[AMDGPU] Assembler: SOP* instruction fixes

s_bitset0_b64, s_bitset1_b64 has 32-bit src0, not 64-bit.
s_rfe_b64 has just one destination operand and no source.
Uncomment S_BITCMP* and S_SETVSKIP, adjust SOPC_* classes for that.
Add s_memrealtime test and change comments in smem.s to follow common style.
Change test for s_memtime to use non-zero register to make it really test encoding.
Add tests for s_buffer_load*.
Add tests for SOPC instructions (same for SI and VI)

Differential Revision: http://reviews.llvm.org/D18040

llvm-svn: 263420

6 files changed, 160 insertions, 42 deletions

diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.td b/llvm/lib/Target/AMDGPU/SIInstrInfo.td
index ac3192d5512..24e4b3d441b 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.td
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.td
@@ -901,6 +901,12 @@ multiclass SOP1_32_64 <sop1 op, string opName, list<dag> pattern> : SOP1_m <
     opName#" $sdst, $src0", pattern
 >;
 
+// 32-bit input, 64-bit output.
+multiclass SOP1_64_32 <sop1 op, string opName, list<dag> pattern> : SOP1_m <
+    op, opName, (outs SReg_64:$sdst), (ins SSrc_32:$src0),
+    opName#" $sdst, $src0", pattern
+>;
+
 class SOP2_Pseudo<string opName, dag outs, dag ins, list<dag> pattern> :
   SOP2<outs, ins, "", pattern>,
   SIMCInstr<opName, SISubtarget.NONE> {
@@ -964,19 +970,26 @@ multiclass SOP2_64_32_32 <sop2 op, string opName, list<dag> pattern> : SOP2_m <
     opName#" $sdst, $src0, $src1", pattern
 >;
 
+class SOPC_Base <bits<7> op, RegisterOperand rc0, RegisterOperand rc1,
+                 string opName, list<dag> pattern = []> : SOPC <
+  op, (outs), (ins rc0:$src0, rc1:$src1),
+  opName#" $src0, $src1", pattern > {
+  let Defs = [SCC];
+}
 class SOPC_Helper <bits<7> op, RegisterOperand rc, ValueType vt,
-                    string opName, PatLeaf cond> : SOPC <
-  op, (outs), (ins rc:$src0, rc:$src1),
-  opName#" $src0, $src1",
+                    string opName, PatLeaf cond> : SOPC_Base <
+  op, rc, rc, opName,
   [(set SCC, (si_setcc_uniform vt:$src0, vt:$src1, cond))] > {
-  let Defs = [SCC];
 }
 
-class SOPC_32<bits<7> op, string opName, PatLeaf cond = COND_NULL>
+class SOPC_CMP_32<bits<7> op, string opName, PatLeaf cond = COND_NULL>
   : SOPC_Helper<op, SSrc_32, i32, opName, cond>;
 
-class SOPC_64<bits<7> op, string opName, PatLeaf cond = COND_NULL>
-  : SOPC_Helper<op, SSrc_64, i64, opName, cond>;
+class SOPC_32<bits<7> op, string opName, list<dag> pattern = []>
+  : SOPC_Base<op, SSrc_32, SSrc_32, opName, pattern>;
+
+class SOPC_64_32<bits<7> op, string opName, list<dag> pattern = []>
+  : SOPC_Base<op, SSrc_64, SSrc_32, opName, pattern>;
 
 class SOPK_Pseudo <string opName, dag outs, dag ins, list<dag> pattern> :
   SOPK <outs, ins, "", pattern>,
diff --git a/llvm/lib/Target/AMDGPU/SIInstructions.td b/llvm/lib/Target/AMDGPU/SIInstructions.td
index e598c146280..fff46836ba9 100644
--- a/llvm/lib/Target/AMDGPU/SIInstructions.td
+++ b/llvm/lib/Target/AMDGPU/SIInstructions.td
@@ -168,13 +168,13 @@ defm S_SEXT_I32_I16 : SOP1_32 <sop1<0x1a, 0x17>, "s_sext_i32_i16",
 >;
 
 defm S_BITSET0_B32 : SOP1_32 <sop1<0x1b, 0x18>, "s_bitset0_b32", []>;
-defm S_BITSET0_B64 : SOP1_64 <sop1<0x1c, 0x19>, "s_bitset0_b64", []>;
+defm S_BITSET0_B64 : SOP1_64_32 <sop1<0x1c, 0x19>, "s_bitset0_b64", []>;
 defm S_BITSET1_B32 : SOP1_32 <sop1<0x1d, 0x1a>, "s_bitset1_b32", []>;
-defm S_BITSET1_B64 : SOP1_64 <sop1<0x1e, 0x1b>, "s_bitset1_b64", []>;
+defm S_BITSET1_B64 : SOP1_64_32 <sop1<0x1e, 0x1b>, "s_bitset1_b64", []>;
 defm S_GETPC_B64 : SOP1_64_0 <sop1<0x1f, 0x1c>, "s_getpc_b64", []>;
 defm S_SETPC_B64 : SOP1_1 <sop1<0x20, 0x1d>, "s_setpc_b64", []>;
 defm S_SWAPPC_B64 : SOP1_64 <sop1<0x21, 0x1e>, "s_swappc_b64", []>;
-defm S_RFE_B64 : SOP1_64 <sop1<0x22, 0x1f>, "s_rfe_b64", []>;
+defm S_RFE_B64 : SOP1_1 <sop1<0x22, 0x1f>, "s_rfe_b64", []>;
 
 let hasSideEffects = 1, Uses = [EXEC], Defs = [EXEC, SCC] in {
 
@@ -344,23 +344,23 @@ defm S_ABSDIFF_I32 : SOP2_32 <sop2<0x2c, 0x2a>, "s_absdiff_i32", []>;
 // SOPC Instructions
 //===----------------------------------------------------------------------===//
 
-def S_CMP_EQ_I32 : SOPC_32 <0x00000000, "s_cmp_eq_i32", COND_EQ>;
-def S_CMP_LG_I32 : SOPC_32 <0x00000001, "s_cmp_lg_i32", COND_NE>;
-def S_CMP_GT_I32 : SOPC_32 <0x00000002, "s_cmp_gt_i32", COND_SGT>;
-def S_CMP_GE_I32 : SOPC_32 <0x00000003, "s_cmp_ge_i32", COND_SGE>;
-def S_CMP_LT_I32 : SOPC_32 <0x00000004, "s_cmp_lt_i32", COND_SLT>;
-def S_CMP_LE_I32 : SOPC_32 <0x00000005, "s_cmp_le_i32", COND_SLE>;
-def S_CMP_EQ_U32 : SOPC_32 <0x00000006, "s_cmp_eq_u32", COND_EQ>;
-def S_CMP_LG_U32 : SOPC_32 <0x00000007, "s_cmp_lg_u32", COND_NE >;
-def S_CMP_GT_U32 : SOPC_32 <0x00000008, "s_cmp_gt_u32", COND_UGT>;
-def S_CMP_GE_U32 : SOPC_32 <0x00000009, "s_cmp_ge_u32", COND_UGE>;
-def S_CMP_LT_U32 : SOPC_32 <0x0000000a, "s_cmp_lt_u32", COND_ULT>;
-def S_CMP_LE_U32 : SOPC_32 <0x0000000b, "s_cmp_le_u32", COND_ULE>;
-////def S_BITCMP0_B32 : SOPC_BITCMP0 <0x0000000c, "s_bitcmp0_b32", []>;
-////def S_BITCMP1_B32 : SOPC_BITCMP1 <0x0000000d, "s_bitcmp1_b32", []>;
-////def S_BITCMP0_B64 : SOPC_BITCMP0 <0x0000000e, "s_bitcmp0_b64", []>;
-////def S_BITCMP1_B64 : SOPC_BITCMP1 <0x0000000f, "s_bitcmp1_b64", []>;
-//def S_SETVSKIP : SOPC_ <0x00000010, "s_setvskip", []>;
+def S_CMP_EQ_I32 : SOPC_CMP_32 <0x00000000, "s_cmp_eq_i32", COND_EQ>;
+def S_CMP_LG_I32 : SOPC_CMP_32 <0x00000001, "s_cmp_lg_i32", COND_NE>;
+def S_CMP_GT_I32 : SOPC_CMP_32 <0x00000002, "s_cmp_gt_i32", COND_SGT>;
+def S_CMP_GE_I32 : SOPC_CMP_32 <0x00000003, "s_cmp_ge_i32", COND_SGE>;
+def S_CMP_LT_I32 : SOPC_CMP_32 <0x00000004, "s_cmp_lt_i32", COND_SLT>;
+def S_CMP_LE_I32 : SOPC_CMP_32 <0x00000005, "s_cmp_le_i32", COND_SLE>;
+def S_CMP_EQ_U32 : SOPC_CMP_32 <0x00000006, "s_cmp_eq_u32", COND_EQ>;
+def S_CMP_LG_U32 : SOPC_CMP_32 <0x00000007, "s_cmp_lg_u32", COND_NE >;
+def S_CMP_GT_U32 : SOPC_CMP_32 <0x00000008, "s_cmp_gt_u32", COND_UGT>;
+def S_CMP_GE_U32 : SOPC_CMP_32 <0x00000009, "s_cmp_ge_u32", COND_UGE>;
+def S_CMP_LT_U32 : SOPC_CMP_32 <0x0000000a, "s_cmp_lt_u32", COND_ULT>;
+def S_CMP_LE_U32 : SOPC_CMP_32 <0x0000000b, "s_cmp_le_u32", COND_ULE>;
+def S_BITCMP0_B32 : SOPC_32 <0x0000000c, "s_bitcmp0_b32">;
+def S_BITCMP1_B32 : SOPC_32 <0x0000000d, "s_bitcmp1_b32">;
+def S_BITCMP0_B64 : SOPC_64_32 <0x0000000e, "s_bitcmp0_b64">;
+def S_BITCMP1_B64 : SOPC_64_32 <0x0000000f, "s_bitcmp1_b64">;
+def S_SETVSKIP : SOPC_32 <0x00000010, "s_setvskip">;
 
