[AMDGPU] Add BFX Formation Combines to RegBankCombiner

llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp

@@ -4636,10 +4636,17 @@ bool CombinerHelper::matchBitfieldExtractFromSExtInReg(
   if (ShiftImm < 0 || ShiftImm + Width > Ty.getScalarSizeInBits())
     return false;
 
+  const RegisterBank *RB = getRegBank(ShiftSrc);
+
   MatchInfo = [=](MachineIRBuilder &B) {
     auto Cst1 = B.buildConstant(ExtractTy, ShiftImm);
     auto Cst2 = B.buildConstant(ExtractTy, Width);
     B.buildSbfx(Dst, ShiftSrc, Cst1, Cst2);
+
+    if (RB) {
+      MRI.setRegBank(Cst1.getReg(0), *RB);
+      MRI.setRegBank(Cst2.getReg(0), *RB);
+    }
   };
   return true;
 }
@@ -4674,10 +4681,18 @@ bool CombinerHelper::matchBitfieldExtractFromAnd(MachineInstr &MI,
     return false;
 
   uint64_t Width = APInt(Size, AndImm).countr_one();
+
+  const RegisterBank *RB = getRegBank(ShiftSrc);
+
   MatchInfo = [=](MachineIRBuilder &B) {
     auto WidthCst = B.buildConstant(ExtractTy, Width);
     auto LSBCst = B.buildConstant(ExtractTy, LSBImm);
     B.buildInstr(TargetOpcode::G_UBFX, {Dst}, {ShiftSrc, LSBCst, WidthCst});
+
+    if (RB) {
+      MRI.setRegBank(WidthCst.getReg(0), *RB);
+      MRI.setRegBank(LSBCst.getReg(0), *RB);
+    }
   };
   return true;
 }
@@ -4724,10 +4739,17 @@ bool CombinerHelper::matchBitfieldExtractFromShr(
   const int64_t Pos = ShrAmt - ShlAmt;
   const int64_t Width = Size - ShrAmt;
 
+  const RegisterBank *RB = getRegBank(ShlSrc);
+
   MatchInfo = [=](MachineIRBuilder &B) {
     auto WidthCst = B.buildConstant(ExtractTy, Width);
     auto PosCst = B.buildConstant(ExtractTy, Pos);
     B.buildInstr(ExtrOpcode, {Dst}, {ShlSrc, PosCst, WidthCst});
+
+    if (RB) {
+      MRI.setRegBank(WidthCst.getReg(0), *RB);
+      MRI.setRegBank(PosCst.getReg(0), *RB);
+    }
   };
   return true;
 }
@@ -4782,10 +4804,17 @@ bool CombinerHelper::matchBitfieldExtractFromShrAnd(
   if (Opcode == TargetOpcode::G_ASHR && Width + ShrAmt == Size)
     return false;
 
+  const RegisterBank *RB = getRegBank(AndSrc);
+
   MatchInfo = [=](MachineIRBuilder &B) {
     auto WidthCst = B.buildConstant(ExtractTy, Width);
     auto PosCst = B.buildConstant(ExtractTy, Pos);
     B.buildInstr(TargetOpcode::G_UBFX, {Dst}, {AndSrc, PosCst, WidthCst});
+
+    if (RB) {
+      MRI.setRegBank(WidthCst.getReg(0), *RB);
+      MRI.setRegBank(PosCst.getReg(0), *RB);
+    }
   };
   return true;
 }

llvm/lib/Target/AMDGPU/AMDGPUCombine.td

@@ -210,5 +210,5 @@ def AMDGPURegBankCombiner : GICombiner<
    fp_minmax_to_clamp, fp_minmax_to_med3, fmed3_intrinsic_to_clamp,
    identity_combines, redundant_and, constant_fold_cast_op,
    cast_of_cast_combines, sext_trunc, zext_of_shift_amount_combines,
-   lower_uniform_sbfx, lower_uniform_ubfx]> {
+   lower_uniform_sbfx, lower_uniform_ubfx, form_bitfield_extract]> {
 }

