[AMDGPU] Force qsads instrs to use different dest register than source registers

[msearles]

Summary

The V_MQSAD_PK_U16_U8, V_QSAD_PK_U16_U8, and V_MQSAD_U32_U8 take more than 1 pass in hardware. For these three instructions, the destination registers must be different than all sources so that the first pass does not overwrite sources for the following passes.

[arsenm]

test/CodeGen/AMDGPU/llvm.amdgcn.mqsad.pk.u16.u8.ll

7–9

You still need a positive check that the instruction was emitted with the proper register class. I also don't think FileCheck supports matching variables that appear later in the same line

[msearles]

test/CodeGen/AMDGPU/llvm.amdgcn.mqsad.pk.u16.u8.ll

7–9

I'll update the test re: register class. As far as FileCheck goes: it seems to work and documentation seems to say that it should work ( "Variables can also be used later on the same line they were defined on" , which is from http://llvm.org/docs/CommandGuide/FileCheck.html#filecheck-variables ).

[msearles]

Add the positive tests back in

[rampitec]

test/CodeGen/AMDGPU/llvm.amdgcn.mqsad.pk.u16.u8.ll

7–9

Do not you need to switch order of checks? If FileCheck first will match positive pattern it will start matching from there, so negative check will not work.

[arsenm]

test/CodeGen/AMDGPU/llvm.amdgcn.mqsad.pk.u16.u8.ll

2

It would be a good idea to add -enable-var-scope to FileCheck

[msearles]

Update tests

[arsenm]

LGTM

lib/Target/AMDGPU/VOP3Instructions.td

 // Double precision division pre-scale.
 def V_DIV_SCALE_F64 : VOP3_Pseudo <"v_div_scale_f64", VOP3b_F64_I1_F64_F64_F64, [], 1> {
   let SchedRW = [WriteDouble, WriteSALU];
   let hasExtraSrcRegAllocReq = 1;
   let AsmMatchConverter = "";
 }
 
 def V_MSAD_U8 : VOP3Inst <"v_msad_u8", VOP3_Profile<VOP_I32_I32_I32_I32>, int_amdgcn_msad_u8>;
+
+let Constraints = "@earlyclobber $vdst" in {
 def V_MQSAD_PK_U16_U8 : VOP3Inst <"v_mqsad_pk_u16_u8", VOP3_Profile<VOP_I64_I64_I32_I64>, int_amdgcn_mqsad_pk_u16_u8>;
+} // End Constraints = "@earlyclobber $vdst"
 
 def V_TRIG_PREOP_F64 : VOP3Inst <"v_trig_preop_f64", VOP3_Profile<VOP_F64_F64_I32>, AMDGPUtrig_preop> {
   let SchedRW = [WriteDouble];
 }
 
 // These instructions only exist on SI and CI
 let SubtargetPredicate = isSICI in {
 def V_LSHL_B64 : VOP3Inst <"v_lshl_b64", VOP3_Profile<VOP_I64_I64_I32>>;
 def V_LSHR_B64 : VOP3Inst <"v_lshr_b64", VOP3_Profile<VOP_I64_I64_I32>>;
 def V_ASHR_I64 : VOP3Inst <"v_ashr_i64", VOP3_Profile<VOP_I64_I64_I32>>;
 def V_MULLIT_F32 : VOP3Inst <"v_mullit_f32", VOP3_Profile<VOP_F32_F32_F32_F32>>;
 } // End SubtargetPredicate = isSICI
 
 let SubtargetPredicate = isVI in {
 def V_LSHLREV_B64 : VOP3Inst <"v_lshlrev_b64", VOP3_Profile<VOP_I64_I32_I64>>;
 def V_LSHRREV_B64 : VOP3Inst <"v_lshrrev_b64", VOP3_Profile<VOP_I64_I32_I64>>;
 def V_ASHRREV_I64 : VOP3Inst <"v_ashrrev_i64", VOP3_Profile<VOP_I64_I32_I64>>;
 } // End SubtargetPredicate = isVI
 
