[SelectionDAG] ComputeNumSignBits of TRUNCATE operations

[RKSimon]

Summary

Add basic ComputeNumSignBits support for TRUNCATE ops for cases where the source's number of sign bits overlaps with the truncated size.

Improves X86 SIGN_EXTEND_IN_REG vector cases which were needlessly sign extending boolean vector results.

[efriedma]

lib/CodeGen/SelectionDAG/SelectionDAG.cpp

2833

Please use variable names which are not "Tmp" and "Tmp2".

[RKSimon]

lib/CodeGen/SelectionDAG/SelectionDAG.cpp

2833

No problem - but these are the default variables that are already used in most of the other cases.

[efriedma]

lib/CodeGen/SelectionDAG/SelectionDAG.cpp

2833

It probably isn't a good idea for the other cases. :)

[RKSimon]

More meaningful variable names.

[efriedma]

LGTM.

lib/CodeGen/SelectionDAG/SelectionDAG.cpp

if (ConstantSDNode *CLHS = isConstOrConstSplat(Op.getOperand(0)))
       }
 
     // Sub can have at most one carry bit.  Thus we know that the output
     // is, at worst, one more bit than the inputs.
     Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
     if (Tmp == 1) return 1;  // Early out.
     return std::min(Tmp, Tmp2)-1;
   case ISD::TRUNCATE:
-    // FIXME: it's tricky to do anything useful for this, but it is an important
-    // case for targets like X86.
+    // Check if the sign bits of source go down as far as the truncated value.
+    Tmp = Op.getOperand(0).getScalarValueSizeInBits();
+    Tmp2 = ComputeNumSignBits(Op.getOperand(0), Depth + 1);
+    if (Tmp2 > (Tmp - VTBits))
+      return Tmp2 - (Tmp - VTBits);
     break;
   case ISD::EXTRACT_ELEMENT: {
     const int KnownSign = ComputeNumSignBits(Op.getOperand(0), Depth+1);
     const int BitWidth = Op.getValueSizeInBits();
     const int Items = Op.getOperand(0).getValueSizeInBits() / BitWidth;
 
     // Get reverse index (starting from 1), Op1 value indexes elements from
     // little end. Sign starts at big end.

test/CodeGen/X86/avx512-mask-op.ll

entry:
   ret i8 %mask_convert
 }
 
 define <4 x i32> @test4(<4 x i64> %x, <4 x i64> %y, <4 x i64> %x1, <4 x i64> %y1) {
 ; KNL-LABEL: test4:
 ; KNL:       ## BB#0:
 ; KNL-NEXT:    vpcmpgtq %ymm1, %ymm0, %ymm0
 ; KNL-NEXT:    vpmovqd %zmm0, %ymm0
-; KNL-NEXT:    vpslld $31, %xmm0, %xmm0
-; KNL-NEXT:    vpsrad $31, %xmm0, %xmm0
 ; KNL-NEXT:    vpcmpgtq %ymm3, %ymm2, %ymm1
 ; KNL-NEXT:    vpmovqd %zmm1, %ymm1
-; KNL-NEXT:    vpslld $31, %xmm1, %xmm1
-; KNL-NEXT:    vpsrad $31, %xmm1, %xmm1
 ; KNL-NEXT:    vpcmpgtd %xmm1, %xmm0, %xmm0
 ; KNL-NEXT:    retq
 ;
 ; SKX-LABEL: test4:
 ; SKX:       ## BB#0:
 ; SKX-NEXT:    vpcmpgtq %ymm1, %ymm0, %k0
 ; SKX-NEXT:    vpcmpgtq %ymm3, %ymm2, %k1
 ; SKX-NEXT:    kandnw %k0, %k1, %k0

test/CodeGen/X86/psubus.ll

 ; AVX1-NEXT:    vpxor %xmm3, %xmm2, %xmm4
 ; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm5
 ; AVX1-NEXT:    vpxor %xmm3, %xmm5, %xmm6
 ; AVX1-NEXT:    vpcmpgtd %xmm4, %xmm6, %xmm4
 ; AVX1-NEXT:    vpxor %xmm3, %xmm1, %xmm6
 ; AVX1-NEXT:    vxorps %xmm3, %xmm0, %xmm3
 ; AVX1-NEXT:    vpcmpgtd %xmm6, %xmm3, %xmm3
 ; AVX1-NEXT:    vpacksswb %xmm4, %xmm3, %xmm3
-; AVX1-NEXT:    vpsllw $15, %xmm3, %xmm3
-; AVX1-NEXT:    vpsraw $15, %xmm3, %xmm3
 ; AVX1-NEXT:    vpsubd %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    vpsubd %xmm5, %xmm2, %xmm1
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm2 = [0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm1, %xmm1
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm0, %xmm0
 ; AVX1-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX1-NEXT:    vpblendvb %xmm3, %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vpandn %xmm0, %xmm3, %xmm0
 ; AVX1-NEXT:    vmovdqu %xmm0, (%rdi)
 ; AVX1-NEXT:    vzeroupper
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: test13:
 ; AVX2:       ## BB#0: ## %vector.ph
 ; AVX2-NEXT:    vmovdqu (%rsi), %ymm0
 ; AVX2-NEXT:    vpmovzxwd {{.*#+}} ymm1 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero,mem[4],zero,mem[5],zero,mem[6],zero,mem[7],zero
 ; AVX2-NEXT:    vpbroadcastd {{.*}}(%rip), %ymm2
 ; AVX2-NEXT:    vpxor %ymm2, %ymm1, %ymm3
 ; AVX2-NEXT:    vpxor %ymm2, %ymm0, %ymm2
 ; AVX2-NEXT:    vpcmpgtd %ymm3, %ymm2, %ymm2
 ; AVX2-NEXT:    vextracti128 $1, %ymm2, %xmm3
 ; AVX2-NEXT:    vpacksswb %xmm3, %xmm2, %xmm2
-; AVX2-NEXT:    vpsllw $15, %xmm2, %xmm2
-; AVX2-NEXT:    vpsraw $15, %xmm2, %xmm2
 ; AVX2-NEXT:    vpsubd %ymm0, %ymm1, %ymm0
 ; AVX2-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,1,4,5,8,9,12,13],zero,zero,zero,zero,zero,zero,zero,zero,ymm0[16,17,20,21,24,25,28,29],zero,zero,zero,zero,zero,zero,zero,zero
 ; AVX2-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,2,3]
-; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vpblendvb %xmm2, %xmm1, %xmm0, %xmm0
+; AVX2-NEXT:    vpandn %xmm0, %xmm2, %xmm0
 ; AVX2-NEXT:    vmovdqu %xmm0, (%rdi)
 ; AVX2-NEXT:    vzeroupper
 ; AVX2-NEXT:    retq
 vector.ph:
   %0 = getelementptr inbounds i16, i16* %head, i64 0
   %1 = bitcast i16* %0 to <8 x i16>*
   %2 = load <8 x i16>, <8 x i16>* %1, align 2
   %3 = getelementptr inbounds i32, i32* %w, i64 0

 ; AVX1-NEXT:    vpxor %xmm3, %xmm2, %xmm4
 ; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm5
 ; AVX1-NEXT:    vpxor %xmm3, %xmm5, %xmm6
 ; AVX1-NEXT:    vpcmpgtd %xmm6, %xmm4, %xmm4
 ; AVX1-NEXT:    vpxor %xmm3, %xmm1, %xmm6
 ; AVX1-NEXT:    vxorps %xmm3, %xmm0, %xmm3
 ; AVX1-NEXT:    vpcmpgtd %xmm3, %xmm6, %xmm3
 ; AVX1-NEXT:    vpacksswb %xmm4, %xmm3, %xmm3
-; AVX1-NEXT:    vpsllw $15, %xmm3, %xmm3
-; AVX1-NEXT:    vpsraw $15, %xmm3, %xmm3
 ; AVX1-NEXT:    vpsubd %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    vpsubd %xmm5, %xmm2, %xmm1
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm2 = [0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm1, %xmm1
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm0, %xmm0
 ; AVX1-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
 ; AVX1-NEXT:    vpand %xmm0, %xmm3, %xmm0
 ; AVX1-NEXT:    vmovdqu %xmm0, (%rdi)
 ; AVX1-NEXT:    vzeroupper
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: test15:
 ; AVX2:       ## BB#0: ## %vector.ph
 ; AVX2-NEXT:    vmovdqu (%rsi), %ymm0
 ; AVX2-NEXT:    vpmovzxwd {{.*#+}} ymm1 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero,mem[4],zero,mem[5],zero,mem[6],zero,mem[7],zero
 ; AVX2-NEXT:    vpbroadcastd {{.*}}(%rip), %ymm2
 ; AVX2-NEXT:    vpxor %ymm2, %ymm0, %ymm3
 ; AVX2-NEXT:    vpxor %ymm2, %ymm1, %ymm2
 ; AVX2-NEXT:    vpcmpgtd %ymm3, %ymm2, %ymm2
 ; AVX2-NEXT:    vextracti128 $1, %ymm2, %xmm3
 ; AVX2-NEXT:    vpacksswb %xmm3, %xmm2, %xmm2
-; AVX2-NEXT:    vpsllw $15, %xmm2, %xmm2
-; AVX2-NEXT:    vpsraw $15, %xmm2, %xmm2
 ; AVX2-NEXT:    vpsubd %ymm0, %ymm1, %ymm0
 ; AVX2-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,1,4,5,8,9,12,13],zero,zero,zero,zero,zero,zero,zero,zero,ymm0[16,17,20,21,24,25,28,29],zero,zero,zero,zero,zero,zero,zero,zero
 ; AVX2-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,2,3]
 ; AVX2-NEXT:    vpand %xmm0, %xmm2, %xmm0
 ; AVX2-NEXT:    vmovdqu %xmm0, (%rdi)
 ; AVX2-NEXT:    vzeroupper
 ; AVX2-NEXT:    retq
 vector.ph:

 ; AVX1-NEXT:    vpxor %xmm3, %xmm2, %xmm4
 ; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm5
 ; AVX1-NEXT:    vpxor %xmm3, %xmm5, %xmm6
 ; AVX1-NEXT:    vpcmpgtd %xmm6, %xmm4, %xmm4
 ; AVX1-NEXT:    vpxor %xmm3, %xmm1, %xmm6
 ; AVX1-NEXT:    vxorps %xmm3, %xmm0, %xmm3
 ; AVX1-NEXT:    vpcmpgtd %xmm3, %xmm6, %xmm3
 ; AVX1-NEXT:    vpacksswb %xmm4, %xmm3, %xmm3
-; AVX1-NEXT:    vpsllw $15, %xmm3, %xmm3
-; AVX1-NEXT:    vpsraw $15, %xmm3, %xmm3
 ; AVX1-NEXT:    vpsubd %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    vpsubd %xmm5, %xmm2, %xmm1
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm2 = [0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15]
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm1, %xmm1
 ; AVX1-NEXT:    vpshufb %xmm2, %xmm0, %xmm0
 ; AVX1-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
 ; AVX1-NEXT:    vpand %xmm0, %xmm3, %xmm0
 ; AVX1-NEXT:    vmovdqu %xmm0, (%rdi)
 ; AVX1-NEXT:    vzeroupper
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: test16:
 ; AVX2:       ## BB#0: ## %vector.ph
 ; AVX2-NEXT:    vmovdqu (%rsi), %ymm0
 ; AVX2-NEXT:    vpmovzxwd {{.*#+}} ymm1 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero,mem[4],zero,mem[5],zero,mem[6],zero,mem[7],zero
 ; AVX2-NEXT:    vpbroadcastd {{.*}}(%rip), %ymm2
 ; AVX2-NEXT:    vpxor %ymm2, %ymm0, %ymm3
 ; AVX2-NEXT:    vpxor %ymm2, %ymm1, %ymm2
 ; AVX2-NEXT:    vpcmpgtd %ymm3, %ymm2, %ymm2
 ; AVX2-NEXT:    vextracti128 $1, %ymm2, %xmm3
 ; AVX2-NEXT:    vpacksswb %xmm3, %xmm2, %xmm2
-; AVX2-NEXT:    vpsllw $15, %xmm2, %xmm2
-; AVX2-NEXT:    vpsraw $15, %xmm2, %xmm2
 ; AVX2-NEXT:    vpsubd %ymm0, %ymm1, %ymm0
 ; AVX2-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,1,4,5,8,9,12,13],zero,zero,zero,zero,zero,zero,zero,zero,ymm0[16,17,20,21,24,25,28,29],zero,zero,zero,zero,zero,zero,zero,zero
 ; AVX2-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,2,3]
 ; AVX2-NEXT:    vpand %xmm0, %xmm2, %xmm0
 ; AVX2-NEXT:    vmovdqu %xmm0, (%rdi)
 ; AVX2-NEXT:    vzeroupper
 ; AVX2-NEXT:    retq
 vector.ph: