SelectionDAG: Select min/max when both are used

Allow two users of the condition if the other user
is also a min/max select. i.e.

%c = icmp slt i32 %x, %y
%min = select i1 %c, i32 %x, i32 %y
%max = select i1 %c, i32 %y, i32 %x

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@269699 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp b/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
index 58751b7..aa39c78 100644 (file)
--- a/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
+++ b/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
@@ -2697,6 +2697,14 @@ void SelectionDAGBuilder::visitFCmp(const User &I) {
   setValue(&I, DAG.getSetCC(getCurSDLoc(), DestVT, Op1, Op2, Condition));
 }
 
+// Check if the condition of the select has one use or two users that are both
+// selects with the same condition.
+bool hasOnlySelectUsers(const Value *Cond) {
+  return std::all_of(Cond->user_begin(), Cond->user_end(), [](const Value *V) {
+    return isa<SelectInst>(V);
+  });
+}
+
 void SelectionDAGBuilder::visitSelect(const User &I) {
   SmallVector<EVT, 4> ValueVTs;
   ComputeValueVTs(DAG.getTargetLoweringInfo(), DAG.getDataLayout(), I.getType(),
@@ -2782,7 +2790,7 @@ void SelectionDAGBuilder::visitSelect(const User &I) {
         // If the underlying comparison instruction is used by any other
         // instruction, the consumed instructions won't be destroyed, so it is
         // not profitable to convert to a min/max.
-        cast<SelectInst>(&I)->getCondition()->hasOneUse()) {
+        hasOnlySelectUsers(cast<SelectInst>(I).getCondition())) {
       OpCode = Opc;
       LHSVal = getValue(LHS);
       RHSVal = getValue(RHS);

diff --git a/test/CodeGen/AMDGPU/select-i1.ll b/test/CodeGen/AMDGPU/select-i1.ll
index 6735394..2406831 100644 (file)
--- a/test/CodeGen/AMDGPU/select-i1.ll
+++ b/test/CodeGen/AMDGPU/select-i1.ll
@@ -1,4 +1,4 @@
-; RUN: llc -march=amdgcn -mcpu=SI -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
+; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
 ; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
 
 ; FIXME: This should go in existing select.ll test, except the current testcase there is broken on SI
@@ -13,3 +13,15 @@ define void @select_i1(i1 addrspace(1)* %out, i32 %cond, i1 %a, i1 %b) nounwind
   ret void
 }
 
+; FUNC-LABEL: {{^}}s_minmax_i1:
+; SI-DAG: buffer_load_ubyte [[COND:v[0-9]+]], off, s{{\[[0-9]+:[0-9]+\]}}, 0 offset:44
+; SI-DAG: buffer_load_ubyte [[A:v[0-9]+]], off, s{{\[[0-9]+:[0-9]+\]}}, 0 offset:45
+; SI-DAG: buffer_load_ubyte [[B:v[0-9]+]], off, s{{\[[0-9]+:[0-9]+\]}}, 0 offset:46
+; SI: v_cmp_eq_i32_e32 vcc, 1, [[COND]]
+; SI: v_cndmask_b32_e32 v{{[0-9]+}}, [[B]], [[A]]
+define void @s_minmax_i1(i1 addrspace(1)* %out, i1 zeroext %cond, i1 zeroext %a, i1 zeroext %b) nounwind {
+  %cmp = icmp slt i1 %cond, false
+  %sel = select i1 %cmp, i1 %a, i1 %b
+  store i1 %sel, i1 addrspace(1)* %out, align 4
+  ret void
+}