 
 let SubtargetPredicate = isCIVI in {
 
+let Constraints = "@earlyclobber $vdst" in {
 def V_QSAD_PK_U16_U8 : VOP3Inst <"v_qsad_pk_u16_u8", VOP3_Profile<VOP_I64_I64_I32_I64>, int_amdgcn_qsad_pk_u16_u8>;
 def V_MQSAD_U32_U8 : VOP3Inst <"v_mqsad_u32_u8", VOP3_Profile<VOP_V4I32_I64_I32_V4I32>, int_amdgcn_mqsad_u32_u8>;
+} // End Constraints = "@earlyclobber $vdst"
 
 let isCommutable = 1 in {
 def V_MAD_U64_U32 : VOP3Inst <"v_mad_u64_u32", VOP3b_I64_I1_I32_I32_I64>;
 def V_MAD_I64_I32 : VOP3Inst <"v_mad_i64_i32", VOP3b_I64_I1_I32_I32_I64>;
 } // End isCommutable = 1
 
 } // End SubtargetPredicate = isCIVI
 

test/CodeGen/AMDGPU/llvm.amdgcn.mqsad.pk.u16.u8.ll

 ; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
 ; RUN: llc -march=amdgcn -mcpu=fiji -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
 
 declare i64 @llvm.amdgcn.mqsad.pk.u16.u8(i64, i32, i64) #0
 
 ; GCN-LABEL: {{^}}v_mqsad_pk_u16_u8:
-; GCN: v_mqsad_pk_u16_u8 v[{{[0-9]+:[0-9]+}}], s[{{[0-9]+:[0-9]+}}], v{{[0-9]+}}, v[{{[0-9]+:[0-9]+}}]
+; GCN: v_mqsad_pk_u16_u8 v[0:1], v[4:5], s{{[0-9]+}}, v[{{[0-9]+:[0-9]+}}]
+; GCN-DAG: v_mov_b32_e32 v5, v1
+; GCN-DAG: v_mov_b32_e32 v4, v0
 define amdgpu_kernel void @v_mqsad_pk_u16_u8(i64 addrspace(1)* %out, i64 %src) {
-  %result= call i64 @llvm.amdgcn.mqsad.pk.u16.u8(i64 %src, i32 100, i64 100) #0
-  store i64 %result, i64 addrspace(1)* %out, align 4
+  %tmp = call i64 asm "v_lsrlrev_b64 $0, $1, 1", "={VGPR4_VGPR5},v"(i64 %src) #0
+  %tmp1 = call i64 @llvm.amdgcn.mqsad.pk.u16.u8(i64 %tmp, i32 100, i64 100) #0
+  %tmp2 = call i64 asm ";; force constraint", "=v,{VGPR4_VGPR5}"(i64 %tmp1) #0
+  store i64 %tmp2, i64 addrspace(1)* %out, align 4
   ret void
 }
 
 ; GCN-LABEL: {{^}}v_mqsad_pk_u16_u8_non_immediate:
-; GCN: v_mqsad_pk_u16_u8 v[{{[0-9]+:[0-9]+}}], s[{{[0-9]+:[0-9]+}}], v{{[0-9]+}}, v[{{[0-9]+:[0-9]+}}]
+; GCN: v_mqsad_pk_u16_u8 v[0:1], v[2:3], v4, v[6:7]
+; GCN-DAG: v_mov_b32_e32 v3, v1
+; GCN-DAG: v_mov_b32_e32 v2, v0   
 define amdgpu_kernel void @v_mqsad_pk_u16_u8_non_immediate(i64 addrspace(1)* %out, i64 %src, i32 %a, i64 %b) {
-  %result= call i64 @llvm.amdgcn.mqsad.pk.u16.u8(i64 %src, i32 %a, i64 %b) #0
-  store i64 %result, i64 addrspace(1)* %out, align 4
+  %tmp = call i64 asm "v_lsrlrev_b64 $0, $1, 1", "={VGPR2_VGPR3},v"(i64 %src) #0
+  %tmp1 = call i32 asm "v_mov_b32 $0, $1", "={VGPR4},v"(i32 %a) #0
+  %tmp2 = call i64 asm "v_lshlrev_b64 $0, $1, 1", "={VGPR6_VGPR7},v"(i64 %b) #0
+  %tmp3 = call i64 @llvm.amdgcn.mqsad.pk.u16.u8(i64 %tmp, i32 %tmp1, i64 %tmp2) #0
+  %tmp4 = call i64 asm ";; force constraint", "=v,{VGPR2_VGPR3}"(i64 %tmp3) #0
+  store i64 %tmp4, i64 addrspace(1)* %out, align 4
   ret void
 }
 
