DS_READ_B32_vi        : DS_Real_vi<0x36, DS_READ_B32>;
 def DS_READ2_B32_vi       : DS_Real_vi<0x37, DS_READ2_B32>;
@@ -856,6 +863,9 @@ def DS_READ_I8_vi         : DS_Real_vi<0x39, DS_READ_I8>;
 def DS_READ_U8_vi         : DS_Real_vi<0x3a, DS_READ_U8>;
 def DS_READ_I16_vi        : DS_Real_vi<0x3b, DS_READ_I16>;
 def DS_READ_U16_vi        : DS_Real_vi<0x3c, DS_READ_U16>;
+def DS_CONSUME_vi         : DS_Real_vi<0xbd, DS_CONSUME>;
+def DS_APPEND_vi          : DS_Real_vi<0xbe, DS_APPEND>;
+def DS_ORDERED_COUNT_vi   : DS_Real_vi<0xbf, DS_ORDERED_COUNT>;
 def DS_SWIZZLE_B32_vi     : DS_Real_vi<0x3d, DS_SWIZZLE_B32>;
 def DS_PERMUTE_B32_vi     : DS_Real_vi<0x3e, DS_PERMUTE_B32>;
 def DS_BPERMUTE_B32_vi    : DS_Real_vi<0x3f, DS_BPERMUTE_B32>;
@@ -897,6 +907,8 @@ def DS_MSKOR_RTN_B64_vi   : DS_Real_vi<0x6c, DS_MSKOR_RTN_B64>;
 def DS_WRXCHG_RTN_B64_vi  : DS_Real_vi<0x6d, DS_WRXCHG_RTN_B64>;
 def DS_WRXCHG2_RTN_B64_vi : DS_Real_vi<0x6e, DS_WRXCHG2_RTN_B64>;
 def DS_WRXCHG2ST64_RTN_B64_vi : DS_Real_vi<0x6f, DS_WRXCHG2ST64_RTN_B64>;
+def DS_CONDXCHG32_RTN_B64_vi   : DS_Real_vi<0x7e, DS_CONDXCHG32_RTN_B64>;
+def DS_GWS_SEMA_RELEASE_ALL_vi : DS_Real_vi<0x98, DS_GWS_SEMA_RELEASE_ALL>;
 def DS_CMPST_RTN_B64_vi   : DS_Real_vi<0x70, DS_CMPST_RTN_B64>;
 def DS_CMPST_RTN_F64_vi   : DS_Real_vi<0x71, DS_CMPST_RTN_F64>;
 def DS_MIN_RTN_F64_vi     : DS_Real_vi<0x72, DS_MIN_RTN_F64>;

diff --git a/llvm/lib/Target/AMDGPU/InstPrinter/AMDGPUInstPrinter.cpp b/llvm/lib/Target/AMDGPU/InstPrinter/AMDGPUInstPrinter.cpp
index 8d04f83..a817ff3 100644 (file)
--- a/llvm/lib/Target/AMDGPU/InstPrinter/AMDGPUInstPrinter.cpp
+++ b/llvm/lib/Target/AMDGPU/InstPrinter/AMDGPUInstPrinter.cpp
@@ -113,7 +113,7 @@ void AMDGPUInstPrinter::printOffset(const MCInst *MI, unsigned OpNo,
                                     raw_ostream &O) {
   uint16_t Imm = MI->getOperand(OpNo).getImm();
   if (Imm != 0) {
-    O << " offset:";
+    O << ((OpNo == 0)? "offset:" : " offset:");
     printU16ImmDecOperand(MI, OpNo, O);
   }
 }

diff --git a/llvm/test/MC/AMDGPU/ds.s b/llvm/test/MC/AMDGPU/ds.s
index d867b13..f234325 100644 (file)
--- a/llvm/test/MC/AMDGPU/ds.s
+++ b/llvm/test/MC/AMDGPU/ds.s
@@ -140,24 +140,32 @@ ds_max_f32 v2, v4
 // VI:   ds_max_f32 v2, v4 ; encoding: [0x00,0x00,0x26,0xd8,0x02,0x04,0x00,0x00]
 
 ds_gws_init v2 gds
-// SICI: ds_gws_init v2 gds ; encoding: [0x00,0x00,0x66,0xd8,0x02,0x00,0x00,0x00]
-// VI:   ds_gws_init v2 gds ; encoding: [0x00,0x00,0x33,0xd8,0x02,0x00,0x00,0x00]
+// SICI: ds_gws_init v2 gds ; encoding: [0x00,0x00,0x66,0xd8,0x00,0x02,0x00,0x00]
+// VI:   ds_gws_init v2 gds ; encoding: [0x00,0x00,0x33,0xd9,0x00,0x02,0x00,0x00]
 
-ds_gws_sema_v v2 gds
-// SICI: ds_gws_sema_v v2 gds ; encoding: [0x00,0x00,0x6a,0xd8,0x02,0x00,0x00,0x00]
-// VI:   ds_gws_sema_v v2 gds ; encoding: [0x00,0x00,0x35,0xd8,0x02,0x00,0x00,0x00]
+ds_gws_init v3 offset:12345 gds
+// SICI: ds_gws_init v3 offset:12345 gds ; encoding: [0x39,0x30,0x66,0xd8,0x00,0x03,0x00,0x00]
+// VI:   ds_gws_init v3 offset:12345 gds ; encoding: [0x39,0x30,0x33,0xd9,0x00,0x03,0x00,0x00]
+
+ds_gws_sema_v gds
+// SICI: ds_gws_sema_v gds ; encoding: [0x00,0x00,0x6a,0xd8,0x00,0x00,0x00,0x00]
+// VI:   ds_gws_sema_v gds ; encoding: [0x00,0x00,0x35,0xd9,0x00,0x00,0x00,0x00]
+
+ds_gws_sema_v offset:257 gds
+// SICI: ds_gws_sema_v offset:257 gds    ; encoding: [0x01,0x01,0x6a,0xd8,0x00,0x00,0x00,0x00]
+// VI:   ds_gws_sema_v offset:257 gds    ; encoding: [0x01,0x01,0x35,0xd9,0x00,0x00,0x00,0x00]
 
 ds_gws_sema_br v2 gds
-// SICI: ds_gws_sema_br v2 gds ; encoding: [0x00,0x00,0x6e,0xd8,0x02,0x00,0x00,0x00]
-// VI:   ds_gws_sema_br v2 gds ; encoding: [0x00,0x00,0x37,0xd8,0x02,0x00,0x00,0x00]
+// SICI: ds_gws_sema_br v2 gds ; encoding: [0x00,0x00,0x6e,0xd8,0x00,0x02,0x00,0x00]
+// VI:   ds_gws_sema_br v2 gds ; encoding: [0x00,0x00,0x37,0xd9,0x00,0x02,0x00,0x00]
 
-ds_gws_sema_p v2 gds
-// SICI: ds_gws_sema_p v2 gds ; encoding: [0x00,0x00,0x72,0xd8,0x02,0x00,0x00,0x00]
-// VI:   ds_gws_sema_p v2 gds ; encoding: [0x00,0x00,0x39,0xd8,0x02,0x00,0x00,0x00]
+ds_gws_sema_p gds
+// SICI: ds_gws_sema_p gds ; encoding: [0x00,0x00,0x72,0xd8,0x00,0x00,0x00,0x00]
+// VI:   ds_gws_sema_p gds ; encoding: [0x00,0x00,0x39,0xd9,0x00,0x00,0x00,0x00]
 
 ds_gws_barrier v2 gds
-// SICI: ds_gws_barrier v2 gds ; encoding: [0x00,0x00,0x76,0xd8,0x02,0x00,0x00,0x00]
-// VI:   ds_gws_barrier v2 gds ; encoding: [0x00,0x00,0x3b,0xd8,0x02,0x00,0x00,0x00]
+// SICI: ds_gws_barrier v2 gds ; encoding: [0x00,0x00,0x76,0xd8,0x00,0x02,0x00,0x00]
+// VI:   ds_gws_barrier v2 gds ; encoding: [0x00,0x00,0x3b,0xd9,0x00,0x02,0x00,0x00]
 
 ds_write_b8 v2, v4
 // SICI: ds_write_b8 v2, v4 ; encoding: [0x00,0x00,0x78,0xd8,0x02,0x04,0x00,0x00]
@@ -284,17 +292,17 @@ ds_read_u16 v8, v2
 // VI:   ds_read_u16 v8, v2 ; encoding: [0x00,0x00,0x78,0xd8,0x02,0x00,0x00,0x08]
 
