R600/SI: Fix implicit vcc operand to v_div_fmas_*

This should allow finally fixing the f64 fdiv implementation.

Test is disabled for VI since there seems to be a problem with one
of the buffer load instructions on it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229236 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/R600/AMDGPUISelLowering.cpp b/lib/Target/R600/AMDGPUISelLowering.cpp
index d577d689f66aaa6e4d2ad6d15e6902fac8958900..d96f03aaeb0545c2e4018d4780178014e9e27152 100644 (file)
--- a/lib/Target/R600/AMDGPUISelLowering.cpp
+++ b/lib/Target/R600/AMDGPUISelLowering.cpp
@@ -907,10 +907,9 @@ SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
     }
 
     case Intrinsic::AMDGPU_div_fmas:
-      // FIXME: Dropping bool parameter. Work is needed to support the implicit
-      // read from VCC.
       return DAG.getNode(AMDGPUISD::DIV_FMAS, DL, VT,
-                         Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
+                         Op.getOperand(1), Op.getOperand(2), Op.getOperand(3),
+                         Op.getOperand(4));
 
     case Intrinsic::AMDGPU_div_fixup:
       return DAG.getNode(AMDGPUISD::DIV_FIXUP, DL, VT,

diff --git a/lib/Target/R600/AMDGPUInstrInfo.td b/lib/Target/R600/AMDGPUInstrInfo.td
index 0e34392bd50d0cbcd5f21cb3b99505386862b848..d657ad05c8c44d17ede31b8682de0a821980c829 100644 (file)
--- a/lib/Target/R600/AMDGPUInstrInfo.td
+++ b/lib/Target/R600/AMDGPUInstrInfo.td
@@ -35,6 +35,11 @@ def AMDGPUDivScaleOp : SDTypeProfile<2, 3,
   [SDTCisFP<0>, SDTCisInt<1>, SDTCisSameAs<0, 2>, SDTCisSameAs<0, 3>, SDTCisSameAs<0, 4>]
 >;
 
+// float, float, float, vcc
+def AMDGPUFmasOp : SDTypeProfile<1, 4,
+  [SDTCisFP<0>, SDTCisSameAs<0, 1>, SDTCisSameAs<0, 2>, SDTCisSameAs<0, 3>, SDTCisInt<4>]
+>;
+
 //===----------------------------------------------------------------------===//
 // AMDGPU DAG Nodes
 //
@@ -153,7 +158,7 @@ def AMDGPUdiv_scale : SDNode<"AMDGPUISD::DIV_SCALE", AMDGPUDivScaleOp>;
 
 //  Special case divide FMA with scale and flags (src0 = Quotient,
 //  src1 = Denominator, src2 = Numerator).
-def AMDGPUdiv_fmas : SDNode<"AMDGPUISD::DIV_FMAS", SDTFPTernaryOp>;
+def AMDGPUdiv_fmas : SDNode<"AMDGPUISD::DIV_FMAS", AMDGPUFmasOp>;
 
 // Single or double precision division fixup.
 // Special case divide fixup and flags(src0 = Quotient, src1 =

diff --git a/lib/Target/R600/SIInstrInfo.td b/lib/Target/R600/SIInstrInfo.td
index cb16be5f2f70c8a2ac8c188532fac46355d264a4..4082b570c41fb4f173453d8bfb2d793118611bcc 100644 (file)
--- a/lib/Target/R600/SIInstrInfo.td
+++ b/lib/Target/R600/SIInstrInfo.td
@@ -1300,6 +1300,28 @@ multiclass VOP3Inst <vop3 op, string opName, VOPProfile P,
   P.NumSrcArgs, P.HasModifiers
 >;
 
