[X86] Don't fold AND(SRL(X,Y),1) -> SETCC(BT(X,Y)) on BMI2 targets

With BMI2 we have SHRX which is a lot quicker than regular x86 shifts.

Fixes #55138

Author: RKSimon

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -47380,11 +47380,13 @@ static SDValue combineAnd(SDNode *N, SelectionDAG &DAG,
             Src.getOpcode() == ISD::TRUNCATE) &&
            Src.getOperand(0)->hasOneUse())
       Src = Src.getOperand(0);
+    bool ContainsNOT = false;
     X86::CondCode X86CC = X86::COND_B;
     // Peek through AND(NOT(SRL(X,Y)),1).
     if (isBitwiseNot(Src)) {
       Src = Src.getOperand(0);
       X86CC = X86::COND_AE;
+      ContainsNOT = true;
     }
     if (Src.getOpcode() == ISD::SRL &&
         !isa<ConstantSDNode>(Src.getOperand(1))) {
@@ -47394,9 +47396,12 @@ static SDValue combineAnd(SDNode *N, SelectionDAG &DAG,
       if (isBitwiseNot(Src)) {
         Src = Src.getOperand(0);
         X86CC = X86CC == X86::COND_AE ? X86::COND_B : X86::COND_AE;
+        ContainsNOT = true;
       }
-      if (SDValue BT = getBT(Src, BitNo, dl, DAG))
-        return DAG.getZExtOrTrunc(getSETCC(X86CC, BT, dl, DAG), dl, VT);
+      // If we have BMI2 then SHRX should be faster for i32/i64 cases.
+      if (!(Subtarget.hasBMI2() && !ContainsNOT && VT.getSizeInBits() >= 32))
+        if (SDValue BT = getBT(Src, BitNo, dl, DAG))
+          return DAG.getZExtOrTrunc(getSETCC(X86CC, BT, dl, DAG), dl, VT);
     }
   }
 

llvm/test/CodeGen/X86/setcc.ll

@@ -1,8 +1,8 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=i686-apple-darwin | FileCheck %s --check-prefixes=X86
-; RUN: llc < %s -mtriple=x86_64-apple-darwin | FileCheck %s --check-prefixes=X64,X64-NOTBM
-; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+tbm | FileCheck %s --check-prefixes=X64,X64-TBM
-; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+bmi2 | FileCheck %s --check-prefixes=X64,X64-NOTBM
+; RUN: llc < %s -mtriple=x86_64-apple-darwin | FileCheck %s --check-prefixes=X64,X64-NOTBM,X64-NOBMI2
+; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+tbm | FileCheck %s --check-prefixes=X64,X64-NOBMI2,X64-TBM
+; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+bmi2 | FileCheck %s --check-prefixes=X64,X64-NOTBM,X64-BMI2
 ; rdar://7329206
 
 define zeroext i16 @t1(i16 zeroext %x) nounwind readnone ssp {
@@ -191,12 +191,18 @@ define i64 @t9(i32 %0, i32 %1) {
 ; X86-NEXT:    xorl %edx, %edx
 ; X86-NEXT:    retl
 ;
-; X64-LABEL: t9:
-; X64:       ## %bb.0:
-; X64-NEXT:    xorl %eax, %eax
-; X64-NEXT:    btl %esi, %edi
-; X64-NEXT:    setb %al
-; X64-NEXT:    retq
+; X64-NOBMI2-LABEL: t9:
+; X64-NOBMI2:       ## %bb.0:
+; X64-NOBMI2-NEXT:    xorl %eax, %eax
+; X64-NOBMI2-NEXT:    btl %esi, %edi
+; X64-NOBMI2-NEXT:    setb %al
+; X64-NOBMI2-NEXT:    retq
+;
+; X64-BMI2-LABEL: t9:
+; X64-BMI2:       ## %bb.0:
+; X64-BMI2-NEXT:    shrxl %esi, %edi, %eax
+; X64-BMI2-NEXT:    andl $1, %eax
+; X64-BMI2-NEXT:    retq
   %3 = lshr i32 %0, %1
   %4 = and i32 %3, 1
   %5 = icmp ne i32 %4, 0
@@ -311,14 +317,22 @@ define i32 @PR55138(i32 %x) {
 ; X86-NEXT:    setb %al
 ; X86-NEXT:    retl
 ;
-; X64-LABEL: PR55138:
-; X64:       ## %bb.0:
-; X64-NEXT:    andl $15, %edi
-; X64-NEXT:    movl $27030, %ecx ## imm = 0x6996
-; X64-NEXT:    xorl %eax, %eax
-; X64-NEXT:    btl %edi, %ecx
-; X64-NEXT:    setb %al
-; X64-NEXT:    retq
+; X64-NOBMI2-LABEL: PR55138:
+; X64-NOBMI2:       ## %bb.0:
+; X64-NOBMI2-NEXT:    andl $15, %edi
+; X64-NOBMI2-NEXT:    movl $27030, %ecx ## imm = 0x6996
+; X64-NOBMI2-NEXT:    xorl %eax, %eax
+; X64-NOBMI2-NEXT:    btl %edi, %ecx
+; X64-NOBMI2-NEXT:    setb %al
+; X64-NOBMI2-NEXT:    retq
+;
+; X64-BMI2-LABEL: PR55138:
+; X64-BMI2:       ## %bb.0:
+; X64-BMI2-NEXT:    andb $15, %dil
+; X64-BMI2-NEXT:    movl $27030, %eax ## imm = 0x6996
+; X64-BMI2-NEXT:    shrxl %edi, %eax, %eax
+; X64-BMI2-NEXT:    andl $1, %eax
+; X64-BMI2-NEXT:    retq
   %urem = and i32 %x, 15
   %shr = lshr i32 27030, %urem
   %and = and i32 %shr, 1