[GlobalISel] Add computeNumSignBits for G_SHUFFLE_VECTOR

The code is similar to computeKnownBits and the code in
SelectionDAG::ComputeNumSignBits

Author: davemgreen

llvm/lib/CodeGen/GlobalISel/GISelValueTracking.cpp

@@ -14,6 +14,7 @@
 #include "llvm/CodeGen/GlobalISel/GISelValueTracking.h"
 #include "llvm/ADT/StringExtras.h"
 #include "llvm/Analysis/ValueTracking.h"
+#include "llvm/Analysis/VectorUtils.h"
 #include "llvm/CodeGen/GlobalISel/GenericMachineInstrs.h"
 #include "llvm/CodeGen/GlobalISel/Utils.h"
 #include "llvm/CodeGen/MachineFrameInfo.h"
@@ -836,6 +837,28 @@ unsigned GISelValueTracking::computeNumSignBits(Register R,
       return TyBits - 1; // Every always-zero bit is a sign bit.
     break;
   }
+  case TargetOpcode::G_SHUFFLE_VECTOR: {
+    // Collect the minimum number of sign bits that are shared by every vector
+    // element referenced by the shuffle.
+    APInt DemandedLHS, DemandedRHS;
+    unsigned NumElts = MRI.getType(MI.getOperand(1).getReg()).getNumElements();
+    if (!getShuffleDemandedElts(NumElts, MI.getOperand(3).getShuffleMask(),
+                                DemandedElts, DemandedLHS, DemandedRHS))
+      return 1;
+
+    if (!!DemandedLHS)
+      FirstAnswer =
+          computeNumSignBits(MI.getOperand(1).getReg(), DemandedLHS, Depth + 1);
+    // If we don't know anything, early out and try computeKnownBits fall-back.
+    if (FirstAnswer == 1)
+      break;
+    if (!!DemandedRHS) {
+      unsigned Tmp2 =
+          computeNumSignBits(MI.getOperand(2).getReg(), DemandedRHS, Depth + 1);
+      FirstAnswer = std::min(FirstAnswer, Tmp2);
+    }
+    break;
+  }
   case TargetOpcode::G_INTRINSIC:
   case TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS:
   case TargetOpcode::G_INTRINSIC_CONVERGENT:

