[AMDGPU] setcc (select cc, CT, CF), CF, eq | ne -> xor cc, -1 | cc

[rampitec]

Summary

This is the common case in the BE when we serialize condition and then rematerialize it. Use either original or inverted condition.

[msearles]

LGTM

[arsenm]

llvm/trunk/lib/Target/AMDGPU/SIISelLowering.cpp

7493 ↗

(On Diff #151603)

Don't these combines exist in DAGCombiner already?

[rampitec]

llvm/trunk/lib/Target/AMDGPU/SIISelLowering.cpp

7493 ↗

(On Diff #151603)

sext part? I did not check, I have just indented existing code here. Although I doubt. The target specific part here is call to isBoolSGPR for both sext and select part.

lib/Target/AMDGPU/SIISelLowering.cpp

SDValue SITargetLowering::performSetCCCombine(SDNode *N,
   if (!CRHS) {
     CRHS = dyn_cast<ConstantSDNode>(LHS);
     if (CRHS) {
       std::swap(LHS, RHS);
       CC = getSetCCSwappedOperands(CC);
     }
   }
 
-  if (CRHS && VT == MVT::i32 && LHS.getOpcode() == ISD::SIGN_EXTEND &&
+  if (CRHS) {
+    if (VT == MVT::i32 && LHS.getOpcode() == ISD::SIGN_EXTEND &&
         isBoolSGPR(LHS.getOperand(0))) {
       // setcc (sext from i1 cc), -1, ne|sgt|ult) => not cc => xor cc, -1
       // setcc (sext from i1 cc), -1, eq|sle|uge) => cc
       // setcc (sext from i1 cc),  0, eq|sge|ule) => not cc => xor cc, -1
       // setcc (sext from i1 cc),  0, ne|ugt|slt) => cc
       if ((CRHS->isAllOnesValue() &&
            (CC == ISD::SETNE || CC == ISD::SETGT || CC == ISD::SETULT)) ||
           (CRHS->isNullValue() &&
            (CC == ISD::SETEQ || CC == ISD::SETGE || CC == ISD::SETULE)))
         return DAG.getNode(ISD::XOR, SL, MVT::i1, LHS.getOperand(0),
                            DAG.getConstant(-1, SL, MVT::i1));
       if ((CRHS->isAllOnesValue() &&
            (CC == ISD::SETEQ || CC == ISD::SETLE || CC == ISD::SETUGE)) ||
           (CRHS->isNullValue() &&
            (CC == ISD::SETNE || CC == ISD::SETUGT || CC == ISD::SETLT)))
         return LHS.getOperand(0);
     }
 
+    uint64_t CRHSVal = CRHS->getZExtValue();
+    if ((CC == ISD::SETEQ || CC == ISD::SETNE) &&
+        LHS.getOpcode() == ISD::SELECT &&
+        isa<ConstantSDNode>(LHS.getOperand(1)) &&
+        isa<ConstantSDNode>(LHS.getOperand(2)) &&
+        LHS.getConstantOperandVal(1) != LHS.getConstantOperandVal(2) &&
+        isBoolSGPR(LHS.getOperand(0))) {
+      // Given CT != FT:
+      // setcc (select cc, CT, CF), CF, eq => xor cc, -1
+      // setcc (select cc, CT, CF), CF, ne => cc
+      // setcc (select cc, CT, CF), CT, ne => xor cc, -1
+      // setcc (select cc, CT, CF), CT, eq => cc
+      uint64_t CT = LHS.getConstantOperandVal(1);
+      uint64_t CF = LHS.getConstantOperandVal(2);
+
+      if ((CF == CRHSVal && CC == ISD::SETEQ) ||
+          (CT == CRHSVal && CC == ISD::SETNE))
+        return DAG.getNode(ISD::XOR, SL, MVT::i1, LHS.getOperand(0),
+                           DAG.getConstant(-1, SL, MVT::i1));
+      if ((CF == CRHSVal && CC == ISD::SETNE) ||
+          (CT == CRHSVal && CC == ISD::SETEQ))
+        return LHS.getOperand(0);
+    }
+  }
+
   if (VT != MVT::f32 && VT != MVT::f64 && (Subtarget->has16BitInsts() &&
                                            VT != MVT::f16))
     return SDValue();
 
   // Match isinf pattern
   // (fcmp oeq (fabs x), inf) -> (fp_class x, (p_infinity | n_infinity))
   if (CC == ISD::SETOEQ && LHS.getOpcode() == ISD::FABS) {
     const ConstantFPSDNode *CRHS = dyn_cast<ConstantFPSDNode>(RHS);

test/CodeGen/AMDGPU/dagcombine-setcc-select.ll

+; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs -O0 < %s | FileCheck -enable-var-scope -check-prefix=GCN %s
+
+; GCN-LABEL: {{^}}eq_t:
+; GCN-DAG: s_load_dword [[X:s[0-9]+]]
+; GCN-DAG: v_mov_b32_e32 [[ONE:v[0-9]+]], 1.0
+; GCN:     v_cmp_lt_f32_e{{32|64}} [[CC:s\[[0-9]+:[0-9]+\]|vcc]], [[X]], [[ONE]]{{$}}
+; GCN-NOT: 0xddd5
+; GCN-NOT: v_cndmask_b32
+; GCN-NOT: v_cmp_eq_u32
+; GCN-NOT: v_cndmask_b32
+; GCN-DAG: v_mov_b32_e32 [[TWO:v[0-9]+]], 2.0
+; GCN-DAG: v_mov_b32_e32 [[FOUR:v[0-9]+]], 4.0
+; GCN:     v_cndmask_b32_e{{32|64}} [[RES:v[0-9]+]], [[TWO]], [[FOUR]], [[CC]]
+; GCN:     store_dword v[{{[0-9:]+}}], [[RES]]{{$}}
+define amdgpu_kernel void @eq_t(float %x) {
+  %c1 = fcmp olt float %x, 1.0
+  %s1 = select i1 %c1, i32 56789, i32 1
+  %c2 = icmp eq i32 %s1, 56789
+  %s2 = select i1 %c2, float 4.0, float 2.0
+  store float %s2, float* undef, align 4
+  ret void
+}
+
+; GCN-LABEL: {{^}}ne_t:
+; GCN-DAG: s_load_dword [[X:s[0-9]+]]
+; GCN-DAG: v_mov_b32_e32 [[ONE:v[0-9]+]], 1.0
+; GCN:     v_cmp_lt_f32_e{{32|64}} [[CC:s\[[0-9]+:[0-9]+\]|vcc]], [[X]], [[ONE]]{{$}}
+; GCN-NOT: 0xddd5
+; GCN-NOT: v_cndmask_b32
+; GCN-NOT: v_cmp_eq_u32
+; GCN-NOT: v_cndmask_b32
+; GCN-DAG: v_mov_b32_e32 [[TWO:v[0-9]+]], 2.0
+; GCN-DAG: v_mov_b32_e32 [[FOUR:v[0-9]+]], 4.0
+; GCN:     v_cndmask_b32_e{{32|64}} [[RES:v[0-9]+]], [[FOUR]], [[TWO]], [[CC]]
+; GCN:     store_dword v[{{[0-9:]+}}], [[RES]]{{$}}
+define amdgpu_kernel void @ne_t(float %x) {
+  %c1 = fcmp olt float %x, 1.0
+  %s1 = select i1 %c1, i32 56789, i32 1
+  %c2 = icmp ne i32 %s1, 56789
+  %s2 = select i1 %c2, float 4.0, float 2.0
+  store float %s2, float* undef, align 4
+  ret void
+}
+
+; GCN-LABEL: {{^}}eq_f:
+; GCN-DAG: s_load_dword [[X:s[0-9]+]]
+; GCN-DAG: v_mov_b32_e32 [[ONE:v[0-9]+]], 1.0
+; GCN:     v_cmp_lt_f32_e{{32|64}} [[CC:s\[[0-9]+:[0-9]+\]|vcc]], [[X]], [[ONE]]{{$}}
+; GCN-NOT: 0xddd5
+; GCN-NOT: v_cndmask_b32