llvm/test/CodeGen/AMDGPU/GlobalISel/ashr.ll

@@ -811,16 +811,15 @@ define amdgpu_ps i32 @s_ashr_v2i16(<2 x i16> inreg %value, <2 x i16> inreg %amou
 ;
 ; GFX8-LABEL: s_ashr_v2i16:
 ; GFX8:       ; %bb.0:
-; GFX8-NEXT:    s_lshr_b32 s2, s0, 16
-; GFX8-NEXT:    s_sext_i32_i16 s0, s0
-; GFX8-NEXT:    s_lshr_b32 s3, s1, 16
-; GFX8-NEXT:    s_ashr_i32 s0, s0, s1
-; GFX8-NEXT:    s_sext_i32_i16 s1, s2
-; GFX8-NEXT:    s_ashr_i32 s1, s1, s3
-; GFX8-NEXT:    s_and_b32 s1, 0xffff, s1
+; GFX8-NEXT:    s_lshr_b32 s2, s1, 16
+; GFX8-NEXT:    s_sext_i32_i16 s3, s0
+; GFX8-NEXT:    s_bfe_i32 s0, s0, 0x100010
+; GFX8-NEXT:    s_ashr_i32 s0, s0, s2
+; GFX8-NEXT:    s_ashr_i32 s1, s3, s1
 ; GFX8-NEXT:    s_and_b32 s0, 0xffff, s0
-; GFX8-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX8-NEXT:    s_or_b32 s0, s0, s1
+; GFX8-NEXT:    s_and_b32 s1, 0xffff, s1
+; GFX8-NEXT:    s_lshl_b32 s0, s0, 16
+; GFX8-NEXT:    s_or_b32 s0, s1, s0
 ; GFX8-NEXT:    ; return to shader part epilog
 ;
 ; GFX9-LABEL: s_ashr_v2i16:
@@ -1014,26 +1013,24 @@ define amdgpu_ps <2 x i32> @s_ashr_v4i16(<4 x i16> inreg %value, <4 x i16> inreg
 ;
 ; GFX8-LABEL: s_ashr_v4i16:
 ; GFX8:       ; %bb.0:
-; GFX8-NEXT:    s_lshr_b32 s4, s0, 16
-; GFX8-NEXT:    s_sext_i32_i16 s0, s0
-; GFX8-NEXT:    s_lshr_b32 s6, s2, 16
-; GFX8-NEXT:    s_ashr_i32 s0, s0, s2
-; GFX8-NEXT:    s_sext_i32_i16 s2, s4
-; GFX8-NEXT:    s_lshr_b32 s5, s1, 16
-; GFX8-NEXT:    s_ashr_i32 s2, s2, s6
-; GFX8-NEXT:    s_sext_i32_i16 s1, s1
-; GFX8-NEXT:    s_lshr_b32 s7, s3, 16
-; GFX8-NEXT:    s_ashr_i32 s1, s1, s3
-; GFX8-NEXT:    s_sext_i32_i16 s3, s5
-; GFX8-NEXT:    s_and_b32 s2, 0xffff, s2
-; GFX8-NEXT:    s_ashr_i32 s3, s3, s7
+; GFX8-NEXT:    s_lshr_b32 s4, s2, 16
+; GFX8-NEXT:    s_sext_i32_i16 s6, s0
+; GFX8-NEXT:    s_bfe_i32 s0, s0, 0x100010
+; GFX8-NEXT:    s_lshr_b32 s5, s3, 16
+; GFX8-NEXT:    s_ashr_i32 s0, s0, s4
+; GFX8-NEXT:    s_sext_i32_i16 s4, s1
+; GFX8-NEXT:    s_bfe_i32 s1, s1, 0x100010
+; GFX8-NEXT:    s_ashr_i32 s2, s6, s2
+; GFX8-NEXT:    s_ashr_i32 s1, s1, s5
 ; GFX8-NEXT:    s_and_b32 s0, 0xffff, s0
-; GFX8-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX8-NEXT:    s_or_b32 s0, s0, s2
-; GFX8-NEXT:    s_and_b32 s2, 0xffff, s3
+; GFX8-NEXT:    s_ashr_i32 s3, s4, s3
+; GFX8-NEXT:    s_and_b32 s2, 0xffff, s2
+; GFX8-NEXT:    s_lshl_b32 s0, s0, 16
 ; GFX8-NEXT:    s_and_b32 s1, 0xffff, s1
-; GFX8-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX8-NEXT:    s_or_b32 s1, s1, s2
+; GFX8-NEXT:    s_or_b32 s0, s2, s0
+; GFX8-NEXT:    s_and_b32 s2, 0xffff, s3
+; GFX8-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX8-NEXT:    s_or_b32 s1, s2, s1
 ; GFX8-NEXT:    ; return to shader part epilog
 ;
 ; GFX9-LABEL: s_ashr_v4i16:
@@ -1223,46 +1220,42 @@ define amdgpu_ps <4 x i32> @s_ashr_v8i16(<8 x i16> inreg %value, <8 x i16> inreg
 ;
 ; GFX8-LABEL: s_ashr_v8i16:
 ; GFX8:       ; %bb.0:
-; GFX8-NEXT:    s_lshr_b32 s8, s0, 16
