[AMDGPU] Recognise bitmask operations as srcmods on integer types (#149110)

Add to the VOP patterns to recognise when or/xor/and are masking only
the most significant bit of i32/v2i32/i64 and replace with the appropriate source modifier.

Author: chrisjbris

llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp

@@ -3059,6 +3059,38 @@ bool AMDGPUDAGToDAGISel::SelectVOP3ModsImpl(SDValue In, SDValue &Src,
     Src = Src.getOperand(0);
   }
 
+  // Convert various sign-bit masks to src mods. Currently disabled for 16-bit
+  // types as the codegen replaces the operand without adding a srcmod.
+  // This is intentionally finding the cases where we are performing float neg
+  // and abs on int types, the goal is not to obtain two's complement neg or
+  // abs.
+  // TODO: Add 16-bit support.
+  unsigned Opc = Src->getOpcode();
+  EVT VT = Src.getValueType();
+  if ((Opc != ISD::AND && Opc != ISD::OR && Opc != ISD::XOR) ||
+      (VT != MVT::i32 && VT != MVT::v2i32 && VT != MVT::i64))
+    return true;
+
+  ConstantSDNode *CRHS = dyn_cast<ConstantSDNode>(Src->getOperand(1));
+  if (!CRHS)
+    return true;
+
+  // Recognise (xor a, 0x80000000) as NEG SrcMod.
+  // Recognise (and a, 0x7fffffff) as ABS SrcMod.
+  // Recognise (or a, 0x80000000) as NEG+ABS SrcModifiers.
+  if (Opc == ISD::XOR && CRHS->getAPIntValue().isSignMask()) {
+    Mods |= SISrcMods::NEG;
+    Src = Src.getOperand(0);
+  } else if (Opc == ISD::AND && AllowAbs &&
+             CRHS->getAPIntValue().isMaxSignedValue()) {
+    Mods |= SISrcMods::ABS;
+    Src = Src.getOperand(0);
+  } else if (Opc == ISD::OR && AllowAbs && CRHS->getAPIntValue().isSignMask()) {
+    Mods |= SISrcMods::ABS;
+    Mods |= SISrcMods::NEG;
+    Src = Src.getOperand(0);
+  }
+
   return true;
 }
 

llvm/test/CodeGen/AMDGPU/fneg-modifier-casting.ll

@@ -349,29 +349,24 @@ define i32 @select_fneg_xor_select_i32(i1 %cond0, i1 %cond1, i32 %arg0, i32 %arg
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GCN-NEXT:    v_and_b32_e32 v0, 1, v0
-; GCN-NEXT:    v_xor_b32_e32 v2, 0x80000000, v2
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, 1, v0
 ; GCN-NEXT:    v_and_b32_e32 v1, 1, v1
-; GCN-NEXT:    v_cndmask_b32_e32 v0, v2, v3, vcc
-; GCN-NEXT:    v_xor_b32_e32 v2, 0x80000000, v0
+; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, 1, v0
+; GCN-NEXT:    v_cndmask_b32_e64 v0, -v2, v3, vcc
 ; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, 1, v1
-; GCN-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
+; GCN-NEXT:    v_cndmask_b32_e64 v0, v0, -v0, vcc
 ; GCN-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX11-LABEL: select_fneg_xor_select_i32:
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    v_and_b32_e32 v0, 1, v0
-; GFX11-NEXT:    v_xor_b32_e32 v2, 0x80000000, v2
 ; GFX11-NEXT:    v_and_b32_e32 v1, 1, v1
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
 ; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX11-NEXT:    v_cndmask_b32_e32 v0, v2, v3, vcc_lo
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_cndmask_b32_e64 v0, -v2, v3, vcc_lo
 ; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc_lo, 1, v1
-; GFX11-NEXT:    v_xor_b32_e32 v2, 0x80000000, v0
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT:    v_cndmask_b32_e64 v0, v0, -v0, vcc_lo
 ; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %fneg0 = xor i32 %arg0, -2147483648
   %select0 = select i1 %cond0, i32 %arg1, i32 %fneg0
@@ -550,31 +545,25 @@ define i64 @select_fneg_xor_select_i64(i1 %cond0, i1 %cond1, i64 %arg0, i64 %arg
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GCN-NEXT:    v_and_b32_e32 v0, 1, v0
-; GCN-NEXT:    v_xor_b32_e32 v3, 0x80000000, v3
-; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, 1, v0
 ; GCN-NEXT:    v_and_b32_e32 v1, 1, v1
+; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, 1, v0
 ; GCN-NEXT:    v_cndmask_b32_e32 v0, v2, v4, vcc
-; GCN-NEXT:    v_cndmask_b32_e32 v2, v3, v5, vcc
-; GCN-NEXT:    v_xor_b32_e32 v3, 0x80000000, v2
+; GCN-NEXT:    v_cndmask_b32_e64 v2, -v3, v5, vcc
 ; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, 1, v1
-; GCN-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
+; GCN-NEXT:    v_cndmask_b32_e64 v1, v2, -v2, vcc
 ; GCN-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX11-LABEL: select_fneg_xor_select_i64:
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    v_and_b32_e32 v0, 1, v0
-; GFX11-NEXT:    v_xor_b32_e32 v3, 0x80000000, v3
-; GFX11-NEXT:    v_and_b32_e32 v1, 1, v1
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX11-NEXT:    v_cndmask_b32_e32 v0, v2, v4, vcc_lo
-; GFX11-NEXT:    v_cndmask_b32_e32 v2, v3, v5, vcc_lo
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_dual_cndmask_b32 v0, v2, v4 :: v_dual_and_b32 v1, 1, v1
+; GFX11-NEXT:    v_cndmask_b32_e64 v2, -v3, v5, vcc_lo
 ; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc_lo, 1, v1
-; GFX11-NEXT:    v_xor_b32_e32 v3, 0x80000000, v2
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc_lo
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT:    v_cndmask_b32_e64 v1, v2, -v2, vcc_lo
 ; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %fneg0 = xor i64 %arg0, 9223372036854775808
   %select0 = select i1 %cond0, i64 %arg1, i64 %fneg0