llvm/test/CodeGen/AArch64/aarch64-dup-ext.ll

@@ -240,14 +240,9 @@ define <2 x i64> @dupzext_v2i16_v2i64(i16 %src, <2 x i16> %b) {
 ; CHECK-GI-NEXT:    and x8, x0, #0xffff
 ; CHECK-GI-NEXT:    and v0.16b, v0.16b, v1.16b
 ; CHECK-GI-NEXT:    dup v1.2d, x8
-; CHECK-GI-NEXT:    fmov x8, d1
-; CHECK-GI-NEXT:    fmov x9, d0
-; CHECK-GI-NEXT:    mov x10, v1.d[1]
-; CHECK-GI-NEXT:    mov x11, v0.d[1]
-; CHECK-GI-NEXT:    mul x8, x8, x9
-; CHECK-GI-NEXT:    mul x9, x10, x11
-; CHECK-GI-NEXT:    mov v0.d[0], x8
-; CHECK-GI-NEXT:    mov v0.d[1], x9
+; CHECK-GI-NEXT:    xtn v1.2s, v1.2d
+; CHECK-GI-NEXT:    xtn v0.2s, v0.2d
+; CHECK-GI-NEXT:    smull v0.2d, v1.2s, v0.2s
 ; CHECK-GI-NEXT:    ret
 entry:
   %in = zext i16 %src to i64
@@ -419,9 +414,10 @@ define <8 x i16> @missing_insert(<8 x i8> %b) {
 ;
 ; CHECK-GI-LABEL: missing_insert:
 ; CHECK-GI:       // %bb.0: // %entry
-; CHECK-GI-NEXT:    sshll v0.8h, v0.8b, #0
-; CHECK-GI-NEXT:    ext v1.16b, v0.16b, v0.16b, #4
-; CHECK-GI-NEXT:    mul v0.8h, v1.8h, v0.8h
+; CHECK-GI-NEXT:    sshll v1.8h, v0.8b, #0
+; CHECK-GI-NEXT:    ext v1.16b, v1.16b, v1.16b, #4
+; CHECK-GI-NEXT:    xtn v1.8b, v1.8h
+; CHECK-GI-NEXT:    smull v0.8h, v1.8b, v0.8b
 ; CHECK-GI-NEXT:    ret
 entry:
   %ext.b = sext <8 x i8> %b to <8 x i16>
@@ -440,10 +436,10 @@ define <8 x i16> @shufsext_v8i8_v8i16(<8 x i8> %src, <8 x i8> %b) {
 ; CHECK-GI-LABEL: shufsext_v8i8_v8i16:
 ; CHECK-GI:       // %bb.0: // %entry
 ; CHECK-GI-NEXT:    sshll v0.8h, v0.8b, #0
-; CHECK-GI-NEXT:    sshll v1.8h, v1.8b, #0
 ; CHECK-GI-NEXT:    rev64 v0.8h, v0.8h
 ; CHECK-GI-NEXT:    ext v0.16b, v0.16b, v0.16b, #8
-; CHECK-GI-NEXT:    mul v0.8h, v0.8h, v1.8h
+; CHECK-GI-NEXT:    xtn v0.8b, v0.8h
+; CHECK-GI-NEXT:    smull v0.8h, v0.8b, v1.8b
 ; CHECK-GI-NEXT:    ret
 entry:
   %in = sext <8 x i8> %src to <8 x i16>
@@ -463,16 +459,9 @@ define <2 x i64> @shufsext_v2i32_v2i64(<2 x i32> %src, <2 x i32> %b) {
 ; CHECK-GI-LABEL: shufsext_v2i32_v2i64:
 ; CHECK-GI:       // %bb.0: // %entry
 ; CHECK-GI-NEXT:    sshll v0.2d, v0.2s, #0
-; CHECK-GI-NEXT:    sshll v1.2d, v1.2s, #0
 ; CHECK-GI-NEXT:    ext v0.16b, v0.16b, v0.16b, #8
-; CHECK-GI-NEXT:    fmov x9, d1
-; CHECK-GI-NEXT:    mov x11, v1.d[1]
-; CHECK-GI-NEXT:    fmov x8, d0
-; CHECK-GI-NEXT:    mov x10, v0.d[1]
-; CHECK-GI-NEXT:    mul x8, x8, x9
-; CHECK-GI-NEXT:    mul x9, x10, x11
-; CHECK-GI-NEXT:    mov v0.d[0], x8
-; CHECK-GI-NEXT:    mov v0.d[1], x9
+; CHECK-GI-NEXT:    xtn v0.2s, v0.2d
+; CHECK-GI-NEXT:    smull v0.2d, v0.2s, v1.2s
 ; CHECK-GI-NEXT:    ret
 entry:
   %in = sext <2 x i32> %src to <2 x i64>
@@ -515,16 +504,9 @@ define <2 x i64> @shufzext_v2i32_v2i64(<2 x i32> %src, <2 x i32> %b) {
 ; CHECK-GI-LABEL: shufzext_v2i32_v2i64:
 ; CHECK-GI:       // %bb.0: // %entry
 ; CHECK-GI-NEXT:    sshll v0.2d, v0.2s, #0
-; CHECK-GI-NEXT:    sshll v1.2d, v1.2s, #0
 ; CHECK-GI-NEXT:    ext v0.16b, v0.16b, v0.16b, #8
-; CHECK-GI-NEXT:    fmov x9, d1
-; CHECK-GI-NEXT:    mov x11, v1.d[1]
-; CHECK-GI-NEXT:    fmov x8, d0
-; CHECK-GI-NEXT:    mov x10, v0.d[1]
-; CHECK-GI-NEXT:    mul x8, x8, x9
-; CHECK-GI-NEXT:    mul x9, x10, x11
-; CHECK-GI-NEXT:    mov v0.d[0], x8
-; CHECK-GI-NEXT:    mov v0.d[1], x9
+; CHECK-GI-NEXT:    xtn v0.2s, v0.2d
+; CHECK-GI-NEXT:    smull v0.2d, v0.2s, v1.2s
 ; CHECK-GI-NEXT:    ret
 entry:
   %in = sext <2 x i32> %src to <2 x i64>

