From: Sanjay Patel <spatel@rotateright.com>
Date: Mon, 11 Apr 2022 11:09:47 +0000 (-0400)
Subject: [SDAG] try to reduce compare of funnel shift equal 0
X-Git-Tag: upstream/15.0.7~10842
X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=2ed15984b49a1af87be37ec8bd6ee3ab7f724767;p=platform%2Fupstream%2Fllvm.git

[SDAG] try to reduce compare of funnel shift equal 0

fshl (or X, Y), X, C ==/!= 0 --> or (shl Y, C), X ==/!= 0
fshl X, (or X, Y), C ==/!= 0 --> or (srl Y, BW-C), X ==/!= 0

This is similar to an existing setcc-of-rotate fold, but the
matching requires more checks for the more general funnel op:
https://alive2.llvm.org/ce/z/Ab2jDd

We are effectively decomposing the funnel shift into logical
shifts, reassociating, and removing a shift.

This should get us the final improvements for x86-64 that were
originally shown in D111530
( https://github.com/llvm/llvm-project/issues/49541 );
x86-32 still shows some SHLD/SHRD, so the pattern is not
matching there yet.

Differential Revision: https://reviews.llvm.org/D122919
---

diff --git a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
index 58f5ae4..d38a5a1 100644
--- a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
@@ -3875,6 +3875,72 @@ static SDValue foldSetCCWithRotate(EVT VT, SDValue N0, SDValue N1,
   return SDValue();
 }
 
+static SDValue foldSetCCWithFunnelShift(EVT VT, SDValue N0, SDValue N1,
+                                        ISD::CondCode Cond, const SDLoc &dl,
+                                        SelectionDAG &DAG) {
+  // If we are testing for all-bits-clear, we might be able to do that with
+  // less shifting since bit-order does not matter.
+  if (Cond != ISD::SETEQ && Cond != ISD::SETNE)
+    return SDValue();
+
+  auto *C1 = isConstOrConstSplat(N1, /* AllowUndefs */ true);
+  if (!C1 || !C1->isZero())
+    return SDValue();
+
+  if (!N0.hasOneUse() ||
+      (N0.getOpcode() != ISD::FSHL && N0.getOpcode() != ISD::FSHR))
+    return SDValue();
+
+  unsigned BitWidth = N0.getScalarValueSizeInBits();
+  auto *ShAmtC = isConstOrConstSplat(N0.getOperand(2));
+  if (!ShAmtC || ShAmtC->getAPIntValue().uge(BitWidth))
+    return SDValue();
+
+  // Canonicalize fshr as fshl to reduce pattern-matching.
+  unsigned ShAmt = ShAmtC->getZExtValue();
+  if (N0.getOpcode() == ISD::FSHR)
+    ShAmt = BitWidth - ShAmt;
+
+  // Match an 'or' with a specific operand 'Other' in either commuted variant.