-attributes #0 = { nounwind readnone }
+attributes #0 = { nounwind readnone }
 No newline at end of file

test/CodeGen/AMDGPU/llvm.amdgcn.mqsad.u32.u8.ll

 ; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
 ; RUN: llc -march=amdgcn -mcpu=fiji -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
 
 declare <4 x i32> @llvm.amdgcn.mqsad.u32.u8(i64, i32, <4 x i32>) #0
 
-; GCN-LABEL: {{^}}v_mqsad_u32_u8_use_non_inline_constant:
-; GCN: v_mqsad_u32_u8 v[{{[0-9]+:[0-9]+}}], s[{{[0-9]+:[0-9]+}}], v{{[0-9]+}}, v[{{[0-9]+:[0-9]+}}]
-define amdgpu_kernel void @v_mqsad_u32_u8_use_non_inline_constant(<4 x i32> addrspace(1)* %out, i64 %src) {
-  %result = call <4 x i32> @llvm.amdgcn.mqsad.u32.u8(i64 %src, i32 100, <4 x i32> <i32 100, i32 100, i32 100, i32 100>) #0
-  store <4 x i32> %result, <4 x i32> addrspace(1)* %out, align 4
+; GCN-LABEL: {{^}}v_mqsad_u32_u8_inline_integer_immediate:
+; GCN-DAG: v_mov_b32_e32 v0, v2
+; GCN-DAG: v_mov_b32_e32 v1, v3   
+; GCN: v_mqsad_u32_u8 v[2:5], v[0:1], v6, v[{{[0-9]+:[0-9]+}}]
+define amdgpu_kernel void @v_mqsad_u32_u8_inline_integer_immediate(<4 x i32> addrspace(1)* %out, i64 %src, i32 %a) {
+  %tmp = call i64 asm "v_lsrlrev_b64 $0, $1, 1", "={VGPR2_VGPR3},v"(i64 %src) #0
+  %tmp1 = call i32 asm "v_mov_b32 $0, $1", "={VGPR4},v"(i32 %a) #0
+  %tmp2 = call <4 x i32> @llvm.amdgcn.mqsad.u32.u8(i64 %tmp, i32 %tmp1, <4 x i32> <i32 10, i32 20, i32 30, i32 40>) #0
+  %tmp3 = call <4 x i32>  asm ";; force constraint", "=v,{VGPR2_VGPR3_VGPR4_VGPR5}"(<4 x i32> %tmp2) #0
+  store <4 x i32> %tmp3, <4 x i32> addrspace(1)* %out, align 4
   ret void
 }
 
 ; GCN-LABEL: {{^}}v_mqsad_u32_u8_non_immediate:
-; GCN: v_mqsad_u32_u8 v[{{[0-9]+:[0-9]+}}], s[{{[0-9]+:[0-9]+}}], v{{[0-9]+}}, v[{{[0-9]+:[0-9]+}}]
+; GCN-DAG: v_mov_b32_e32 v0, v2
+; GCN-DAG: v_mov_b32_e32 v1, v3   
+; GCN: v_mqsad_u32_u8 v[2:5], v[0:1], v6, v[{{[0-9]+:[0-9]+}}]
 define amdgpu_kernel void @v_mqsad_u32_u8_non_immediate(<4 x i32> addrspace(1)* %out, i64 %src, i32 %a, <4 x i32> %b) {
-  %result = call <4 x i32> @llvm.amdgcn.mqsad.u32.u8(i64 %src, i32 %a, <4 x i32> %b) #0
-  store <4 x i32> %result, <4 x i32> addrspace(1)* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}v_mqsad_u32_u8_inline_integer_immediate:
-; GCN: v_mqsad_u32_u8 v[{{[0-9]+:[0-9]+}}], s[{{[0-9]+:[0-9]+}}], v{{[0-9]+}}, v[{{[0-9]+:[0-9]+}}]
-define amdgpu_kernel void @v_mqsad_u32_u8_inline_integer_immediate(<4 x i32> addrspace(1)* %out, i64 %src, i32 %a) {
-  %result = call <4 x i32> @llvm.amdgcn.mqsad.u32.u8(i64 %src, i32 %a, <4 x i32> <i32 10, i32 20, i32 30, i32 40>) #0
-  store <4 x i32> %result, <4 x i32> addrspace(1)* %out, align 4
+  %tmp = call i64 asm "v_lsrlrev_b64 $0, $1, 1", "={VGPR2_VGPR3},v"(i64 %src) #0
+  %tmp1 = call i32 asm "v_mov_b32 $0, $1", "={VGPR4},v"(i32 %a) #0
+  %tmp2 = call <4 x i32> @llvm.amdgcn.mqsad.u32.u8(i64 %tmp, i32 %tmp1, <4 x i32> %b) #0
+  %tmp3 = call <4 x i32>  asm ";; force constraint", "=v,{VGPR2_VGPR3_VGPR4_VGPR5}"(<4 x i32> %tmp2) #0
+  store <4 x i32> %tmp3, <4 x i32> addrspace(1)* %out, align 4
   ret void
 }
 