 //===----------------------------------------------------------------------===//
 // SOPK Instructions
diff --git a/llvm/test/MC/AMDGPU/smem.s b/llvm/test/MC/AMDGPU/smem.s
index 8fa964ca8d1..ab2dcf4b8a1 100644
--- a/llvm/test/MC/AMDGPU/smem.s
+++ b/llvm/test/MC/AMDGPU/smem.s
@@ -3,9 +3,13 @@
 // RUN: not llvm-mc -arch=amdgcn -mcpu=bonaire %s 2>&1 | FileCheck -check-prefix=NOSI %s
 
 s_dcache_wb
-; VI: s_dcache_wb  ; encoding: [0x00,0x00,0x84,0xc0,0x00,0x00,0x00,0x00]
-; NOSI: error: instruction not supported on this GPU
+// VI: s_dcache_wb  ; encoding: [0x00,0x00,0x84,0xc0,0x00,0x00,0x00,0x00]
+// NOSI: error: instruction not supported on this GPU
 
 s_dcache_wb_vol
-; VI: s_dcache_wb_vol ; encoding: [0x00,0x00,0x8c,0xc0,0x00,0x00,0x00,0x00]
-; NOSI: error: instruction not supported on this GPU
+// VI: s_dcache_wb_vol ; encoding: [0x00,0x00,0x8c,0xc0,0x00,0x00,0x00,0x00]
+// NOSI: error: instruction not supported on this GPU
+
+s_memrealtime s[4:5]
+// VI: s_memrealtime s[4:5] ; encoding: [0x00,0x01,0x94,0xc0,0x00,0x00,0x00,0x00]
+// NOSI: error: instruction not supported on this GPU
diff --git a/llvm/test/MC/AMDGPU/smrd.s b/llvm/test/MC/AMDGPU/smrd.s
index 60c8289401e..8ac4f916cb3 100644
--- a/llvm/test/MC/AMDGPU/smrd.s
+++ b/llvm/test/MC/AMDGPU/smrd.s
@@ -80,6 +80,58 @@ s_load_dwordx16 s[88:103], s[2:3], s4
 // GCN: s_load_dwordx16 s[88:103], s[2:3], s4 ; encoding: [0x04,0x02,0x2c,0xc1]
 // NOVI: error: invalid operand for instruction
 