 
-//ds_consume v8
-// FIXMESICI: ds_consume v8 ; encoding: [0x00,0x00,0xf4,0xd8,0x00,0x00,0x00,0x08]
-// FIXMEVI:   ds_consume v8 ; encoding: [0x00,0x00,0x7a,0xd8,0x00,0x00,0x00,0x08]
+ds_consume v8
+// SICI: ds_consume v8 ; encoding: [0x00,0x00,0xf4,0xd8,0x00,0x00,0x00,0x08]
+// VI:   ds_consume v8 ; encoding: [0x00,0x00,0x7a,0xd9,0x00,0x00,0x00,0x08]
 
-//ds_append v8
-// FIXMESICI: ds_append v8 ; encoding: [0x00,0x00,0xf8,0xd8,0x00,0x00,0x00,0x08]
-// FIXMEVI:   ds_append v8 ; encoding: [0x00,0x00,0x7c,0xd8,0x00,0x00,0x00,0x08]
+ds_append v8
+// SICI: ds_append v8 ; encoding: [0x00,0x00,0xf8,0xd8,0x00,0x00,0x00,0x08]
+// VI:   ds_append v8 ; encoding: [0x00,0x00,0x7c,0xd9,0x00,0x00,0x00,0x08]
 
-//ds_ordered_count v8, v2 gds
-// FIXMESICI: ds_ordered_count v8, v2 gds ; encoding: [0x00,0x00,0xfe,0xd8,0x02,0x00,0x00,0x08]
-// FIXMEVI:   ds_ordered_count v8, v2 gds ; encoding: [0x00,0x00,0x7f,0xd8,0x02,0x00,0x00,0x08]
+ds_ordered_count v8, v2 gds
+// SICI: ds_ordered_count v8, v2 gds ; encoding: [0x00,0x00,0xfe,0xd8,0x02,0x00,0x00,0x08]
+// VI:   ds_ordered_count v8, v2 gds ; encoding: [0x00,0x00,0x7f,0xd9,0x02,0x00,0x00,0x08]
 
 ds_add_u64 v2, v[4:5]
 // SICI: ds_add_u64 v2, v[4:5] ; encoding: [0x00,0x00,0x00,0xd9,0x02,0x04,0x00,0x00]

diff --git a/llvm/test/MC/AMDGPU/expressions.s b/llvm/test/MC/AMDGPU/expressions.s
index 9fc9566..e593bcd 100644 (file)
--- a/llvm/test/MC/AMDGPU/expressions.s
+++ b/llvm/test/MC/AMDGPU/expressions.s
@@ -11,7 +11,7 @@ s_mov_b32 s0, global
 
 // Use a token with the same name as a global
 ds_gws_init v2 gds
-// VI: ds_gws_init v2 gds ; encoding: [0x00,0x00,0x33,0xd8,0x02,0x00,0x00,0x00]
+// VI: ds_gws_init v2 gds ; encoding: [0x00,0x00,0x33,0xd9,0x00,0x02,0x00,0x00]
 
 // Use a global with the same name as a token
 s_mov_b32 s0, gds

diff --git a/llvm/test/MC/AMDGPU/gfx7_asm_all.s b/llvm/test/MC/AMDGPU/gfx7_asm_all.s
index 09c4b36..d1d864c 100644 (file)
--- a/llvm/test/MC/AMDGPU/gfx7_asm_all.s
+++ b/llvm/test/MC/AMDGPU/gfx7_asm_all.s
@@ -456,19 +456,25 @@ ds_max_f32 v1, v2 offset:65535 gds
 // CHECK: [0xff,0xff,0x4e,0xd8,0x01,0x02,0x00,0x00]
 
 ds_gws_init v1 gds
-// CHECK: [0x00,0x00,0x66,0xd8,0x01,0x00,0x00,0x00]
+// CHECK: [0x00,0x00,0x66,0xd8,0x00,0x01,0x00,0x00]
 