 ; GCN-LABEL: {{^}}v_mqsad_u32_u8_inline_fp_immediate:
-; GCN: v_mqsad_u32_u8 v[{{[0-9]+:[0-9]+}}], s[{{[0-9]+:[0-9]+}}], v{{[0-9]+}}, v[{{[0-9]+:[0-9]+}}]
+; GCN-DAG: v_mov_b32_e32 v0, v2
+; GCN-DAG: v_mov_b32_e32 v1, v3   
+; GCN: v_mqsad_u32_u8 v[2:5], v[0:1], v6, v[{{[0-9]+:[0-9]+}}]
 define amdgpu_kernel void @v_mqsad_u32_u8_inline_fp_immediate(<4 x i32> addrspace(1)* %out, i64 %src, i32 %a) {
-  %result = call <4 x i32> @llvm.amdgcn.mqsad.u32.u8(i64 %src, i32 %a, <4 x i32> <i32 1065353216, i32 0, i32 0, i32 0>) #0
-  store <4 x i32> %result, <4 x i32> addrspace(1)* %out, align 4
+  %tmp = call i64 asm "v_lsrlrev_b64 $0, $1, 1", "={VGPR2_VGPR3},v"(i64 %src) #0
+  %tmp1 = call i32 asm "v_mov_b32 $0, $1", "={VGPR4},v"(i32 %a) #0
+  %tmp2 = call <4 x i32> @llvm.amdgcn.mqsad.u32.u8(i64 %tmp, i32 %tmp1, <4 x i32> <i32 1065353216, i32 0, i32 0, i32 0>) #0
+  %tmp3 = call <4 x i32>  asm ";; force constraint", "=v,{VGPR2_VGPR3_VGPR4_VGPR5}"(<4 x i32> %tmp2) #0
+  store <4 x i32> %tmp3, <4 x i32> addrspace(1)* %out, align 4
   ret void
 }
 
 ; GCN-LABEL: {{^}}v_mqsad_u32_u8_use_sgpr_vgpr:
-; GCN: v_mqsad_u32_u8 v[{{[0-9]+:[0-9]+}}], s[{{[0-9]+:[0-9]+}}], v{{[0-9]+}}, v[{{[0-9]+:[0-9]+}}]
+; GCN-DAG: v_mov_b32_e32 v0, v2
+; GCN-DAG: v_mov_b32_e32 v1, v3   
+; GCN: v_mqsad_u32_u8 v[2:5], v[0:1], v6, v[{{[0-9]+:[0-9]+}}]
 define amdgpu_kernel void @v_mqsad_u32_u8_use_sgpr_vgpr(<4 x i32> addrspace(1)* %out, i64 %src, i32 %a, <4 x i32> addrspace(1)* %input) {
   %in = load <4 x i32>, <4 x i32> addrspace(1) * %input
-
-  %result = call <4 x i32> @llvm.amdgcn.mqsad.u32.u8(i64 %src, i32 %a, <4 x i32> %in) #0
-  store <4 x i32> %result, <4 x i32> addrspace(1)* %out, align 4
+  %tmp = call i64 asm "v_lsrlrev_b64 $0, $1, 1", "={VGPR2_VGPR3},v"(i64 %src) #0
+  %tmp1 = call i32 asm "v_mov_b32 $0, $1", "={VGPR4},v"(i32 %a) #0
+  %tmp2 = call <4 x i32> @llvm.amdgcn.mqsad.u32.u8(i64 %tmp, i32 %tmp1, <4 x i32> %in) #0
+  %tmp3 = call <4 x i32>  asm ";; force constraint", "=v,{VGPR2_VGPR3_VGPR4_VGPR5}"(<4 x i32> %tmp2) #0
+  store <4 x i32> %tmp3, <4 x i32> addrspace(1)* %out, align 4
   ret void
 }
 