+s_buffer_load_dword s1, s[4:7], 1
+// GCN: s_buffer_load_dword s1, s[4:7], 0x1 ; encoding: [0x01,0x85,0x00,0xc2]
+// VI:	s_buffer_load_dword s1, s[4:7], 0x1    ; encoding: [0x42,0x00,0x22,0xc0,0x01,0x00,0x00,0x00]
+
+s_buffer_load_dword s1, s[4:7], s4
+// GCN: s_buffer_load_dword s1, s[4:7], s4 ; encoding: [0x04,0x84,0x00,0xc2]
+// VI:	s_buffer_load_dword s1, s[4:7], s4     ; encoding: [0x42,0x00,0x20,0xc0,0x04,0x00,0x00,0x00]
+
+s_buffer_load_dwordx2 s[8:9], s[4:7], 1
+// GCN: s_buffer_load_dwordx2 s[8:9], s[4:7], 0x1 ; encoding: [0x01,0x05,0x44,0xc2]
+// VI:	s_buffer_load_dwordx2 s[8:9], s[4:7], 0x1 ; encoding: [0x02,0x02,0x26,0xc0,0x01,0x00,0x00,0x00]
+
+s_buffer_load_dwordx2 s[8:9], s[4:7], s4
+// GCN: s_buffer_load_dwordx2 s[8:9], s[4:7], s4 ; encoding: [0x04,0x04,0x44,0xc2]
+// VI:	s_buffer_load_dwordx2 s[8:9], s[4:7], s4 ; encoding: [0x02,0x02,0x24,0xc0,0x04,0x00,0x00,0x00]
+
+s_buffer_load_dwordx4 s[8:11], s[4:7], 1
+// GCN: s_buffer_load_dwordx4 s[8:11], s[4:7], 0x1 ; encoding: [0x01,0x05,0x84,0xc2]
+// VI:	s_buffer_load_dwordx4 s[8:11], s[4:7], 0x1 ; encoding: [0x02,0x02,0x2a,0xc0,0x01,0x00,0x00,0x00]
+
+s_buffer_load_dwordx4 s[8:11], s[4:7], s4
+// GCN: s_buffer_load_dwordx4 s[8:11], s[4:7], s4 ; encoding: [0x04,0x04,0x84,0xc2]
+// VI:	s_buffer_load_dwordx4 s[8:11], s[4:7], s4 ; encoding: [0x02,0x02,0x28,0xc0,0x04,0x00,0x00,0x00]
+
+s_buffer_load_dwordx4 s[100:103], s[4:7], s4
+// GCN: s_buffer_load_dwordx4 s[100:103], s[4:7], s4 ; encoding: [0x04,0x04,0xb2,0xc2]
+// NOVI: error: invalid operand for instruction
+
+s_buffer_load_dwordx8 s[8:15], s[4:7], 1
+// GCN: s_buffer_load_dwordx8 s[8:15], s[4:7], 0x1 ; encoding: [0x01,0x05,0xc4,0xc2]
+// VI:	s_buffer_load_dwordx8 s[8:15], s[4:7], 0x1 ; encoding: [0x02,0x02,0x2e,0xc0,0x01,0x00,0x00,0x00]
+
+s_buffer_load_dwordx8 s[8:15], s[4:7], s4
+// GCN: s_buffer_load_dwordx8 s[8:15], s[4:7], s4 ; encoding: [0x04,0x04,0xc4,0xc2]
+// VI:	s_buffer_load_dwordx8 s[8:15], s[4:7], s4 ; encoding: [0x02,0x02,0x2c,0xc0,0x04,0x00,0x00,0x00]
+
+s_buffer_load_dwordx8 s[96:103], s[4:7], s4
+// GCN: s_buffer_load_dwordx8 s[96:103], s[4:7], s4 ; encoding: [0x04,0x04,0xf0,0xc2]
+// NOVI: error: invalid operand for instruction
+
+s_buffer_load_dwordx16 s[16:31], s[4:7], 1
+// GCN: s_buffer_load_dwordx16 s[16:31], s[4:7], 0x1 ; encoding: [0x01,0x05,0x08,0xc3]
+// VI:	s_buffer_load_dwordx16 s[16:31], s[4:7], 0x1 ; encoding: [0x02,0x04,0x32,0xc0,0x01,0x00,0x00,0x00]
+
+s_buffer_load_dwordx16 s[16:31], s[4:7], s4
+// GCN: s_buffer_load_dwordx16 s[16:31], s[4:7], s4 ; encoding: [0x04,0x04,0x08,0xc3]
+// VI:	s_buffer_load_dwordx16 s[16:31], s[4:7], s4 ; encoding: [0x02,0x04,0x30,0xc0,0x04,0x00,0x00,0x00]
+
+s_buffer_load_dwordx16 s[88:103], s[4:7], s4
+// GCN: s_buffer_load_dwordx16 s[88:103], s[4:7], s4 ; encoding: [0x04,0x04,0x2c,0xc3]
+// NOVI: error: invalid operand for instruction
+
 s_dcache_inv
 // GCN: s_dcache_inv ; encoding: [0x00,0x00,0xc0,0xc7]
 // VI:	s_dcache_inv                    ; encoding: [0x00,0x00,0x80,0xc0,0x00,0x00,0x00,0x00]
@@ -89,6 +141,6 @@ s_dcache_inv_vol
 // NOSI: error: instruction not supported on this GPU
 // VI: s_dcache_inv_vol                ; encoding: [0x00,0x00,0x88,0xc0,0x00,0x00,0x00,0x00]
 