llvm/test/CodeGen/AArch64/aarch64-matrix-umull-smull.ll

@@ -843,30 +843,18 @@ define void @sink_v4i64_1(ptr %p, ptr %d, i64 %n, <2 x i32> %a) {
 ; CHECK-GI-NEXT:    sshll v0.2d, v0.2s, #0
 ; CHECK-GI-NEXT:    mov x8, xzr
 ; CHECK-GI-NEXT:    dup v0.2d, v0.d[1]
-; CHECK-GI-NEXT:    mov x9, v0.d[1]
-; CHECK-GI-NEXT:    fmov x10, d0
+; CHECK-GI-NEXT:    xtn v0.2s, v0.2d
 ; CHECK-GI-NEXT:  .LBB7_1: // %loop
 ; CHECK-GI-NEXT:    // =>This Inner Loop Header: Depth=1
-; CHECK-GI-NEXT:    ldr q0, [x0]
+; CHECK-GI-NEXT:    ldr q1, [x0]
 ; CHECK-GI-NEXT:    subs x2, x2, #8
 ; CHECK-GI-NEXT:    add x8, x8, #8
-; CHECK-GI-NEXT:    sshll v1.2d, v0.2s, #0
-; CHECK-GI-NEXT:    sshll2 v0.2d, v0.4s, #0
-; CHECK-GI-NEXT:    fmov x11, d1
-; CHECK-GI-NEXT:    mov x12, v1.d[1]
-; CHECK-GI-NEXT:    fmov x13, d0
-; CHECK-GI-NEXT:    mov x14, v0.d[1]
-; CHECK-GI-NEXT:    mul x11, x11, x10
-; CHECK-GI-NEXT:    mul x13, x13, x10
-; CHECK-GI-NEXT:    mul x12, x12, x9
-; CHECK-GI-NEXT:    mov v0.d[0], x11
-; CHECK-GI-NEXT:    mul x11, x14, x9
-; CHECK-GI-NEXT:    mov v1.d[0], x13
-; CHECK-GI-NEXT:    mov v0.d[1], x12
-; CHECK-GI-NEXT:    mov v1.d[1], x11
-; CHECK-GI-NEXT:    shrn v0.2s, v0.2d, #15
-; CHECK-GI-NEXT:    shrn2 v0.4s, v1.2d, #15
-; CHECK-GI-NEXT:    str q0, [x0], #32
+; CHECK-GI-NEXT:    mov d2, v1.d[1]
+; CHECK-GI-NEXT:    smull v1.2d, v1.2s, v0.2s
+; CHECK-GI-NEXT:    smull v2.2d, v2.2s, v0.2s
+; CHECK-GI-NEXT:    shrn v1.2s, v1.2d, #15
+; CHECK-GI-NEXT:    shrn2 v1.4s, v2.2d, #15
+; CHECK-GI-NEXT:    str q1, [x0], #32
 ; CHECK-GI-NEXT:    b.ne .LBB7_1
 ; CHECK-GI-NEXT:  // %bb.2: // %exit
 ; CHECK-GI-NEXT:    ret
@@ -979,18 +967,19 @@ define void @sink_v16s16_8(ptr %p, ptr %d, i64 %n, <16 x i8> %a) {
 ; CHECK-GI:       // %bb.0: // %entry
 ; CHECK-GI-NEXT:    sshll2 v0.8h, v0.16b, #0
 ; CHECK-GI-NEXT:    mov x8, xzr
+; CHECK-GI-NEXT:    dup v0.8h, v0.h[2]
+; CHECK-GI-NEXT:    xtn v0.8b, v0.8h
 ; CHECK-GI-NEXT:  .LBB9_1: // %loop
 ; CHECK-GI-NEXT:    // =>This Inner Loop Header: Depth=1
 ; CHECK-GI-NEXT:    ldr q1, [x0]
 ; CHECK-GI-NEXT:    subs x2, x2, #8
 ; CHECK-GI-NEXT:    add x8, x8, #8
-; CHECK-GI-NEXT:    sshll v2.8h, v1.8b, #0
-; CHECK-GI-NEXT:    sshll2 v1.8h, v1.16b, #0
-; CHECK-GI-NEXT:    mul v2.8h, v2.8h, v0.h[2]
-; CHECK-GI-NEXT:    mul v1.8h, v1.8h, v0.h[2]
-; CHECK-GI-NEXT:    sshr v2.8h, v2.8h, #15
+; CHECK-GI-NEXT:    mov d2, v1.d[1]
+; CHECK-GI-NEXT:    smull v1.8h, v1.8b, v0.8b
+; CHECK-GI-NEXT:    smull v2.8h, v2.8b, v0.8b
 ; CHECK-GI-NEXT:    sshr v1.8h, v1.8h, #15
-; CHECK-GI-NEXT:    uzp1 v1.16b, v2.16b, v1.16b
+; CHECK-GI-NEXT:    sshr v2.8h, v2.8h, #15
+; CHECK-GI-NEXT:    uzp1 v1.16b, v1.16b, v2.16b
 ; CHECK-GI-NEXT:    str q1, [x0], #32
 ; CHECK-GI-NEXT:    b.ne .LBB9_1
 ; CHECK-GI-NEXT:  // %bb.2: // %exit