-attributes #0 = { nounwind readnone }
+attributes #0 = { nounwind readnone }
 No newline at end of file

test/CodeGen/AMDGPU/llvm.amdgcn.qsad.pk.u16.u8.ll

 ; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
 ; RUN: llc -march=amdgcn -mcpu=fiji -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
 
 declare i64 @llvm.amdgcn.qsad.pk.u16.u8(i64, i32, i64) #0
 
 ; GCN-LABEL: {{^}}v_qsad_pk_u16_u8:
-; GCN: v_qsad_pk_u16_u8 v[{{[0-9]+:[0-9]+}}], s[{{[0-9]+:[0-9]+}}], v{{[0-9]+}}, v[{{[0-9]+:[0-9]+}}]
+; GCN: v_qsad_pk_u16_u8 v[0:1], v[4:5], s{{[0-9]+}}, v[{{[0-9]+:[0-9]+}}]
+; GCN-DAG: v_mov_b32_e32 v5, v1
+; GCN-DAG: v_mov_b32_e32 v4, v0
 define amdgpu_kernel void @v_qsad_pk_u16_u8(i64 addrspace(1)* %out, i64 %src) {
-  %result= call i64 @llvm.amdgcn.qsad.pk.u16.u8(i64 %src, i32 100, i64 100) #0
-  store i64 %result, i64 addrspace(1)* %out, align 4
+  %tmp = call i64 asm "v_lsrlrev_b64 $0, $1, 1", "={VGPR4_VGPR5},v"(i64 %src) #0
+  %tmp1 = call i64 @llvm.amdgcn.qsad.pk.u16.u8(i64 %tmp, i32 100, i64 100) #0
+  %tmp2 = call i64 asm ";; force constraint", "=v,{VGPR4_VGPR5}"(i64 %tmp1) #0
+  store i64 %tmp2, i64 addrspace(1)* %out, align 4
   ret void
 }
 
 ; GCN-LABEL: {{^}}v_qsad_pk_u16_u8_non_immediate:
-; GCN: v_qsad_pk_u16_u8 v[{{[0-9]+:[0-9]+}}], s[{{[0-9]+:[0-9]+}}], v{{[0-9]+}}, v[{{[0-9]+:[0-9]+}}]
+; GCN: v_qsad_pk_u16_u8 v[0:1], v[2:3], v4, v[6:7]
+; GCN-DAG: v_mov_b32_e32 v3, v1
+; GCN-DAG: v_mov_b32_e32 v2, v0
 define amdgpu_kernel void @v_qsad_pk_u16_u8_non_immediate(i64 addrspace(1)* %out, i64 %src, i32 %a, i64 %b) {
-  %result= call i64 @llvm.amdgcn.qsad.pk.u16.u8(i64 %src, i32 %a, i64 %b) #0
-  store i64 %result, i64 addrspace(1)* %out, align 4
+  %tmp = call i64 asm "v_lsrlrev_b64 $0, $1, 1", "={VGPR2_VGPR3},v"(i64 %src) #0
+  %tmp1 = call i32 asm "v_mov_b32 $0, $1", "={VGPR4},v"(i32 %a) #0
+  %tmp2 = call i64 asm "v_lshlrev_b64 $0, $1, 1", "={VGPR6_VGPR7},v"(i64 %b) #0
+  %tmp3 = call i64 @llvm.amdgcn.qsad.pk.u16.u8(i64 %tmp, i32 %tmp1, i64 %tmp2) #0
+  %tmp4 = call i64 asm ";; force constraint", "=v,{VGPR2_VGPR3}"(i64 %tmp3) #0
+  store i64 %tmp4, i64 addrspace(1)* %out, align 4
   ret void
 }
 
 attributes #0 = { nounwind readnone }