-s_memtime s[0:1]
-// GCN: s_memtime s[0:1] ; encoding: [0x00,0x00,0x80,0xc7]
-// VI:	s_memtime s[0:1]                ; encoding: [0x00,0x00,0x90,0xc0,0x00,0x00,0x00,0x00]
+s_memtime s[4:5]
+// GCN: s_memtime s[4:5] ; encoding: [0x00,0x00,0x82,0xc7]
+// VI:	s_memtime s[4:5]                ; encoding: [0x00,0x01,0x90,0xc0,0x00,0x00,0x00,0x00]
diff --git a/llvm/test/MC/AMDGPU/sop1.s b/llvm/test/MC/AMDGPU/sop1.s
index 6bfbfd574c1..5488e726c17 100644
--- a/llvm/test/MC/AMDGPU/sop1.s
+++ b/llvm/test/MC/AMDGPU/sop1.s
@@ -121,14 +121,14 @@ s_sext_i32_i16 s1, s2
 s_bitset0_b32 s1, s2
 // CHECK: s_bitset0_b32 s1, s2 ; encoding: [0x02,0x1b,0x81,0xbe]
 
-s_bitset0_b64 s[2:3], s[4:5]
-// CHECK: s_bitset0_b64 s[2:3], s[4:5] ; encoding: [0x04,0x1c,0x82,0xbe]
+s_bitset0_b64 s[2:3], s4
+// CHECK: s_bitset0_b64 s[2:3], s4 ; encoding: [0x04,0x1c,0x82,0xbe]
 
 s_bitset1_b32 s1, s2
 // CHECK: s_bitset1_b32 s1, s2 ; encoding: [0x02,0x1d,0x81,0xbe]
 
-s_bitset1_b64 s[2:3], s[4:5]
-// CHECK: s_bitset1_b64 s[2:3], s[4:5] ; encoding: [0x04,0x1e,0x82,0xbe]
+s_bitset1_b64 s[2:3], s4
+// CHECK: s_bitset1_b64 s[2:3], s4 ; encoding: [0x04,0x1e,0x82,0xbe]
 
 s_getpc_b64 s[2:3]
 // CHECK: s_getpc_b64 s[2:3] ; encoding: [0x00,0x1f,0x82,0xbe]
@@ -139,8 +139,8 @@ s_setpc_b64 s[4:5]
 s_swappc_b64 s[2:3], s[4:5]
 // CHECK: s_swappc_b64 s[2:3], s[4:5] ; encoding: [0x04,0x21,0x82,0xbe]
 
-s_rfe_b64 s[2:3], s[4:5]
-// CHECK: s_rfe_b64 s[2:3], s[4:5] ; encoding: [0x04,0x22,0x82,0xbe]
+s_rfe_b64 s[4:5]
+// CHECK: s_rfe_b64 s[4:5] ; encoding: [0x04,0x22,0x80,0xbe]
 
 s_and_saveexec_b64 s[2:3], s[4:5]
 // CHECK: s_and_saveexec_b64 s[2:3], s[4:5] ; encoding: [0x04,0x24,0x82,0xbe]
diff --git a/llvm/test/MC/AMDGPU/sopc.s b/llvm/test/MC/AMDGPU/sopc.s
index 0899c1a2eed..16416879342 100644
--- a/llvm/test/MC/AMDGPU/sopc.s
+++ b/llvm/test/MC/AMDGPU/sopc.s
@@ -1,9 +1,58 @@
 // RUN: llvm-mc -arch=amdgcn -show-encoding %s | FileCheck %s
-// RUN: llvm-mc -arch=amdgcn -show-encoding %s | FileCheck %s
+// RUN: llvm-mc -arch=amdgcn -mcpu=SI -show-encoding %s | FileCheck %s
+// RUN: llvm-mc -arch=amdgcn -mcpu=fiji -show-encoding %s | FileCheck %s
 
 //===----------------------------------------------------------------------===//
-// Instructions
+// SOPC Instructions
 //===----------------------------------------------------------------------===//
 
 s_cmp_eq_i32 s1, s2
 // CHECK: s_cmp_eq_i32 s1, s2 ; encoding: [0x01,0x02,0x00,0xbf]
+
+s_cmp_lg_i32 s1, s2
+// CHECK: s_cmp_lg_i32 s1, s2 ; encoding: [0x01,0x02,0x01,0xbf]
+
+s_cmp_gt_i32 s1, s2
+// CHECK: s_cmp_gt_i32 s1, s2 ; encoding: [0x01,0x02,0x02,0xbf]
+
+s_cmp_ge_i32 s1, s2
+// CHECK: s_cmp_ge_i32 s1, s2 ; encoding: [0x01,0x02,0x03,0xbf]
+
+s_cmp_lt_i32 s1, s2
+// CHECK: s_cmp_lt_i32 s1, s2 ; encoding: [0x01,0x02,0x04,0xbf]
+
+s_cmp_le_i32 s1, s2
+// CHECK: s_cmp_le_i32 s1, s2 ; encoding: [0x01,0x02,0x05,0xbf]
+
+s_cmp_eq_u32 s1, s2
+// CHECK: s_cmp_eq_u32 s1, s2 ; encoding: [0x01,0x02,0x06,0xbf]
+
+s_cmp_lg_u32 s1, s2
+// CHECK: s_cmp_lg_u32 s1, s2 ; encoding: [0x01,0x02,0x07,0xbf]
+
+s_cmp_gt_u32 s1, s2
+// CHECK: s_cmp_gt_u32 s1, s2 ; encoding: [0x01,0x02,0x08,0xbf]
+
+s_cmp_ge_u32 s1, s2
+// CHECK: s_cmp_ge_u32 s1, s2 ; encoding: [0x01,0x02,0x09,0xbf]
+
+s_cmp_lt_u32 s1, s2
+// CHECK: s_cmp_lt_u32 s1, s2 ; encoding: [0x01,0x02,0x0a,0xbf]
+
+s_cmp_le_u32 s1, s2
+// CHECK: s_cmp_le_u32 s1, s2 ; encoding: [0x01,0x02,0x0b,0xbf]
+
+s_bitcmp0_b32 s1, s2
+// CHECK: s_bitcmp0_b32 s1, s2 ; encoding: [0x01,0x02,0x0c,0xbf]
+
+s_bitcmp1_b32 s1, s2
+// CHECK: s_bitcmp1_b32 s1, s2 ; encoding: [0x01,0x02,0x0d,0xbf]
+
+s_bitcmp0_b64 s[2:3], s4
+// CHECK: s_bitcmp0_b64 s[2:3], s4 ; encoding: [0x02,0x04,0x0e,0xbf]
+
+s_bitcmp1_b64 s[2:3], s4
+// CHECK: s_bitcmp1_b64 s[2:3], s4 ; encoding: [0x02,0x04,0x0f,0xbf]
+
+s_setvskip s3, s5
+// CHECK: s_setvskip s3, s5 ; encoding: [0x03,0x05,0x10,0xbf]