[llvm] 9485d98 - [InstCombine] Disable generation of fshl/fshr for rotates

[Paulo Matos via llvm-commits]

Author: Paulo Matos
Date: 2023-06-01T15:31:49+02:00
New Revision: 9485d983ac0c56d412c958b40f4e150a3c30ca7c

URL: https://github.com/llvm/llvm-project/commit/9485d983ac0c56d412c958b40f4e150a3c30ca7c
DIFF: https://github.com/llvm/llvm-project/commit/9485d983ac0c56d412c958b40f4e150a3c30ca7c.diff

LOG: [InstCombine] Disable generation of fshl/fshr for rotates

Disable conversion of funnel shifts (fshl/fshr) into rotates
unless one of the operands is known to be a constant value.

Reviewed By: nikic

Differential Revision: https://reviews.llvm.org/D150670

Added: 
    

Modified: 
    llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp
    llvm/test/Transforms/InstCombine/fsh.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp b/llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp
index 3e0c0247ec065..56e4515a28b16 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp
@@ -912,9 +912,26 @@ Value *InstCombinerImpl::SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
 
         APInt DemandedMaskLHS(DemandedMask.lshr(ShiftAmt));
         APInt DemandedMaskRHS(DemandedMask.shl(BitWidth - ShiftAmt));
-        if (SimplifyDemandedBits(I, 0, DemandedMaskLHS, LHSKnown, Depth + 1) ||
-            SimplifyDemandedBits(I, 1, DemandedMaskRHS, RHSKnown, Depth + 1))
-          return I;
+        if (I->getOperand(0) != I->getOperand(1)) {
+          if (SimplifyDemandedBits(I, 0, DemandedMaskLHS, LHSKnown,
+                                   Depth + 1) ||
+              SimplifyDemandedBits(I, 1, DemandedMaskRHS, RHSKnown, Depth + 1))
+            return I;
+        } else { // fshl is a rotate
+        // Avoid converting rotate into funnel shift. 
+        // Only simplify if one operand is constant.
+          KnownBits LHSKnown = computeKnownBits(I->getOperand(0), Depth + 1, I);
+          if (DemandedMaskLHS.isSubsetOf(LHSKnown.Zero | LHSKnown.One)) {
+            replaceOperand(*I, 0, Constant::getIntegerValue(VTy, LHSKnown.One));
+            return I;
+          }
+
+          KnownBits RHSKnown = computeKnownBits(I->getOperand(1), Depth + 1, I);
+          if (DemandedMaskRHS.isSubsetOf(RHSKnown.Zero | RHSKnown.One)) {
+            replaceOperand(*I, 1, Constant::getIntegerValue(VTy, RHSKnown.One));
+            return I;
+          }
+        }
 
         Known.Zero = LHSKnown.Zero.shl(ShiftAmt) |
                      RHSKnown.Zero.lshr(BitWidth - ShiftAmt);

diff  --git a/llvm/test/Transforms/InstCombine/fsh.ll b/llvm/test/Transforms/InstCombine/fsh.ll
index 489f6e686680b..48bf296993f6a 100644
--- a/llvm/test/Transforms/InstCombine/fsh.ll
+++ b/llvm/test/Transforms/InstCombine/fsh.ll
@@ -440,12 +440,10 @@ define <2 x i32> @fshr_undef_shift_amount_vec(<2 x i32> %x, <2 x i32> %y) {
   ret <2 x i32> %r
 }
 
-; TODO: Don't let SimplifyDemandedBits split up a rotate - keep the same operand.
-
 define i32 @rotl_common_demanded(i32 %a0) {
 ; CHECK-LABEL: @rotl_common_demanded(
 ; CHECK-NEXT:    [[X:%.*]] = xor i32 [[A0:%.*]], 2
-; CHECK-NEXT:    [[R:%.*]] = call i32 @llvm.fshl.i32(i32 [[X]], i32 [[A0]], i32 8)
+; CHECK-NEXT:    [[R:%.*]] = call i32 @llvm.fshl.i32(i32 [[X]], i32 [[X]], i32 8)
 ; CHECK-NEXT:    ret i32 [[R]]
 ;
   %x = xor i32 %a0, 2
@@ -456,7 +454,7 @@ define i32 @rotl_common_demanded(i32 %a0) {
 define i33 @rotr_common_demanded(i33 %a0) {
 ; CHECK-LABEL: @rotr_common_demanded(
 ; CHECK-NEXT:    [[X:%.*]] = xor i33 [[A0:%.*]], 2
-; CHECK-NEXT:    [[R:%.*]] = call i33 @llvm.fshl.i33(i33 [[X]], i33 [[A0]], i33 25)
+; CHECK-NEXT:    [[R:%.*]] = call i33 @llvm.fshl.i33(i33 [[X]], i33 [[X]], i33 25)
 ; CHECK-NEXT:    ret i33 [[R]]
 ;
   %x = xor i33 %a0, 2
@@ -704,6 +702,26 @@ define i32 @fshl_mask_args_
diff erent(i32 %a) {
   ret i32 %t3
 }
 
+define i32 @fsh_andconst_rotate(i32 %a) {
+; CHECK-LABEL: @fsh_andconst_rotate(
+; CHECK-NEXT:    [[T2:%.*]] = lshr i32 [[A:%.*]], 16
+; CHECK-NEXT:    ret i32 [[T2]]
+;
+  %t1 = and i32 %a, 4294901760 ; 0xffff0000
+  %t2 = call i32 @llvm.fshl.i32(i32 %t1, i32 %t1, i32 16)
+  ret i32 %t2
+}
+
+define i32 @fsh_orconst_rotate(i32 %a) {
+; CHECK-LABEL: @fsh_orconst_rotate(
+; CHECK-NEXT:    [[T2:%.*]] = call i32 @llvm.fshl.i32(i32 [[A:%.*]], i32 -268435456, i32 4)
+; CHECK-NEXT:    ret i32 [[T2]]
+;
+  %t1 = or i32 %a, 4026531840 ; 0xf0000000
+  %t2 = call i32 @llvm.fshl.i32(i32 %t1, i32 %t1, i32 4)
+  ret i32 %t2
+}
+
 define <2 x i31> @fshr_mask_args_same_vector(<2 x i31> %a) {
 ; CHECK-LABEL: @fshr_mask_args_same_vector(
 ; CHECK-NEXT:    [[T3:%.*]] = shl <2 x i31> [[A:%.*]], <i31 10, i31 10>