From 4e0008dcbe9fce99b9727e8bbeb129efc7bf2d80 Mon Sep 17 00:00:00 2001
From: Nathan Chancellor <nathan@kernel.org>
Date: Tue, 22 Mar 2022 17:32:33 -0700
Subject: [PATCH] Revert "[InstCombine] try to narrow shifted bswap-of-zext"

This reverts commit 9e9bda2e8f5b88715bad767a4b7740df32b040d2.

This causes a backend error when building the Linux kernel for arm64.
See https://reviews.llvm.org/D122166 for a simplified reproducer.
---
 .../InstCombine/InstCombineShifts.cpp         | 15 ---------
 llvm/test/Transforms/InstCombine/lshr.ll      | 31 +++++++++----------
 2 files changed, 15 insertions(+), 31 deletions(-)

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp b/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp
index 403979b4a6a3..03214118a2cf 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp
@@ -1173,21 +1173,6 @@ Instruction *InstCombinerImpl::visitLShr(BinaryOperator &I) {
         MulC->logBase2() == ShAmtC)
       return BinaryOperator::CreateAnd(X, ConstantInt::get(Ty, *MulC - 2));
 
-    // Try to narrow a bswap:
-    // (bswap (zext X)) >> C --> zext (bswap X >> C')
-    // In the case where the shift amount equals the bitwidth difference, the
-    // shift is eliminated.
-    if (match(Op0, m_OneUse(m_Intrinsic<Intrinsic::bswap>(
-                       m_OneUse(m_ZExt(m_Value(X))))))) {
-      // TODO: If the shift amount is less than the zext, we could shift left.
-      unsigned WidthDiff = BitWidth - X->getType()->getScalarSizeInBits();
-      if (ShAmtC >= WidthDiff) {
-        Value *NarrowSwap = Builder.CreateUnaryIntrinsic(Intrinsic::bswap, X);
-        Value *NewShift = Builder.CreateLShr(NarrowSwap, ShAmtC - WidthDiff);
-        return new ZExtInst(NewShift, Ty);
-      }
-    }
-
     // If the shifted-out value is known-zero, then this is an exact shift.
     if (!I.isExact() &&
         MaskedValueIsZero(Op0, APInt::getLowBitsSet(BitWidth, ShAmtC), 0, &I)) {
diff --git a/llvm/test/Transforms/InstCombine/lshr.ll b/llvm/test/Transforms/InstCombine/lshr.ll
index a46f3f460f5e..175430189ce1 100644
--- a/llvm/test/Transforms/InstCombine/lshr.ll
+++ b/llvm/test/Transforms/InstCombine/lshr.ll
@@ -831,8 +831,9 @@ define i1 @icmp_sge(i32 %x, i32 %y) {
 
 define i32 @narrow_bswap(i16 %x) {
 ; CHECK-LABEL: @narrow_bswap(
-; CHECK-NEXT:    [[TMP1:%.*]] = call i16 @llvm.bswap.i16(i16 [[X:%.*]])
-; CHECK-NEXT:    [[S:%.*]] = zext i16 [[TMP1]] to i32
+; CHECK-NEXT:    [[Z:%.*]] = zext i16 [[X:%.*]] to i32
+; CHECK-NEXT:    [[B:%.*]] = call i32 @llvm.bswap.i32(i32 [[Z]])
+; CHECK-NEXT:    [[S:%.*]] = lshr exact i32 [[B]], 16
 ; CHECK-NEXT:    ret i32 [[S]]
 ;
   %z = zext i16 %x to i32
@@ -843,8 +844,9 @@ define i32 @narrow_bswap(i16 %x) {
 
 define i128 @narrow_bswap_extra_wide(i16 %x) {
 ; CHECK-LABEL: @narrow_bswap_extra_wide(
-; CHECK-NEXT:    [[TMP1:%.*]] = call i16 @llvm.bswap.i16(i16 [[X:%.*]])
-; CHECK-NEXT:    [[S:%.*]] = zext i16 [[TMP1]] to i128
+; CHECK-NEXT:    [[Z:%.*]] = zext i16 [[X:%.*]] to i128
+; CHECK-NEXT:    [[B:%.*]] = call i128 @llvm.bswap.i128(i128 [[Z]])
+; CHECK-NEXT:    [[S:%.*]] = lshr exact i128 [[B]], 112
 ; CHECK-NEXT:    ret i128 [[S]]
 ;
   %z = zext i16 %x to i128
@@ -853,8 +855,6 @@ define i128 @narrow_bswap_extra_wide(i16 %x) {
   ret i128 %s
 }
 
-; TODO: The bswap can be narrowed followed by shl.
-
 define i32 @narrow_bswap_undershift(i16 %x) {
 ; CHECK-LABEL: @narrow_bswap_undershift(
 ; CHECK-NEXT:    [[Z:%.*]] = zext i16 [[X:%.*]] to i32
@@ -870,8 +870,9 @@ define i32 @narrow_bswap_undershift(i16 %x) {
 
 define <2 x i64> @narrow_bswap_splat(<2 x i16> %x) {
 ; CHECK-LABEL: @narrow_bswap_splat(
-; CHECK-NEXT:    [[TMP1:%.*]] = call <2 x i16> @llvm.bswap.v2i16(<2 x i16> [[X:%.*]])
-; CHECK-NEXT:    [[S:%.*]] = zext <2 x i16> [[TMP1]] to <2 x i64>
+; CHECK-NEXT:    [[Z:%.*]] = zext <2 x i16> [[X:%.*]] to <2 x i64>
+; CHECK-NEXT:    [[B:%.*]] = call <2 x i64> @llvm.bswap.v2i64(<2 x i64> [[Z]])
+; CHECK-NEXT:    [[S:%.*]] = lshr exact <2 x i64> [[B]], <i64 48, i64 48>
 ; CHECK-NEXT:    ret <2 x i64> [[S]]
 ;
   %z = zext <2 x i16> %x to <2 x i64>
@@ -880,8 +881,6 @@ define <2 x i64> @narrow_bswap_splat(<2 x i16> %x) {
   ret <2 x i64> %s
 }
 
-; TODO: poison/undef in the shift amount is ok to propagate.
-
 define <2 x i64> @narrow_bswap_splat_poison_elt(<2 x i16> %x) {
 ; CHECK-LABEL: @narrow_bswap_splat_poison_elt(
 ; CHECK-NEXT:    [[Z:%.*]] = zext <2 x i16> [[X:%.*]] to <2 x i64>
@@ -897,9 +896,9 @@ define <2 x i64> @narrow_bswap_splat_poison_elt(<2 x i16> %x) {
 
 define <2 x i64> @narrow_bswap_overshift(<2 x i32> %x) {
 ; CHECK-LABEL: @narrow_bswap_overshift(
-; CHECK-NEXT:    [[TMP1:%.*]] = call <2 x i32> @llvm.bswap.v2i32(<2 x i32> [[X:%.*]])
-; CHECK-NEXT:    [[TMP2:%.*]] = lshr <2 x i32> [[TMP1]], <i32 16, i32 16>
-; CHECK-NEXT:    [[S:%.*]] = zext <2 x i32> [[TMP2]] to <2 x i64>
+; CHECK-NEXT:    [[Z:%.*]] = zext <2 x i32> [[X:%.*]] to <2 x i64>
+; CHECK-NEXT:    [[B:%.*]] = call <2 x i64> @llvm.bswap.v2i64(<2 x i64> [[Z]])
+; CHECK-NEXT:    [[S:%.*]] = lshr <2 x i64> [[B]], <i64 48, i64 48>
 ; CHECK-NEXT:    ret <2 x i64> [[S]]
 ;
   %z = zext <2 x i32> %x to <2 x i64>
@@ -910,9 +909,9 @@ define <2 x i64> @narrow_bswap_overshift(<2 x i32> %x) {
 
 define i128 @narrow_bswap_overshift2(i96 %x) {
 ; CHECK-LABEL: @narrow_bswap_overshift2(
-; CHECK-NEXT:    [[TMP1:%.*]] = call i96 @llvm.bswap.i96(i96 [[X:%.*]])
-; CHECK-NEXT:    [[TMP2:%.*]] = lshr i96 [[TMP1]], 29
-; CHECK-NEXT:    [[S:%.*]] = zext i96 [[TMP2]] to i128
+; CHECK-NEXT:    [[Z:%.*]] = zext i96 [[X:%.*]] to i128
+; CHECK-NEXT:    [[B:%.*]] = call i128 @llvm.bswap.i128(i128 [[Z]])
+; CHECK-NEXT:    [[S:%.*]] = lshr i128 [[B]], 61
 ; CHECK-NEXT:    ret i128 [[S]]
 ;
   %z = zext i96 %x to i128