+// Special case for v_div_fmas_{f32|f64}, since it seems to be the
+// only VOP instruction that implicitly reads VCC.
+multiclass VOP3_VCC_Inst <vop3 op, string opName,
+                          VOPProfile P,
+                          SDPatternOperator node = null_frag> : VOP3_Helper <
+  op, opName,
+  P.Outs,
+  (ins InputModsNoDefault:$src0_modifiers, P.Src0RC64:$src0,
+       InputModsNoDefault:$src1_modifiers, P.Src1RC64:$src1,
+       InputModsNoDefault:$src2_modifiers, P.Src2RC64:$src2,
+       ClampMod:$clamp,
+       omod:$omod),
+  " $dst, $src0_modifiers, $src1_modifiers, $src2_modifiers"#"$clamp"#"$omod",
+  [(set P.DstVT:$dst,
+            (node (P.Src0VT (VOP3Mods0 P.Src0VT:$src0, i32:$src0_modifiers,
+                                       i1:$clamp, i32:$omod)),
+                  (P.Src1VT (VOP3Mods P.Src1VT:$src1, i32:$src1_modifiers)),
+                  (P.Src2VT (VOP3Mods P.Src2VT:$src2, i32:$src2_modifiers)),
+                  (i1 VCC)))],
+  3, 1
+>;
+
 multiclass VOP3b_Helper <vop op, RegisterClass vrc, RegisterOperand arc,
                     string opName, list<dag> pattern> :
   VOP3b_3_m <

diff --git a/lib/Target/R600/SIInstructions.td b/lib/Target/R600/SIInstructions.td
index 032d6c2abfcffc18cfd5cd9b8bdd60095a1a593a..0f14dc31136041d5f06bba110b5473d69ed6d1cd 100644 (file)
--- a/lib/Target/R600/SIInstructions.td
+++ b/lib/Target/R600/SIInstructions.td
@@ -1744,14 +1744,27 @@ let SchedRW = [WriteDouble, WriteSALU] in {
 defm V_DIV_SCALE_F64 : VOP3b_64 <vop3<0x16e, 0x1e1>, "v_div_scale_f64", []>;
 } // let SchedRW = [WriteDouble]
 
-let isCommutable = 1 in {
-defm V_DIV_FMAS_F32 : VOP3Inst <vop3<0x16f, 0x1e2>, "v_div_fmas_f32",
+let isCommutable = 1, Uses = [VCC] in {
+
+// v_div_fmas_f32:
+//   result = src0 * src1 + src2
+//   if (vcc)
+//     result *= 2^32
+//
+defm V_DIV_FMAS_F32 : VOP3_VCC_Inst <vop3<0x16f, 0x1e2>, "v_div_fmas_f32",
   VOP_F32_F32_F32_F32, AMDGPUdiv_fmas
 >;
+
 let SchedRW = [WriteDouble] in {
-defm V_DIV_FMAS_F64 : VOP3Inst <vop3<0x170, 0x1e3>, "v_div_fmas_f64",
+// v_div_fmas_f64:
+//   result = src0 * src1 + src2
+//   if (vcc)
+//     result *= 2^64
+//
+defm V_DIV_FMAS_F64 : VOP3_VCC_Inst <vop3<0x170, 0x1e3>, "v_div_fmas_f64",
   VOP_F64_F64_F64_F64, AMDGPUdiv_fmas
 >;
+
 } // End SchedRW = [WriteDouble]
 } // End isCommutable = 1
 

diff --git a/test/CodeGen/R600/llvm.AMDGPU.div_fmas.ll b/test/CodeGen/R600/llvm.AMDGPU.div_fmas.ll
index a15c10fab513494d1f6ae4ff01154a9403f46269..ceffd05421520bcfddfef91216953ceb2ba9c407 100644 (file)
--- a/test/CodeGen/R600/llvm.AMDGPU.div_fmas.ll
+++ b/test/CodeGen/R600/llvm.AMDGPU.div_fmas.ll
@@ -1,6 +1,10 @@
 ; RUN: llc -march=amdgcn -mcpu=SI -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=SI %s
-; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VI %s
+; XUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VI %s
 