-; GFX8-NEXT:    s_sext_i32_i16 s0, s0
-; GFX8-NEXT:    s_lshr_b32 s12, s4, 16
-; GFX8-NEXT:    s_ashr_i32 s0, s0, s4
-; GFX8-NEXT:    s_sext_i32_i16 s4, s8
-; GFX8-NEXT:    s_lshr_b32 s9, s1, 16
-; GFX8-NEXT:    s_ashr_i32 s4, s4, s12
-; GFX8-NEXT:    s_sext_i32_i16 s1, s1
-; GFX8-NEXT:    s_lshr_b32 s13, s5, 16
-; GFX8-NEXT:    s_ashr_i32 s1, s1, s5
-; GFX8-NEXT:    s_sext_i32_i16 s5, s9
-; GFX8-NEXT:    s_and_b32 s4, 0xffff, s4
-; GFX8-NEXT:    s_lshr_b32 s10, s2, 16
-; GFX8-NEXT:    s_ashr_i32 s5, s5, s13
-; GFX8-NEXT:    s_sext_i32_i16 s2, s2
+; GFX8-NEXT:    s_lshr_b32 s8, s4, 16
+; GFX8-NEXT:    s_sext_i32_i16 s12, s0
+; GFX8-NEXT:    s_bfe_i32 s0, s0, 0x100010
+; GFX8-NEXT:    s_lshr_b32 s9, s5, 16
+; GFX8-NEXT:    s_ashr_i32 s0, s0, s8
+; GFX8-NEXT:    s_sext_i32_i16 s8, s1
+; GFX8-NEXT:    s_bfe_i32 s1, s1, 0x100010
+; GFX8-NEXT:    s_lshr_b32 s10, s6, 16
+; GFX8-NEXT:    s_ashr_i32 s4, s12, s4
+; GFX8-NEXT:    s_ashr_i32 s5, s8, s5
+; GFX8-NEXT:    s_ashr_i32 s1, s1, s9
+; GFX8-NEXT:    s_sext_i32_i16 s8, s2
+; GFX8-NEXT:    s_bfe_i32 s2, s2, 0x100010
 ; GFX8-NEXT:    s_and_b32 s0, 0xffff, s0
-; GFX8-NEXT:    s_lshl_b32 s4, s4, 16
-; GFX8-NEXT:    s_lshr_b32 s14, s6, 16
-; GFX8-NEXT:    s_ashr_i32 s2, s2, s6
-; GFX8-NEXT:    s_sext_i32_i16 s6, s10
-; GFX8-NEXT:    s_or_b32 s0, s0, s4
-; GFX8-NEXT:    s_and_b32 s4, 0xffff, s5
-; GFX8-NEXT:    s_lshr_b32 s11, s3, 16
-; GFX8-NEXT:    s_ashr_i32 s6, s6, s14
-; GFX8-NEXT:    s_sext_i32_i16 s3, s3
+; GFX8-NEXT:    s_lshr_b32 s11, s7, 16
+; GFX8-NEXT:    s_ashr_i32 s6, s8, s6
+; GFX8-NEXT:    s_ashr_i32 s2, s2, s10
+; GFX8-NEXT:    s_sext_i32_i16 s8, s3
+; GFX8-NEXT:    s_bfe_i32 s3, s3, 0x100010
+; GFX8-NEXT:    s_and_b32 s4, 0xffff, s4
+; GFX8-NEXT:    s_lshl_b32 s0, s0, 16
 ; GFX8-NEXT:    s_and_b32 s1, 0xffff, s1
-; GFX8-NEXT:    s_lshl_b32 s4, s4, 16
-; GFX8-NEXT:    s_lshr_b32 s15, s7, 16
-; GFX8-NEXT:    s_ashr_i32 s3, s3, s7
-; GFX8-NEXT:    s_sext_i32_i16 s7, s11
-; GFX8-NEXT:    s_or_b32 s1, s1, s4
-; GFX8-NEXT:    s_and_b32 s4, 0xffff, s6
-; GFX8-NEXT:    s_ashr_i32 s7, s7, s15
+; GFX8-NEXT:    s_ashr_i32 s3, s3, s11
+; GFX8-NEXT:    s_or_b32 s0, s4, s0
+; GFX8-NEXT:    s_and_b32 s4, 0xffff, s5
+; GFX8-NEXT:    s_lshl_b32 s1, s1, 16
 ; GFX8-NEXT:    s_and_b32 s2, 0xffff, s2
-; GFX8-NEXT:    s_lshl_b32 s4, s4, 16
-; GFX8-NEXT:    s_or_b32 s2, s2, s4
-; GFX8-NEXT:    s_and_b32 s4, 0xffff, s7
+; GFX8-NEXT:    s_ashr_i32 s7, s8, s7
+; GFX8-NEXT:    s_or_b32 s1, s4, s1
+; GFX8-NEXT:    s_and_b32 s4, 0xffff, s6
+; GFX8-NEXT:    s_lshl_b32 s2, s2, 16
 ; GFX8-NEXT:    s_and_b32 s3, 0xffff, s3
-; GFX8-NEXT:    s_lshl_b32 s4, s4, 16
-; GFX8-NEXT:    s_or_b32 s3, s3, s4
+; GFX8-NEXT:    s_or_b32 s2, s4, s2
+; GFX8-NEXT:    s_and_b32 s4, 0xffff, s7
+; GFX8-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX8-NEXT:    s_or_b32 s3, s4, s3
 ; GFX8-NEXT:    ; return to shader part epilog
 ;
 ; GFX9-LABEL: s_ashr_v8i16: