[ARM] shouldFoldMaskToVariableShiftPair should be true for scalars up to the biggest legal type (llvm#158070)

For ARM, we want to do this up to 32-bits. Otherwise the code ends up
bigger and bloated.

Author: AZero13

llvm/lib/Target/ARM/ARMISelLowering.h

@@ -775,6 +775,16 @@ class VectorType;
     bool shouldFoldConstantShiftPairToMask(const SDNode *N,
                                            CombineLevel Level) const override;
 
+    /// Return true if it is profitable to fold a pair of shifts into a mask.
+    bool shouldFoldMaskToVariableShiftPair(SDValue Y) const override {
+      EVT VT = Y.getValueType();
+
+      if (VT.isVector())
+        return false;
+
+      return VT.getScalarSizeInBits() <= 32;
+    }
+
     bool shouldFoldSelectWithIdentityConstant(unsigned BinOpcode, EVT VT,
                                               unsigned SelectOpcode, SDValue X,
                                               SDValue Y) const override;

llvm/test/CodeGen/ARM/and-mask-variable.ll

@@ -0,0 +1,90 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=thumbv7m-eabi  %s -o - | FileCheck %s --check-prefix V7M
+; RUN: llc -mtriple=armv7a-eabi  %s -o -   | FileCheck %s --check-prefix V7A
+; RUN: llc -mtriple=thumbv7a-eabi  %s -o -   | FileCheck %s --check-prefix V7A-T
+; RUN: llc -mtriple=armv6m-eabi  %s -o -   | FileCheck %s --check-prefix V6M
+
+define i32 @mask_pair(i32 %x, i32 %y) {
+; V7M-LABEL: mask_pair:
+; V7M:       @ %bb.0:
+; V7M-NEXT:    lsrs r0, r1
+; V7M-NEXT:    lsls r0, r1
+; V7M-NEXT:    bx lr
+;
+; V7A-LABEL: mask_pair:
+; V7A:       @ %bb.0:
+; V7A-NEXT:    lsr r0, r0, r1
+; V7A-NEXT:    lsl r0, r0, r1
+; V7A-NEXT:    bx lr
+;
+; V7A-T-LABEL: mask_pair:
+; V7A-T:       @ %bb.0:
+; V7A-T-NEXT:    lsrs r0, r1
+; V7A-T-NEXT:    lsls r0, r1
+; V7A-T-NEXT:    bx lr
+;
+; V6M-LABEL: mask_pair:
+; V6M:       @ %bb.0:
+; V6M-NEXT:    lsrs r0, r1
+; V6M-NEXT:    lsls r0, r1
+; V6M-NEXT:    bx lr
+  %shl = shl nsw i32 -1, %y
+  %and = and i32 %shl, %x
+  ret i32 %and
+}
+
+define i64 @mask_pair_64(i64 %x, i64 %y) {
+; V7M-LABEL: mask_pair_64:
+; V7M:       @ %bb.0:
+; V7M-NEXT:    mov.w r3, #-1
+; V7M-NEXT:    lsl.w r12, r3, r2
+; V7M-NEXT:    subs r2, #32
+; V7M-NEXT:    it pl
+; V7M-NEXT:    movpl.w r12, #0
+; V7M-NEXT:    it pl
+; V7M-NEXT:    lslpl r3, r2
+; V7M-NEXT:    and.w r0, r0, r12
+; V7M-NEXT:    ands r1, r3
+; V7M-NEXT:    bx lr
+;
+; V7A-LABEL: mask_pair_64:
+; V7A:       @ %bb.0:
+; V7A-NEXT:    subs r12, r2, #32
+; V7A-NEXT:    mvn r3, #0
+; V7A-NEXT:    lsl r2, r3, r2
+; V7A-NEXT:    lslpl r3, r3, r12
+; V7A-NEXT:    movwpl r2, #0
+; V7A-NEXT:    and r1, r3, r1
+; V7A-NEXT:    and r0, r2, r0
+; V7A-NEXT:    bx lr
+;
+; V7A-T-LABEL: mask_pair_64:
+; V7A-T:       @ %bb.0:
+; V7A-T-NEXT:    mov.w r3, #-1
+; V7A-T-NEXT:    lsl.w r12, r3, r2
+; V7A-T-NEXT:    subs r2, #32
+; V7A-T-NEXT:    it pl
+; V7A-T-NEXT:    movpl.w r12, #0
+; V7A-T-NEXT:    it pl
+; V7A-T-NEXT:    lslpl r3, r2
+; V7A-T-NEXT:    and.w r0, r0, r12
+; V7A-T-NEXT:    ands r1, r3
+; V7A-T-NEXT:    bx lr
+;
+; V6M-LABEL: mask_pair_64:
+; V6M:       @ %bb.0:
+; V6M-NEXT:    .save {r4, r5, r7, lr}
+; V6M-NEXT:    push {r4, r5, r7, lr}
+; V6M-NEXT:    mov r4, r1
+; V6M-NEXT:    mov r5, r0
+; V6M-NEXT:    movs r0, #0
+; V6M-NEXT:    mvns r0, r0
+; V6M-NEXT:    mov r1, r0
+; V6M-NEXT:    bl __aeabi_llsl
+; V6M-NEXT:    ands r0, r5
+; V6M-NEXT:    ands r1, r4
+; V6M-NEXT:    pop {r4, r5, r7, pc}
+  %shl = shl nsw i64 -1, %y
+  %and = and i64 %shl, %x
+  ret i64 %and
+}