-ds_gws_sema_v v1 gds
-// CHECK: [0x00,0x00,0x6a,0xd8,0x01,0x00,0x00,0x00]
+ds_gws_sema_v gds
+// CHECK: [0x00,0x00,0x6a,0xd8,0x00,0x00,0x00,0x00]
 
 ds_gws_sema_br v1 gds
-// CHECK: [0x00,0x00,0x6e,0xd8,0x01,0x00,0x00,0x00]
+// CHECK: [0x00,0x00,0x6e,0xd8,0x00,0x01,0x00,0x00]
 
-ds_gws_sema_p v1 gds
-// CHECK: [0x00,0x00,0x72,0xd8,0x01,0x00,0x00,0x00]
+ds_gws_sema_p gds
+// CHECK: [0x00,0x00,0x72,0xd8,0x00,0x00,0x00,0x00]
 
 ds_gws_barrier v1 gds
-// CHECK: [0x00,0x00,0x76,0xd8,0x01,0x00,0x00,0x00]
+// CHECK: [0x00,0x00,0x76,0xd8,0x00,0x01,0x00,0x00]
+
+ds_gws_sema_release_all offset:65535 gds
+// CHECK: [0xff,0xff,0x62,0xd8,0x00,0x00,0x00,0x00]
+
+ds_gws_sema_release_all gds
+// CHECK: [0x00,0x00,0x62,0xd8,0x00,0x00,0x00,0x00]
 
 ds_write_b8 v1, v2 offset:65535
 // CHECK: [0xff,0xff,0x78,0xd8,0x01,0x02,0x00,0x00]
@@ -2660,6 +2666,24 @@ ds_max_src2_f64 v1 offset:4
 ds_max_src2_f64 v1 offset:65535 gds
 // CHECK: [0xff,0xff,0x4e,0xdb,0x01,0x00,0x00,0x00]
 
+ds_wrap_rtn_b32 v255, v1, v2, v3 offset:65535
+// CHECK: [0xff,0xff,0xd0,0xd8,0x01,0x02,0x03,0xff]
+
+ds_wrap_rtn_b32 v255, v1, v2, v3 offset:65535 gds
+// CHECK: [0xff,0xff,0xd2,0xd8,0x01,0x02,0x03,0xff]
+
+ds_wrap_rtn_b32 v255, v1, v2, v3
+// CHECK: [0x00,0x00,0xd0,0xd8,0x01,0x02,0x03,0xff]
+
+ds_condxchg32_rtn_b64 v[5:6], v1, v[2:3]
+// CHECK: [0x00,0x00,0xf8,0xd9,0x01,0x02,0x00,0x05]
+
+ds_condxchg32_rtn_b64 v[5:6], v1, v[2:3] gds
+// CHECK: [0x00,0x00,0xfa,0xd9,0x01,0x02,0x00,0x05]
+
+ds_condxchg32_rtn_b64 v[5:6], v1, v[254:255] offset:65535
+// CHECK: [0xff,0xff,0xf8,0xd9,0x01,0xfe,0x00,0x05]
+
 exp mrt0, v0, v0, v0, v0
 // CHECK: [0x0f,0x00,0x00,0xf8,0x00,0x00,0x00,0x00]
 

diff --git a/llvm/test/MC/AMDGPU/gfx8_asm_all.s b/llvm/test/MC/AMDGPU/gfx8_asm_all.s
index 409cafd..2580e6c 100644 (file)
--- a/llvm/test/MC/AMDGPU/gfx8_asm_all.s
+++ b/llvm/test/MC/AMDGPU/gfx8_asm_all.s
@@ -2678,6 +2678,90 @@ ds_max_src2_f64 v1 offset:4
 ds_max_src2_f64 v1 offset:65535 gds
 // CHECK: [0xff,0xff,0xa7,0xd9,0x01,0x00,0x00,0x00]
 
+ds_and_src2_b32 v1
+// CHECK: [0x00,0x00,0x12,0xd9,0x01,0x00,0x00,0x00]
+
+ds_and_src2_b32 v1 gds
+// CHECK: [0x00,0x00,0x13,0xd9,0x01,0x00,0x00,0x00]
+
+ds_and_src2_b32 v255 offset:65535
+// CHECK: [0xff,0xff,0x12,0xd9,0xff,0x00,0x00,0x00]
+
+ds_append v5
+// CHECK: [0x00,0x00,0x7c,0xd9,0x00,0x00,0x00,0x05]
+
+ds_append v5 gds
+// CHECK: [0x00,0x00,0x7d,0xd9,0x00,0x00,0x00,0x05]
+
+ds_append v255 offset:65535
+// CHECK: [0xff,0xff,0x7c,0xd9,0x00,0x00,0x00,0xff]
+
+ds_consume v5
+// CHECK: [0x00,0x00,0x7a,0xd9,0x00,0x00,0x00,0x05]
+
+ds_consume v5 gds
+// CHECK: [0x00,0x00,0x7b,0xd9,0x00,0x00,0x00,0x05]
+
+ds_consume v255 offset:65535
+// CHECK: [0xff,0xff,0x7a,0xd9,0x00,0x00,0x00,0xff]
+
+ds_ordered_count v5, v1 gds
+// CHECK: [0x00,0x00,0x7f,0xd9,0x01,0x00,0x00,0x05]
+
+ds_ordered_count v5, v255 offset:65535 gds
+// CHECK: [0xff,0xff,0x7f,0xd9,0xff,0x00,0x00,0x05]
+
+ds_ordered_count v5, v255 gds
+// CHECK: [0x00,0x00,0x7f,0xd9,0xff,0x00,0x00,0x05]
+
+ds_gws_barrier v1 gds
+// CHECK: [0x00,0x00,0x3b,0xd9,0x00,0x01,0x00,0x00]
+
+ds_gws_barrier v255 offset:65535 gds
+// CHECK: [0xff,0xff,0x3b,0xd9,0x00,0xff,0x00,0x00]
+
+ds_gws_init v1 gds
+// CHECK: [0x00,0x00,0x33,0xd9,0x00,0x01,0x00,0x00]
+
+ds_gws_init v255 offset:65535 gds
+// CHECK: [0xff,0xff,0x33,0xd9,0x00,0xff,0x00,0x00]
+
+ds_gws_sema_br v1 gds
+// CHECK: [0x00,0x00,0x37,0xd9,0x00,0x01,0x00,0x00]
+
+ds_gws_sema_br v255 offset:65535 gds
+// CHECK: [0xff,0xff,0x37,0xd9,0x00,0xff,0x00,0x00]
+
+ds_gws_sema_p offset:65535 gds
+// CHECK: [0xff,0xff,0x39,0xd9,0x00,0x00,0x00,0x00]
+
+ds_gws_sema_p gds
+// CHECK: [0x00,0x00,0x39,0xd9,0x00,0x00,0x00,0x00]
+
+ds_gws_sema_release_all offset:65535 gds
+// CHECK: [0xff,0xff,0x31,0xd9,0x00,0x00,0x00,0x00]
+
+ds_gws_sema_release_all gds
+// CHECK: [0x00,0x00,0x31,0xd9,0x00,0x00,0x00,0x00]
+
+ds_gws_sema_v offset:65535 gds
+// CHECK: [0xff,0xff,0x35,0xd9,0x00,0x00,0x00,0x00]
+
+ds_gws_sema_v gds
+// CHECK: [0x00,0x00,0x35,0xd9,0x00,0x00,0x00,0x00]
+
+ds_wrap_rtn_b32 v5, v255, v2, v3 gds
+// CHECK: [0x00,0x00,0x69,0xd8,0xff,0x02,0x03,0x05]
+
+ds_wrap_rtn_b32 v5, v255, v2, v255 offset:65535
+// CHECK: [0xff,0xff,0x68,0xd8,0xff,0x02,0xff,0x05]
+
+ds_condxchg32_rtn_b64 v[5:6], v1, v[254:255] offset:65535 gds
+// CHECK: [0xff,0xff,0xfd,0xd8,0x01,0xfe,0x00,0x05]
+
+ds_condxchg32_rtn_b64 v[5:6], v1, v[254:255]
+// CHECK: [0x00,0x00,0xfc,0xd8,0x01,0xfe,0x00,0x05]
+
 exp mrt0, v0, v0, v0, v0
 // CHECK: [0x0f,0x00,0x00,0xc4,0x00,0x00,0x00,0x00]
 

diff --git a/llvm/test/MC/Disassembler/AMDGPU/ds_vi.txt b/llvm/test/MC/Disassembler/AMDGPU/ds_vi.txt
index 84d55cd..6d910ea 100644 (file)
--- a/llvm/test/MC/Disassembler/AMDGPU/ds_vi.txt
+++ b/llvm/test/MC/Disassembler/AMDGPU/ds_vi.txt
@@ -81,20 +81,26 @@
 # VI:   ds_max_f32 v2, v4 ; encoding: [0x00,0x00,0x26,0xd8,0x02,0x04,0x00,0x00]
 0x00 0x00 0x26 0xd8 0x02 0x04 0x00 0x00
 
-# VI:   ds_gws_init v2 gds ; encoding: [0x00,0x00,0x33,0xd8,0x02,0x00,0x00,0x00]
-0x00 0x00 0x33 0xd8 0x02 0x00 0x00 0x00
+# VI:   ds_gws_init v2 gds ; encoding: [0x00,0x00,0x33,0xd9,0x00,0x02,0x00,0x00]
+0x00 0x00 0x33 0xd9 0x00 0x02 0x00,0x00
 
-# VI:   ds_gws_sema_v v2 gds ; encoding: [0x00,0x00,0x35,0xd8,0x02,0x00,0x00,0x00]
-0x00 0x00 0x35 0xd8 0x02 0x00 0x00 0x00
+# VI:   ds_gws_init v3 offset:12345 gds ; encoding: [0x39,0x30,0x33,0xd9,0x00,0x03,0x00,0x00]
+0x39 0x30 0x33 0xd9 0x00 0x03 0x00 0x00
 
-# VI:   ds_gws_sema_br v2 gds ; encoding: [0x00,0x00,0x37,0xd8,0x02,0x00,0x00,0x00]
-0x00 0x00 0x37 0xd8 0x02 0x00 0x00 0x00
+# VI:   ds_gws_sema_v gds ; encoding: [0x00,0x00,0x35,0xd9,0x00,0x00,0x00,0x00]
+0x00 0x00 0x35 0xd9 0x00 0x00 0x00 0x00
 
-# VI:   ds_gws_sema_p v2 gds ; encoding: [0x00,0x00,0x39,0xd8,0x02,0x00,0x00,0x00]
-0x00 0x00 0x39 0xd8 0x02 0x00 0x00 0x00
+# VI:   ds_gws_sema_v offset:257 gds    ; encoding: [0x01,0x01,0x35,0xd9,0x00,0x00,0x00,0x00]
+0x01 0x01 0x35 0xd9 0x00 0x00 0x00 0x00
 
-# VI:   ds_gws_barrier v2 gds ; encoding: [0x00,0x00,0x3b,0xd8,0x02,0x00,0x00,0x00]
-0x00 0x00 0x3b 0xd8 0x02 0x00 0x00 0x00
+# VI:   ds_gws_sema_br v2 gds ; encoding: [0x00,0x00,0x37,0xd9,0x00,0x02,0x00,0x00]
+0x00 0x00 0x37 0xd9 0x00 0x02 0x00 0x00
+
+# VI:   ds_gws_sema_p gds ; encoding: [0x00,0x00,0x39,0xd9,0x00,0x00,0x00,0x00]
+0x00 0x00 0x39 0xd9 0x00 0x00 0x00 0x00
+
+# VI:   ds_gws_barrier v2 gds ; encoding: [0x00,0x00,0x3b,0xd9,0x00,0x02,0x00,0x00]
+0x00 0x00 0x3b 0xd9 0x00 0x02 0x00 0x00
 
 # VI:   ds_write_b8 v2, v4 ; encoding: [0x00,0x00,0x3c,0xd8,0x02,0x04,0x00,0x00]
 0x00 0x00 0x3c 0xd8 0x02 0x04 0x00 0x00