diff --git a/test/CodeGen/AMDGPU/sminmax.ll b/test/CodeGen/AMDGPU/sminmax.ll
index bea3ac1..20f96fc 100644 (file)
--- a/test/CodeGen/AMDGPU/sminmax.ll
+++ b/test/CodeGen/AMDGPU/sminmax.ll
@@ -128,3 +128,76 @@ define void @v_abs_v4i32(<4 x i32> addrspace(1)* %out, <4 x i32> addrspace(1)* %
   store <4 x i32> %res2, <4 x i32> addrspace(1)* %out, align 4
   ret void
 }
+
+; FUNC-LABEL: {{^}}s_min_max_i32:
+; GCN: s_load_dword [[VAL0:s[0-9]+]]
+; GCN: s_load_dword [[VAL1:s[0-9]+]]
+
+; GCN-DAG: s_min_i32 s{{[0-9]+}}, [[VAL0]], [[VAL1]]
+; GCN-DAG: s_max_i32 s{{[0-9]+}}, [[VAL0]], [[VAL1]]
+define void @s_min_max_i32(i32 addrspace(1)* %out0, i32 addrspace(1)* %out1, i32 %val0, i32 %val1) nounwind {
+  %cond0 = icmp sgt i32 %val0, %val1
+  %sel0 = select i1 %cond0, i32 %val0, i32 %val1
+  %sel1 = select i1 %cond0, i32 %val1, i32 %val0
+
+  store volatile i32 %sel0, i32 addrspace(1)* %out0, align 4
+  store volatile i32 %sel1, i32 addrspace(1)* %out1, align 4
+  ret void
+}
+
+; FUNC-LABEL: {{^}}v_min_max_i32:
+; GCN: buffer_load_dword [[VAL0:v[0-9]+]]
+; GCN: buffer_load_dword [[VAL1:v[0-9]+]]
+
+; GCN-DAG: v_min_i32_e32 v{{[0-9]+}}, [[VAL1]], [[VAL0]]
+; GCN-DAG: v_max_i32_e32 v{{[0-9]+}}, [[VAL1]], [[VAL0]]
+define void @v_min_max_i32(i32 addrspace(1)* %out0, i32 addrspace(1)* %out1, i32 addrspace(1)* %ptr0, i32 addrspace(1)* %ptr1) nounwind {
+  %val0 = load volatile i32, i32 addrspace(1)* %ptr0
+  %val1 = load volatile i32, i32 addrspace(1)* %ptr1
+
+  %cond0 = icmp sgt i32 %val0, %val1
+  %sel0 = select i1 %cond0, i32 %val0, i32 %val1
+  %sel1 = select i1 %cond0, i32 %val1, i32 %val0
+
+  store volatile i32 %sel0, i32 addrspace(1)* %out0, align 4
+  store volatile i32 %sel1, i32 addrspace(1)* %out1, align 4
+  ret void
+}
+
+; FUNC-LABEL: {{^}}s_min_max_v4i32:
+; GCN-DAG: s_min_i32
+; GCN-DAG: s_min_i32
+; GCN-DAG: s_min_i32
+; GCN-DAG: s_min_i32
+; GCN-DAG: s_max_i32
+; GCN-DAG: s_max_i32
+; GCN-DAG: s_max_i32
+; GCN-DAG: s_max_i32
+define void @s_min_max_v4i32(<4 x i32> addrspace(1)* %out0, <4 x i32> addrspace(1)* %out1, <4 x i32> %val0, <4 x i32> %val1) nounwind {
+  %cond0 = icmp sgt <4 x i32> %val0, %val1
+  %sel0 = select <4 x i1> %cond0, <4 x i32> %val0, <4 x i32> %val1
+  %sel1 = select <4 x i1> %cond0, <4 x i32> %val1, <4 x i32> %val0
+
+  store volatile <4 x i32> %sel0, <4 x i32> addrspace(1)* %out0, align 4
+  store volatile <4 x i32> %sel1, <4 x i32> addrspace(1)* %out1, align 4
+  ret void
+}
+
+; FUNC-LABEL: {{^}}v_min_max_i32_user:
+; GCN: v_cmp_gt_i32_e32
+; GCN-DAG: v_cndmask_b32_e32
+; GCN-DAG: v_cndmask_b32_e32
+; GCN-DAG: v_cndmask_b32_e64 v{{[0-9]+}}, 0, 1, vcc
+define void @v_min_max_i32_user(i32 addrspace(1)* %out0, i32 addrspace(1)* %out1, i32 addrspace(1)* %ptr0, i32 addrspace(1)* %ptr1) nounwind {
+  %val0 = load volatile i32, i32 addrspace(1)* %ptr0
+  %val1 = load volatile i32, i32 addrspace(1)* %ptr1
+
+  %cond0 = icmp sgt i32 %val0, %val1
+  %sel0 = select i1 %cond0, i32 %val0, i32 %val1
+  %sel1 = select i1 %cond0, i32 %val1, i32 %val0
+
+  store volatile i32 %sel0, i32 addrspace(1)* %out0, align 4
+  store volatile i32 %sel1, i32 addrspace(1)* %out1, align 4
+  store volatile i1 %cond0, i1 addrspace(1)* undef
+  ret void
+}