+; GCN-NOT: v_cmp_eq_u32
+; GCN-NOT: v_cndmask_b32
+; GCN-DAG: v_mov_b32_e32 [[TWO:v[0-9]+]], 2.0
+; GCN-DAG: v_mov_b32_e32 [[FOUR:v[0-9]+]], 4.0
+; GCN:     v_cndmask_b32_e{{32|64}} [[RES:v[0-9]+]], [[FOUR]], [[TWO]], [[CC]]
+; GCN:     store_dword v[{{[0-9:]+}}], [[RES]]{{$}}
+define amdgpu_kernel void @eq_f(float %x) {
+  %c1 = fcmp olt float %x, 1.0
+  %s1 = select i1 %c1, i32 1, i32 56789
+  %c2 = icmp eq i32 %s1, 56789
+  %s2 = select i1 %c2, float 4.0, float 2.0
+  store float %s2, float* undef, align 4
+  ret void
+}
+
+; GCN-LABEL: {{^}}ne_f:
+; GCN-DAG: s_load_dword [[X:s[0-9]+]]
+; GCN-DAG: v_mov_b32_e32 [[ONE:v[0-9]+]], 1.0
+; GCN:     v_cmp_lt_f32_e{{32|64}} [[CC:s\[[0-9]+:[0-9]+\]|vcc]], [[X]], [[ONE]]{{$}}
+; GCN-NOT: 0xddd5
+; GCN-NOT: v_cndmask_b32
+; GCN-NOT: v_cmp_eq_u32
+; GCN-NOT: v_cndmask_b32
+; GCN-DAG: v_mov_b32_e32 [[TWO:v[0-9]+]], 2.0
+; GCN-DAG: v_mov_b32_e32 [[FOUR:v[0-9]+]], 4.0
+; GCN:     v_cndmask_b32_e{{32|64}} [[RES:v[0-9]+]], [[TWO]], [[FOUR]], [[CC]]
+; GCN:     store_dword v[{{[0-9:]+}}], [[RES]]{{$}}
+define amdgpu_kernel void @ne_f(float %x) {
+  %c1 = fcmp olt float %x, 1.0
+  %s1 = select i1 %c1, i32 1, i32 56789
+  %c2 = icmp ne i32 %s1, 56789
+  %s2 = select i1 %c2, float 4.0, float 2.0
+  store float %s2, float* undef, align 4
+  ret void
+}
+
+; GCN-LABEL: {{^}}different_constants:
+; GCN-DAG: s_load_dword [[X:s[0-9]+]]
+; GCN-DAG: v_mov_b32_e32 [[ONE:v[0-9]+]], 1.0
+; GCN-DAG: v_cmp_lt_f32_e{{32|64}} [[CC1:s\[[0-9]+:[0-9]+\]|vcc]], [[X]], [[ONE]]{{$}}
+; GCN-DAG: v_cndmask_b32_e{{32|64}} [[CND1:v[0-9]+]], v{{[0-9]+}}, v{{[0-9]+}}, [[CC1]]
+; GCN-DAG: v_cmp_eq_u32_e{{32|64}} [[CC2:s\[[0-9]+:[0-9]+\]|vcc]], v{{[0-9]+}}, v{{[0-9]+}}{{$}}
+; GCN-DAG: v_mov_b32_e32 [[TWO:v[0-9]+]], 2.0
+; GCN-DAG: v_mov_b32_e32 [[FOUR:v[0-9]+]], 4.0
+; GCN:     v_cndmask_b32_e{{32|64}} [[RES:v[0-9]+]], [[TWO]], [[FOUR]], [[CC2]]
+; GCN:     store_dword v[{{[0-9:]+}}], [[RES]]{{$}}
+define amdgpu_kernel void @different_constants(float %x) {
+  %c1 = fcmp olt float %x, 1.0
+  %s1 = select i1 %c1, i32 56789, i32 1
+  %c2 = icmp eq i32 %s1, 5678
+  %s2 = select i1 %c2, float 4.0, float 2.0
+  store float %s2, float* undef, align 4
+  ret void
+}