[SelectionDAG] Improve knownbits handling of UMIN/UMAX (PR31293)

[RKSimon]

Summary

This patch improves the knownbits logic for unsigned integer min/max opcodes.

For UMIN we know that the result will have the maximum of the inputs' known leading zero bits in the result, similarly for UMAX the maximum of the inputs' leading one bits.

This is particularly useful for simplifying clamping patterns,. e.g. as SSE doesn't have a uitofp instruction we want to use sitofp instead where possible and for that we need to confirm that the top bit is not set.

[arsenm]

lib/CodeGen/SelectionDAG/SelectionDAG.cpp

2712–2715 ↗

(On Diff #84818)

Should it also be the minimum of the known leading 0s?

[RKSimon]

Thanks Matt

lib/CodeGen/SelectionDAG/SelectionDAG.cpp

2712–2715 ↗

(On Diff #84818)

If one input is 0xFFFFxxxx and the other is 0xFxxxxxxx then we know that the UMAX result is at least 0xFFFF0000.

llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAG.cpp

void SelectionDAG::computeKnownBits(SDValue Op, APInt &KnownZero,
   }
   case ISD::BSWAP: {
     computeKnownBits(Op.getOperand(0), KnownZero2, KnownOne2, DemandedElts,
                      Depth + 1);
     KnownZero = KnownZero2.byteSwap();
     KnownOne = KnownOne2.byteSwap();
     break;
   }
-  case ISD::SMIN:
-  case ISD::SMAX:
-  case ISD::UMIN:
+  case ISD::UMIN: {
+    computeKnownBits(Op.getOperand(0), KnownZero, KnownOne, DemandedElts,
+                     Depth + 1);
+    computeKnownBits(Op.getOperand(1), KnownZero2, KnownOne2, DemandedElts,
+                     Depth + 1);
+
+    // UMIN - we know that the result will have the maximum of the
+    // known zero leading bits of the inputs.
+    unsigned LeadZero = KnownZero.countLeadingOnes();
+    LeadZero = std::max(LeadZero, KnownZero2.countLeadingOnes());
+
+    KnownZero &= KnownZero2;
+    KnownOne &= KnownOne2;
+    KnownZero |= APInt::getHighBitsSet(BitWidth, LeadZero);
+    break;
+  }
   case ISD::UMAX: {
     computeKnownBits(Op.getOperand(0), KnownZero, KnownOne, DemandedElts,
                      Depth + 1);
+    computeKnownBits(Op.getOperand(1), KnownZero2, KnownOne2, DemandedElts,
+                     Depth + 1);
+
+    // UMAX - we know that the result will have the maximum of the
+    // known one leading bits of the inputs.
+    unsigned LeadOne = KnownOne.countLeadingOnes();
+    LeadOne = std::max(LeadOne, KnownOne2.countLeadingOnes());
+
+    KnownZero &= KnownZero2;
+    KnownOne &= KnownOne2;
+    KnownOne |= APInt::getHighBitsSet(BitWidth, LeadOne);
+    break;
+  }
+  case ISD::SMIN:
+  case ISD::SMAX: {
+    computeKnownBits(Op.getOperand(0), KnownZero, KnownOne, DemandedElts,
+                     Depth + 1);
     // If we don't know any bits, early out.
     if (!KnownOne && !KnownZero)
       break;
     computeKnownBits(Op.getOperand(1), KnownZero2, KnownOne2, DemandedElts,
                      Depth + 1);
     KnownZero &= KnownZero2;
     KnownOne &= KnownOne2;
     break;

llvm/trunk/test/CodeGen/X86/known-bits-vector.ll

 declare <4 x i32> @llvm.x86.sse41.pmaxsd(<4 x i32>, <4 x i32>) nounwind readnone
 declare <4 x i32> @llvm.x86.sse41.pminsd(<4 x i32>, <4 x i32>) nounwind readnone
 
 define <4 x float> @knownbits_umin_shuffle_uitofp(<4 x i32> %a0) {
 ; X32-LABEL: knownbits_umin_shuffle_uitofp:
 ; X32:       # BB#0:
 ; X32-NEXT:    vpminud {{\.LCPI.*}}, %xmm0, %xmm0
 ; X32-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,0,3,3]
-; X32-NEXT:    vpblendw {{.*#+}} xmm1 = xmm0[0],mem[1],xmm0[2],mem[3],xmm0[4],mem[5],xmm0[6],mem[7]
-; X32-NEXT:    vpsrld $16, %xmm0, %xmm0
-; X32-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0],mem[1],xmm0[2],mem[3],xmm0[4],mem[5],xmm0[6],mem[7]
-; X32-NEXT:    vaddps {{\.LCPI.*}}, %xmm0, %xmm0
-; X32-NEXT:    vaddps %xmm0, %xmm1, %xmm0
+; X32-NEXT:    vcvtdq2ps %xmm0, %xmm0
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: knownbits_umin_shuffle_uitofp:
 ; X64:       # BB#0:
 ; X64-NEXT:    vpminud {{.*}}(%rip), %xmm0, %xmm0
 ; X64-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,0,3,3]
-; X64-NEXT:    vpblendw {{.*#+}} xmm1 = xmm0[0],mem[1],xmm0[2],mem[3],xmm0[4],mem[5],xmm0[6],mem[7]
-; X64-NEXT:    vpsrld $16, %xmm0, %xmm0
-; X64-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0],mem[1],xmm0[2],mem[3],xmm0[4],mem[5],xmm0[6],mem[7]
-; X64-NEXT:    vaddps {{.*}}(%rip), %xmm0, %xmm0
-; X64-NEXT:    vaddps %xmm0, %xmm1, %xmm0
+; X64-NEXT:    vcvtdq2ps %xmm0, %xmm0
 ; X64-NEXT:    retq
   %1 = call <4 x i32> @llvm.x86.sse41.pminud(<4 x i32> %a0, <4 x i32> <i32 65535, i32 -1, i32 -1, i32 262143>)
   %2 = shufflevector <4 x i32> %1, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 3, i32 3>
   %3 = uitofp <4 x i32> %2 to <4 x float>
   ret <4 x float> %3
 }
 declare <4 x i32> @llvm.x86.sse41.pmaxud(<4 x i32>, <4 x i32>) nounwind readnone
 declare <4 x i32> @llvm.x86.sse41.pminud(<4 x i32>, <4 x i32>) nounwind readnone
 
 define <4 x i32> @knownbits_umax_shuffle_ashr(<4 x i32> %a0) {
 ; X32-LABEL: knownbits_umax_shuffle_ashr:
 ; X32:       # BB#0:
 ; X32-NEXT:    vpmaxud {{\.LCPI.*}}, %xmm0, %xmm0
 ; X32-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[1,1,2,2]
-; X32-NEXT:    vpsrad $31, %xmm0, %xmm0
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: knownbits_umax_shuffle_ashr:
 ; X64:       # BB#0:
 ; X64-NEXT:    vpmaxud {{.*}}(%rip), %xmm0, %xmm0
 ; X64-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[1,1,2,2]
-; X64-NEXT:    vpsrad $31, %xmm0, %xmm0
 ; X64-NEXT:    retq
   %1 = call <4 x i32> @llvm.x86.sse41.pmaxud(<4 x i32> %a0, <4 x i32> <i32 65535, i32 -1, i32 -1, i32 262143>)
   %2 = shufflevector <4 x i32> %1, <4 x i32> undef, <4 x i32> <i32 1, i32 1, i32 2, i32 2>
   %3 = ashr <4 x i32> %2, <i32 31, i32 31, i32 31, i32 31>
   ret <4 x i32> %3
 }
 
 define <4 x float> @knownbits_mask_umax_shuffle_uitofp(<4 x i32> %a0) {