+; FIXME: Enable for VI.
+
+declare i32 @llvm.r600.read.tidig.x() nounwind readnone
+declare void @llvm.AMDGPU.barrier.global() nounwind noduplicate
 declare float @llvm.AMDGPU.div.fmas.f32(float, float, float, i1) nounwind readnone
 declare double @llvm.AMDGPU.div.fmas.f64(double, double, double, i1) nounwind readnone
 
@@ -13,7 +17,8 @@ declare double @llvm.AMDGPU.div.fmas.f64(double, double, double, i1) nounwind re
 ; VI-DAG: s_load_dword [[SB:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0x30
 ; GCN-DAG: v_mov_b32_e32 [[VC:v[0-9]+]], [[SC]]
 ; GCN-DAG: v_mov_b32_e32 [[VB:v[0-9]+]], [[SB]]
-; GCN: v_div_fmas_f32 [[RESULT:v[0-9]+]], [[SA]], [[VB]], [[VC]]
+; GCN-DAG: v_mov_b32_e32 [[VA:v[0-9]+]], [[SA]]
+; GCN: v_div_fmas_f32 [[RESULT:v[0-9]+]], [[VA]], [[VB]], [[VC]]
 ; GCN: buffer_store_dword [[RESULT]],
 ; GCN: s_endpgm
 define void @test_div_fmas_f32(float addrspace(1)* %out, float %a, float %b, float %c, i1 %d) nounwind {
@@ -29,3 +34,104 @@ define void @test_div_fmas_f64(double addrspace(1)* %out, double %a, double %b,
   store double %result, double addrspace(1)* %out, align 8
   ret void
 }
+
+; GCN-LABEL: {{^}}test_div_fmas_f32_cond_to_vcc:
+; SI: v_cmp_eq_i32_e64 vcc, s{{[0-9]+}}, 0
+; SI: v_div_fmas_f32 {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}
+define void @test_div_fmas_f32_cond_to_vcc(float addrspace(1)* %out, float %a, float %b, float %c, i32 %i) nounwind {
+  %cmp = icmp eq i32 %i, 0
+  %result = call float @llvm.AMDGPU.div.fmas.f32(float %a, float %b, float %c, i1 %cmp) nounwind readnone
+  store float %result, float addrspace(1)* %out, align 4
+  ret void
+}
+
+; GCN-LABEL: {{^}}test_div_fmas_f32_imm_false_cond_to_vcc:
+; SI: s_mov_b64 vcc, 0
+; SI: v_div_fmas_f32 {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}
+define void @test_div_fmas_f32_imm_false_cond_to_vcc(float addrspace(1)* %out, float %a, float %b, float %c) nounwind {
+  %result = call float @llvm.AMDGPU.div.fmas.f32(float %a, float %b, float %c, i1 false) nounwind readnone
+  store float %result, float addrspace(1)* %out, align 4
+  ret void
+}
+
+; GCN-LABEL: {{^}}test_div_fmas_f32_imm_true_cond_to_vcc:
+; SI: s_mov_b64 vcc, -1
+; SI: v_div_fmas_f32 {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}
+define void @test_div_fmas_f32_imm_true_cond_to_vcc(float addrspace(1)* %out, float %a, float %b, float %c) nounwind {
+  %result = call float @llvm.AMDGPU.div.fmas.f32(float %a, float %b, float %c, i1 true) nounwind readnone
+  store float %result, float addrspace(1)* %out, align 4
+  ret void
+}
+
+; GCN-LABEL: {{^}}test_div_fmas_f32_logical_cond_to_vcc:
+; SI-DAG: buffer_load_dword [[A:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
+; SI-DAG: buffer_load_dword [[B:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:4{{$}}
+; SI-DAG: buffer_load_dword [[C:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
+
+; SI-DAG: v_cmp_eq_i32_e64 [[CMP0:s\[[0-9]+:[0-9]+\]]], v{{[0-9]+}}, 0
+; SI-DAG: v_cmp_ne_i32_e64 [[CMP1:s\[[0-9]+:[0-9]+\]]], s{{[0-9]+}}, 0
+; SI: s_and_b64 vcc, [[CMP0]], [[CMP1]]
+; SI: v_div_fmas_f32 {{v[0-9]+}}, [[B]], [[A]], [[C]]
+; SI: s_endpgm
+define void @test_div_fmas_f32_logical_cond_to_vcc(float addrspace(1)* %out, float addrspace(1)* %in, i32 %d) nounwind {
+  %tid = call i32 @llvm.r600.read.tidig.x() nounwind readnone
+  %gep.a = getelementptr float addrspace(1)* %in, i32 %tid
+  %gep.b = getelementptr float addrspace(1)* %gep.a, i32 1
+  %gep.c = getelementptr float addrspace(1)* %gep.a, i32 2
+  %gep.out = getelementptr float addrspace(1)* %out, i32 2
+
+  %a = load float addrspace(1)* %gep.a
+  %b = load float addrspace(1)* %gep.b
+  %c = load float addrspace(1)* %gep.c
+
+  %cmp0 = icmp eq i32 %tid, 0
+  %cmp1 = icmp ne i32 %d, 0
+  %and = and i1 %cmp0, %cmp1
+
+  %result = call float @llvm.AMDGPU.div.fmas.f32(float %a, float %b, float %c, i1 %and) nounwind readnone
+  store float %result, float addrspace(1)* %gep.out, align 4
+  ret void
+}
+
+; GCN-LABEL: {{^}}test_div_fmas_f32_i1_phi_vcc:
+; SI: v_cmp_eq_i32_e64 [[CMPTID:s\[[0-9]+:[0-9]+\]]], v{{[0-9]+}}, 0
+; SI: s_and_saveexec_b64 [[CMPTID]], [[CMPTID]]
+; SI: s_xor_b64 [[CMPTID]], exec, [[CMPTID]]
+
+; SI: buffer_load_dword [[LOAD:v[0-9]+]]
+; SI: v_cmp_ne_i32_e64 [[CMPLOAD:s\[[0-9]+:[0-9]+\]]], [[LOAD]], 0
+; SI: v_cndmask_b32_e64 {{v[0-9]+}}, 0, -1, [[CMPLOAD]]
+
+
+; SI: BB6_2:
+; SI: s_or_b64 exec, exec, [[CMPTID]]
+; SI: v_cmp_ne_i32_e32 vcc, 0, v0
+; SI: v_div_fmas_f32 {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}
+; SI: buffer_store_dword
+; SI: s_endpgm
+define void @test_div_fmas_f32_i1_phi_vcc(float addrspace(1)* %out, float addrspace(1)* %in, i32 addrspace(1)* %dummy) nounwind {
+entry:
+  %tid = call i32 @llvm.r600.read.tidig.x() nounwind readnone
+  %gep.out = getelementptr float addrspace(1)* %out, i32 2
+  %gep.a = getelementptr float addrspace(1)* %in, i32 %tid
+  %gep.b = getelementptr float addrspace(1)* %gep.a, i32 1
+  %gep.c = getelementptr float addrspace(1)* %gep.a, i32 2
+
+  %a = load float addrspace(1)* %gep.a
+  %b = load float addrspace(1)* %gep.b
+  %c = load float addrspace(1)* %gep.c
+
+  %cmp0 = icmp eq i32 %tid, 0
+  br i1 %cmp0, label %bb, label %exit
+
+bb:
+  %val = load i32 addrspace(1)* %dummy
+  %cmp1 = icmp ne i32 %val, 0
+  br label %exit
+
+exit:
+  %cond = phi i1 [false, %entry], [%cmp1, %bb]
+  %result = call float @llvm.AMDGPU.div.fmas.f32(float %a, float %b, float %c, i1 %cond) nounwind readnone
+  store float %result, float addrspace(1)* %gep.out, align 4
+  ret void
+}