[X86] X86DAGToDAGISel::matchBitExtract(): pattern b: truncation awareness

[lebedev.ri]

Summary

(Not so) boringly identical to pattern a (D62786)
Not yet sure how do deal with the last pattern c.

[lebedev.ri]

Rebased

[lebedev.ri]

Rebased, extra test exposed an alternative way to still end up
with unrecognized pattern with truncation, so handle it too.

This resulted in inlining isBitwiseNot(), not sure if it would be better to extend it with all that instead?

[RKSimon]

lib/CodeGen/SelectionDAG/SelectionDAG.cpp

8623 ↗

(On Diff #202612)

Could you not create an AllOnes equivalent MaskedValueIsZero instead? MaskedValueIsAllOnes?

[lebedev.ri]

lib/CodeGen/SelectionDAG/SelectionDAG.cpp

8623 ↗

(On Diff #202612)

Err, s/not//?
"Could you create an AllOnes equivalent of MaskedValueIsZero instead? MaskedValueIsAllOnes?"

[lebedev.ri]

Introduce SelectionDAG::MaskedValueIsAllOnes().
Is this better?

[RKSimon]

lib/CodeGen/SelectionDAG/SelectionDAG.cpp

8623 ↗

(On Diff #202612)

I don't think we need a DemandedElts variant so this should work.

/// MaskedValueIsAllOnes - Return true if 'V & Mask' is known to be all ones.
bool MaskedValueIsAllOnes(SDValue Op, const APInt &Mask, unsigned Depth = 0) const;

bool SelectionDAG::MaskedValueIsAllOnes(SDValue V, const APInt &Mask, unsigned Depth) const {
  return Mask.isSubsetOf(computeKnownBits(V, Depth).One);
}

[lebedev.ri]

Drop demandedelements variant

[RKSimon]

LGTM with one minor

lib/CodeGen/SelectionDAG/SelectionDAG.cpp

2248 ↗

(On Diff #202617)

Remove this line - its MaskedValueIsZero specific

[lebedev.ri]

Rebased.

lib/Target/X86/X86ISelDAGToDAG.cpp

if (M0->getOpcode() != ISD::SHL || !checkOneUse(M0))
       return false;
     if (!isOneConstant(M0->getOperand(0)))
       return false;
     NBits = M0->getOperand(1);
     return true;
   };
 
   // b) x & ~(-1 << nbits)
-  auto matchPatternB = [&checkOneUse, &NBits](SDValue Mask) -> bool {
+  auto matchPatternB = [&checkOneUse, &peekThroughOneUseTruncation,
+                        &NBits](SDValue Mask) -> bool {
     // Match `~()`. Must only have one use!
     if (!isBitwiseNot(Mask) || !checkOneUse(Mask))
       return false;
-    // Match `-1 << nbits`. Must only have one use!
-    SDValue M0 = Mask->getOperand(0);
+    // Match `-1 << nbits`. Might be truncated. Must only have one use!
+    SDValue M0 = peekThroughOneUseTruncation(Mask->getOperand(0));
     if (M0->getOpcode() != ISD::SHL || !checkOneUse(M0))
       return false;
     if (!isAllOnesConstant(M0->getOperand(0)))
       return false;
     NBits = M0->getOperand(1);
     return true;
   };
 

test/CodeGen/X86/extract-bits.ll

 ; X64-NOBMI-NEXT:    shlq %cl, %rax
 ; X64-NOBMI-NEXT:    notl %eax
 ; X64-NOBMI-NEXT:    andl %edi, %eax
 ; X64-NOBMI-NEXT:    # kill: def $eax killed $eax killed $rax
 ; X64-NOBMI-NEXT:    retq
 ;
 ; X64-BMI1NOTBM-LABEL: bextr64_32_b0:
 ; X64-BMI1NOTBM:       # %bb.0:
-; X64-BMI1NOTBM-NEXT:    movq %rsi, %rcx
-; X64-BMI1NOTBM-NEXT:    # kill: def $cl killed $cl killed $rcx
-; X64-BMI1NOTBM-NEXT:    shrq %cl, %rdi
-; X64-BMI1NOTBM-NEXT:    movq $-1, %rax
-; X64-BMI1NOTBM-NEXT:    movl %edx, %ecx
-; X64-BMI1NOTBM-NEXT:    shlq %cl, %rax
-; X64-BMI1NOTBM-NEXT:    andnl %edi, %eax, %eax
+; X64-BMI1NOTBM-NEXT:    shll $8, %edx
+; X64-BMI1NOTBM-NEXT:    movzbl %sil, %eax
+; X64-BMI1NOTBM-NEXT:    orl %edx, %eax
+; X64-BMI1NOTBM-NEXT:    bextrq %rax, %rdi, %rax
+; X64-BMI1NOTBM-NEXT:    # kill: def $eax killed $eax killed $rax
 ; X64-BMI1NOTBM-NEXT:    retq
 ;
 ; X64-BMI1BMI2-LABEL: bextr64_32_b0:
 ; X64-BMI1BMI2:       # %bb.0:
-; X64-BMI1BMI2-NEXT:    # kill: def $edx killed $edx def $rdx
 ; X64-BMI1BMI2-NEXT:    shrxq %rsi, %rdi, %rax
-; X64-BMI1BMI2-NEXT:    movq $-1, %rcx
-; X64-BMI1BMI2-NEXT:    shlxq %rdx, %rcx, %rcx
-; X64-BMI1BMI2-NEXT:    andnl %eax, %ecx, %eax
+; X64-BMI1BMI2-NEXT:    bzhil %edx, %eax, %eax
 ; X64-BMI1BMI2-NEXT:    retq
   %shiftedval = lshr i64 %val, %numskipbits
   %widenumlowbits = zext i8 %numlowbits to i64
   %notmask = shl nsw i64 -1, %widenumlowbits
   %mask = xor i64 %notmask, -1
   %wideres = and i64 %shiftedval, %mask
   %res = trunc i64 %wideres to i32
   ret i32 %res

test/CodeGen/X86/extract-lowbits.ll

 ; X64-NOBMI-NEXT:    shlq %cl, %rax
 ; X64-NOBMI-NEXT:    notl %eax
 ; X64-NOBMI-NEXT:    andl %edi, %eax
 ; X64-NOBMI-NEXT:    # kill: def $eax killed $eax killed $rax
 ; X64-NOBMI-NEXT:    retq
 ;
 ; X64-BMI1NOTBM-LABEL: bzhi64_32_b0:
 ; X64-BMI1NOTBM:       # %bb.0:
-; X64-BMI1NOTBM-NEXT:    movl %esi, %ecx
-; X64-BMI1NOTBM-NEXT:    movq $-1, %rax
-; X64-BMI1NOTBM-NEXT:    # kill: def $cl killed $cl killed $ecx
-; X64-BMI1NOTBM-NEXT:    shlq %cl, %rax
-; X64-BMI1NOTBM-NEXT:    andnl %edi, %eax, %eax
+; X64-BMI1NOTBM-NEXT:    shll $8, %esi
+; X64-BMI1NOTBM-NEXT:    bextrl %esi, %edi, %eax
 ; X64-BMI1NOTBM-NEXT:    retq
 ;
 ; X64-BMI1BMI2-LABEL: bzhi64_32_b0:
 ; X64-BMI1BMI2:       # %bb.0:
-; X64-BMI1BMI2-NEXT:    # kill: def $esi killed $esi def $rsi
-; X64-BMI1BMI2-NEXT:    movq $-1, %rax
-; X64-BMI1BMI2-NEXT:    shlxq %rsi, %rax, %rax
-; X64-BMI1BMI2-NEXT:    andnl %edi, %eax, %eax
+; X64-BMI1BMI2-NEXT:    bzhil %esi, %edi, %eax
 ; X64-BMI1BMI2-NEXT:    retq
   %widenumlowbits = zext i8 %numlowbits to i64
   %notmask = shl nsw i64 -1, %widenumlowbits
   %mask = xor i64 %notmask, -1
   %wideres = and i64 %val, %mask
   %res = trunc i64 %wideres to i32
   ret i32 %res
 }