+  SDValue X, Y;
+  auto matchOr = [&X, &Y](SDValue Or, SDValue Other) {
+    if (Or.getOpcode() != ISD::OR || !Or.hasOneUse())
+      return false;
+    if (Or.getOperand(0) == Other) {
+      X = Or.getOperand(0);
+      Y = Or.getOperand(1);
+      return true;
+    }
+    if (Or.getOperand(1) == Other) {
+      X = Or.getOperand(1);
+      Y = Or.getOperand(0);
+      return true;
+    }
+    return false;
+  };
+
+  EVT OpVT = N0.getValueType();
+  EVT ShAmtVT = N0.getOperand(2).getValueType();
+  SDValue F0 = N0.getOperand(0);
+  SDValue F1 = N0.getOperand(1);
+  if (matchOr(F0, F1)) {
+    // fshl (or X, Y), X, C ==/!= 0 --> or (shl Y, C), X ==/!= 0
+    SDValue NewShAmt = DAG.getConstant(ShAmt, dl, ShAmtVT);
+    SDValue Shift = DAG.getNode(ISD::SHL, dl, OpVT, Y, NewShAmt);
+    SDValue NewOr = DAG.getNode(ISD::OR, dl, OpVT, Shift, X);
+    return DAG.getSetCC(dl, VT, NewOr, N1, Cond);
+  }
+  if (matchOr(F1, F0)) {
+    // fshl X, (or X, Y), C ==/!= 0 --> or (srl Y, BW-C), X ==/!= 0
+    SDValue NewShAmt = DAG.getConstant(BitWidth - ShAmt, dl, ShAmtVT);
+    SDValue Shift = DAG.getNode(ISD::SRL, dl, OpVT, Y, NewShAmt);
+    SDValue NewOr = DAG.getNode(ISD::OR, dl, OpVT, Shift, X);
+    return DAG.getSetCC(dl, VT, NewOr, N1, Cond);
+  }
+
+  return SDValue();
+}
+
 /// Try to simplify a setcc built with the specified operands and cc. If it is
 /// unable to simplify it, return a null SDValue.
 SDValue TargetLowering::SimplifySetCC(EVT VT, SDValue N0, SDValue N1,
@@ -3914,6 +3980,9 @@ SDValue TargetLowering::SimplifySetCC(EVT VT, SDValue N0, SDValue N1,
   if (SDValue V = foldSetCCWithRotate(VT, N0, N1, Cond, dl, DAG))
     return V;
 
+  if (SDValue V = foldSetCCWithFunnelShift(VT, N0, N1, Cond, dl, DAG))
+    return V;
+
   if (auto *N1C = isConstOrConstSplat(N1)) {
     const APInt &C1 = N1C->getAPIntValue();
 
diff --git a/llvm/test/CodeGen/AArch64/setcc-fsh.ll b/llvm/test/CodeGen/AArch64/setcc-fsh.ll
index d70f0b4..aaf3a6c 100644
--- a/llvm/test/CodeGen/AArch64/setcc-fsh.ll
+++ b/llvm/test/CodeGen/AArch64/setcc-fsh.ll
@@ -9,8 +9,7 @@ declare <4 x i32> @llvm.fshl.v4i32(<4 x i32>, <4 x i32>, <4 x i32>)
 define i1 @fshl_or_eq_0(i32 %x, i32 %y) {
 ; CHECK-LABEL: fshl_or_eq_0:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ror w8, w0, #27
-; CHECK-NEXT:    orr w8, w8, w1, lsl #5
+; CHECK-NEXT:    orr w8, w0, w1, lsl #5
 ; CHECK-NEXT:    cmp w8, #0
 ; CHECK-NEXT:    cset w0, eq
 ; CHECK-NEXT:    ret
@@ -23,8 +22,7 @@ define i1 @fshl_or_eq_0(i32 %x, i32 %y) {
 define i1 @fshl_or_commute_eq_0(i32 %x, i32 %y) {
 ; CHECK-LABEL: fshl_or_commute_eq_0:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ror w8, w0, #27
-; CHECK-NEXT:    orr w8, w8, w1, lsl #5
+; CHECK-NEXT:    orr w8, w0, w1, lsl #5
 ; CHECK-NEXT:    cmp w8, #0
 ; CHECK-NEXT:    cset w0, eq
 ; CHECK-NEXT:    ret
@@ -37,10 +35,8 @@ define i1 @fshl_or_commute_eq_0(i32 %x, i32 %y) {
 define <4 x i1> @fshl_or2_eq_0(<4 x i32> %x, <4 x i32> %y) {
 ; CHECK-LABEL: fshl_or2_eq_0:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    orr v1.16b, v0.16b, v1.16b
-; CHECK-NEXT:    shl v0.4s, v0.4s, #25
 ; CHECK-NEXT:    ushr v1.4s, v1.4s, #7
-; CHECK-NEXT:    orr v0.16b, v0.16b, v1.16b
+; CHECK-NEXT:    orr v0.16b, v1.16b, v0.16b
 ; CHECK-NEXT:    cmeq v0.4s, v0.4s, #0
 ; CHECK-NEXT:    xtn v0.4h, v0.4s
 ; CHECK-NEXT:    ret
@@ -53,10 +49,8 @@ define <4 x i1> @fshl_or2_eq_0(<4 x i32> %x, <4 x i32> %y) {
 define <4 x i1> @fshl_or2_commute_eq_0(<4 x i32> %x, <4 x i32> %y) {
 ; CHECK-LABEL: fshl_or2_commute_eq_0:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    orr v1.16b, v1.16b, v0.16b
-; CHECK-NEXT:    shl v0.4s, v0.4s, #25
 ; CHECK-NEXT:    ushr v1.4s, v1.4s, #7
-; CHECK-NEXT:    orr v0.16b, v0.16b, v1.16b
+; CHECK-NEXT:    orr v0.16b, v1.16b, v0.16b
 ; CHECK-NEXT:    cmeq v0.4s, v0.4s, #0
 ; CHECK-NEXT:    xtn v0.4h, v0.4s
 ; CHECK-NEXT:    ret
@@ -69,9 +63,7 @@ define <4 x i1> @fshl_or2_commute_eq_0(<4 x i32> %x, <4 x i32> %y) {
 define i1 @fshr_or_eq_0(i16 %x, i16 %y) {
 ; CHECK-LABEL: fshr_or_eq_0:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    lsl w8, w0, #16
-; CHECK-NEXT:    orr w9, w0, w1
-; CHECK-NEXT:    extr w8, w9, w8, #24
+; CHECK-NEXT:    orr w8, w0, w1, lsl #8
 ; CHECK-NEXT:    tst w8, #0xffff
 ; CHECK-NEXT:    cset w0, eq
 ; CHECK-NEXT:    ret
@@ -84,9 +76,7 @@ define i1 @fshr_or_eq_0(i16 %x, i16 %y) {
 define i1 @fshr_or_commute_eq_0(i16 %x, i16 %y) {
 ; CHECK-LABEL: fshr_or_commute_eq_0:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    lsl w8, w0, #16
-; CHECK-NEXT:    orr w9, w1, w0
-; CHECK-NEXT:    extr w8, w9, w8, #24
+; CHECK-NEXT:    orr w8, w0, w1, lsl #8
 ; CHECK-NEXT:    tst w8, #0xffff
 ; CHECK-NEXT:    cset w0, eq
 ; CHECK-NEXT:    ret
@@ -99,8 +89,7 @@ define i1 @fshr_or_commute_eq_0(i16 %x, i16 %y) {
 define i1 @fshr_or2_eq_0(i64 %x, i64 %y) {
 ; CHECK-LABEL: fshr_or2_eq_0:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ror x8, x0, #3
-; CHECK-NEXT:    orr x8, x8, x1, lsr #3
+; CHECK-NEXT:    orr x8, x0, x1, lsr #3
 ; CHECK-NEXT:    cmp x8, #0
 ; CHECK-NEXT:    cset w0, eq
 ; CHECK-NEXT:    ret
@@ -113,8 +102,7 @@ define i1 @fshr_or2_eq_0(i64 %x, i64 %y) {
 define i1 @fshl_or_ne_0(i32 %x, i32 %y) {
 ; CHECK-LABEL: fshl_or_ne_0:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ror w8, w0, #25
-; CHECK-NEXT:    orr w8, w8, w1, lsl #7
+; CHECK-NEXT:    orr w8, w0, w1, lsl #7
 ; CHECK-NEXT:    cmp w8, #0
 ; CHECK-NEXT:    cset w0, ne
 ; CHECK-NEXT:    ret
@@ -127,8 +115,7 @@ define i1 @fshl_or_ne_0(i32 %x, i32 %y) {
 define i1 @fshl_or_commute_ne_0(i32 %x, i32 %y) {
 ; CHECK-LABEL: fshl_or_commute_ne_0:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ror w8, w0, #25
-; CHECK-NEXT:    orr w8, w8, w1, lsl #7
+; CHECK-NEXT:    orr w8, w0, w1, lsl #7
 ; CHECK-NEXT:    cmp w8, #0
 ; CHECK-NEXT:    cset w0, ne
 ; CHECK-NEXT:    ret
@@ -141,10 +128,8 @@ define i1 @fshl_or_commute_ne_0(i32 %x, i32 %y) {
 define <4 x i1> @fshl_or2_ne_0(<4 x i32> %x, <4 x i32> %y) {
 ; CHECK-LABEL: fshl_or2_ne_0:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    orr v1.16b, v0.16b, v1.16b
-; CHECK-NEXT:    shl v0.4s, v0.4s, #5
 ; CHECK-NEXT:    ushr v1.4s, v1.4s, #27
-; CHECK-NEXT:    orr v0.16b, v0.16b, v1.16b
+; CHECK-NEXT:    orr v0.16b, v1.16b, v0.16b
 ; CHECK-NEXT:    cmtst v0.4s, v0.4s, v0.4s
 ; CHECK-NEXT:    xtn v0.4h, v0.4s
 ; CHECK-NEXT:    ret
@@ -157,10 +142,8 @@ define <4 x i1> @fshl_or2_ne_0(<4 x i32> %x, <4 x i32> %y) {
 define <4 x i1> @fshl_or2_commute_ne_0(<4 x i32> %x, <4 x i32> %y) {
 ; CHECK-LABEL: fshl_or2_commute_ne_0:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    orr v1.16b, v1.16b, v0.16b
-; CHECK-NEXT:    shl v0.4s, v0.4s, #5
 ; CHECK-NEXT:    ushr v1.4s, v1.4s, #27
-; CHECK-NEXT:    orr v0.16b, v0.16b, v1.16b
+; CHECK-NEXT:    orr v0.16b, v1.16b, v0.16b
 ; CHECK-NEXT:    cmtst v0.4s, v0.4s, v0.4s
 ; CHECK-NEXT:    xtn v0.4h, v0.4s
 ; CHECK-NEXT:    ret
@@ -173,8 +156,7 @@ define <4 x i1> @fshl_or2_commute_ne_0(<4 x i32> %x, <4 x i32> %y) {
 define i1 @fshr_or_ne_0(i64 %x, i64 %y) {
 ; CHECK-LABEL: fshr_or_ne_0:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    orr w8, w0, w1
-; CHECK-NEXT:    extr x8, x8, x0, #1
+; CHECK-NEXT:    orr x8, x0, x1, lsl #63
 ; CHECK-NEXT:    cmp x8, #0
 ; CHECK-NEXT:    cset w0, ne
 ; CHECK-NEXT:    ret
@@ -187,8 +169,7 @@ define i1 @fshr_or_ne_0(i64 %x, i64 %y) {
 define i1 @fshr_or_commute_ne_0(i64 %x, i64 %y) {
 ; CHECK-LABEL: fshr_or_commute_ne_0:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    orr w8, w1, w0
-; CHECK-NEXT:    extr x8, x8, x0, #1
+; CHECK-NEXT:    orr x8, x0, x1, lsl #63
 ; CHECK-NEXT:    cmp x8, #0
 ; CHECK-NEXT:    cset w0, ne
 ; CHECK-NEXT:    ret
@@ -201,9 +182,8 @@ define i1 @fshr_or_commute_ne_0(i64 %x, i64 %y) {
 define i1 @fshr_or2_ne_0(i16 %x, i16 %y) {
 ; CHECK-LABEL: fshr_or2_ne_0:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    orr w8, w0, w1
-; CHECK-NEXT:    lsl w8, w8, #16
-; CHECK-NEXT:    extr w8, w0, w8, #18
+; CHECK-NEXT:    and w8, w1, #0xfffc
+; CHECK-NEXT:    orr w8, w0, w8, lsr #2
 ; CHECK-NEXT:    tst w8, #0xffff
 ; CHECK-NEXT:    cset w0, ne
 ; CHECK-NEXT:    ret
@@ -216,9 +196,8 @@ define i1 @fshr_or2_ne_0(i16 %x, i16 %y) {
 define i1 @fshr_or2_commute_ne_0(i16 %x, i16 %y) {
 ; CHECK-LABEL: fshr_or2_commute_ne_0:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    orr w8, w1, w0
-; CHECK-NEXT:    lsl w8, w8, #16
-; CHECK-NEXT:    extr w8, w0, w8, #18
+; CHECK-NEXT:    and w8, w1, #0xfffc
+; CHECK-NEXT:    orr w8, w0, w8, lsr #2
 ; CHECK-NEXT:    tst w8, #0xffff
 ; CHECK-NEXT:    cset w0, ne
 ; CHECK-NEXT:    ret
diff --git a/llvm/test/CodeGen/X86/icmp-shift-opt.ll b/llvm/test/CodeGen/X86/icmp-shift-opt.ll
index 0afe8d3..23524d2 100644
--- a/llvm/test/CodeGen/X86/icmp-shift-opt.ll
+++ b/llvm/test/CodeGen/X86/icmp-shift-opt.ll
@@ -53,8 +53,8 @@ define i128 @opt_setcc_lt_power_of_2(i128 %a) nounwind {
 ; X64-NEXT:    addq $1, %rax
 ; X64-NEXT:    adcq $0, %rdx
 ; X64-NEXT:    movq %rax, %rcx
+; X64-NEXT:    shrq $60, %rcx
 ; X64-NEXT:    orq %rdx, %rcx
-; X64-NEXT:    shrdq $60, %rdx, %rcx
 ; X64-NEXT:    jne .LBB0_1
 ; X64-NEXT:  # %bb.2: # %exit
 ; X64-NEXT:    retq
@@ -90,8 +90,8 @@ define i1 @opt_setcc_srl_eq_zero(i128 %a) nounwind {
 ;
 ; X64-LABEL: opt_setcc_srl_eq_zero:
 ; X64:       # %bb.0:
+; X64-NEXT:    shrq $17, %rdi
 ; X64-NEXT:    orq %rsi, %rdi
-; X64-NEXT:    shrdq $17, %rsi, %rdi
 ; X64-NEXT:    sete %al
 ; X64-NEXT:    retq
    %srl = lshr i128 %a, 17
@@ -119,8 +119,8 @@ define i1 @opt_setcc_srl_ne_zero(i128 %a) nounwind {
 ;
 ; X64-LABEL: opt_setcc_srl_ne_zero:
 ; X64:       # %bb.0:
+; X64-NEXT:    shrq $17, %rdi
 ; X64-NEXT:    orq %rsi, %rdi
-; X64-NEXT:    shrdq $17, %rsi, %rdi
 ; X64-NEXT:    setne %al
 ; X64-NEXT:    retq
    %srl = lshr i128 %a, 17
@@ -148,8 +148,8 @@ define i1 @opt_setcc_shl_eq_zero(i128 %a) nounwind {
 ;
 ; X64-LABEL: opt_setcc_shl_eq_zero:
 ; X64:       # %bb.0:
+; X64-NEXT:    shlq $17, %rsi
 ; X64-NEXT:    orq %rdi, %rsi
-; X64-NEXT:    shldq $17, %rdi, %rsi
 ; X64-NEXT:    sete %al
 ; X64-NEXT:    retq
    %shl = shl i128 %a, 17
@@ -177,8 +177,8 @@ define i1 @opt_setcc_shl_ne_zero(i128 %a) nounwind {
 ;
 ; X64-LABEL: opt_setcc_shl_ne_zero:
 ; X64:       # %bb.0:
+; X64-NEXT:    shlq $17, %rsi
 ; X64-NEXT:    orq %rdi, %rsi
-; X64-NEXT:    shldq $17, %rdi, %rsi
 ; X64-NEXT:    setne %al
 ; X64-NEXT:    retq
    %shl = shl i128 %a, 17
@@ -255,8 +255,8 @@ define i1 @opt_setcc_expanded_shl_correct_shifts(i64 %a, i64 %b) nounwind {
 ;
 ; X64-LABEL: opt_setcc_expanded_shl_correct_shifts:
 ; X64:       # %bb.0:
+; X64-NEXT:    shlq $17, %rdi
 ; X64-NEXT:    orq %rsi, %rdi
-; X64-NEXT:    shldq $17, %rsi, %rdi
 ; X64-NEXT:    sete %al
 ; X64-NEXT:    retq
   %shl.a = shl i64 %a, 17
diff --git a/llvm/test/CodeGen/X86/setcc-fsh.ll b/llvm/test/CodeGen/X86/setcc-fsh.ll
index 59319a7..7ab6395 100644
--- a/llvm/test/CodeGen/X86/setcc-fsh.ll
+++ b/llvm/test/CodeGen/X86/setcc-fsh.ll
@@ -265,8 +265,8 @@ define <4 x i1> @or_rotl_ne_eq0(<4 x i32> %x, <4 x i32> %y) nounwind {
 define i1 @fshl_or_eq_0(i32 %x, i32 %y) {
 ; CHECK-LABEL: fshl_or_eq_0:
 ; CHECK:       # %bb.0:
+; CHECK-NEXT:    shll $5, %esi
 ; CHECK-NEXT:    orl %edi, %esi
-; CHECK-NEXT:    shldl $5, %edi, %esi
 ; CHECK-NEXT:    sete %al
 ; CHECK-NEXT:    retq
   %or = or i32 %x, %y
@@ -278,8 +278,8 @@ define i1 @fshl_or_eq_0(i32 %x, i32 %y) {
 define i1 @fshl_or_commute_eq_0(i32 %x, i32 %y) {
 ; CHECK-LABEL: fshl_or_commute_eq_0:
 ; CHECK:       # %bb.0:
+; CHECK-NEXT:    shll $5, %esi
 ; CHECK-NEXT:    orl %edi, %esi
-; CHECK-NEXT:    shldl $5, %edi, %esi
 ; CHECK-NEXT:    sete %al
 ; CHECK-NEXT:    retq
   %or = or i32 %y, %x
@@ -291,12 +291,10 @@ define i1 @fshl_or_commute_eq_0(i32 %x, i32 %y) {
 define <4 x i1> @fshl_or2_eq_0(<4 x i32> %x, <4 x i32> %y) {
 ; CHECK-LABEL: fshl_or2_eq_0:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    por %xmm0, %xmm1
+; CHECK-NEXT:    pxor %xmm2, %xmm2
 ; CHECK-NEXT:    psrld $7, %xmm1
-; CHECK-NEXT:    pslld $25, %xmm0
 ; CHECK-NEXT:    por %xmm1, %xmm0
-; CHECK-NEXT:    pxor %xmm1, %xmm1
-; CHECK-NEXT:    pcmpeqd %xmm1, %xmm0
+; CHECK-NEXT:    pcmpeqd %xmm2, %xmm0
 ; CHECK-NEXT:    retq
   %or = or <4 x i32> %x, %y
   %f = call <4 x i32> @llvm.fshl.v4i32(<4 x i32> %x, <4 x i32> %or, <4 x i32> <i32 25, i32 25, i32 25, i32 25>)
@@ -307,12 +305,10 @@ define <4 x i1> @fshl_or2_eq_0(<4 x i32> %x, <4 x i32> %y) {
 define <4 x i1> @fshl_or2_commute_eq_0(<4 x i32> %x, <4 x i32> %y) {
 ; CHECK-LABEL: fshl_or2_commute_eq_0:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    por %xmm0, %xmm1
+; CHECK-NEXT:    pxor %xmm2, %xmm2
 ; CHECK-NEXT:    psrld $7, %xmm1
-; CHECK-NEXT:    pslld $25, %xmm0
 ; CHECK-NEXT:    por %xmm1, %xmm0
-; CHECK-NEXT:    pxor %xmm1, %xmm1
-; CHECK-NEXT:    pcmpeqd %xmm1, %xmm0
+; CHECK-NEXT:    pcmpeqd %xmm2, %xmm0
 ; CHECK-NEXT:    retq
   %or = or <4 x i32> %y, %x
   %f = call <4 x i32> @llvm.fshl.v4i32(<4 x i32> %x, <4 x i32> %or, <4 x i32> <i32 25, i32 25, i32 25, i32 25>)
@@ -323,8 +319,8 @@ define <4 x i1> @fshl_or2_commute_eq_0(<4 x i32> %x, <4 x i32> %y) {
 define i1 @fshr_or_eq_0(i16 %x, i16 %y) {
 ; CHECK-LABEL: fshr_or_eq_0:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    orl %edi, %esi
-; CHECK-NEXT:    shldw $8, %di, %si
+; CHECK-NEXT:    shll $8, %esi
+; CHECK-NEXT:    orw %di, %si
 ; CHECK-NEXT:    sete %al
 ; CHECK-NEXT:    retq
   %or = or i16 %x, %y
@@ -336,8 +332,8 @@ define i1 @fshr_or_eq_0(i16 %x, i16 %y) {
 define i1 @fshr_or_commute_eq_0(i16 %x, i16 %y) {
 ; CHECK-LABEL: fshr_or_commute_eq_0:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    orl %edi, %esi
-; CHECK-NEXT:    shldw $8, %di, %si
+; CHECK-NEXT:    shll $8, %esi
+; CHECK-NEXT:    orw %di, %si
 ; CHECK-NEXT:    sete %al
 ; CHECK-NEXT:    retq
   %or = or i16 %y, %x
@@ -349,8 +345,8 @@ define i1 @fshr_or_commute_eq_0(i16 %x, i16 %y) {
 define i1 @fshr_or2_eq_0(i64 %x, i64 %y) {
 ; CHECK-LABEL: fshr_or2_eq_0:
 ; CHECK:       # %bb.0:
+; CHECK-NEXT:    shrq $3, %rsi
 ; CHECK-NEXT:    orq %rdi, %rsi
-; CHECK-NEXT:    shrdq $3, %rdi, %rsi
 ; CHECK-NEXT:    sete %al
 ; CHECK-NEXT:    retq
   %or = or i64 %x, %y
@@ -362,8 +358,8 @@ define i1 @fshr_or2_eq_0(i64 %x, i64 %y) {
 define i1 @fshr_or2_commute_eq_0(i64 %x, i64 %y) {
 ; CHECK-LABEL: fshr_or2_commute_eq_0:
 ; CHECK:       # %bb.0:
+; CHECK-NEXT:    shrq $3, %rsi
 ; CHECK-NEXT:    orq %rdi, %rsi
-; CHECK-NEXT:    shrdq $3, %rdi, %rsi
 ; CHECK-NEXT:    sete %al
 ; CHECK-NEXT:    retq
   %or = or i64 %y, %x
@@ -375,8 +371,8 @@ define i1 @fshr_or2_commute_eq_0(i64 %x, i64 %y) {
 define i1 @fshl_or_ne_0(i32 %x, i32 %y) {
 ; CHECK-LABEL: fshl_or_ne_0:
 ; CHECK:       # %bb.0:
+; CHECK-NEXT:    shll $7, %esi
 ; CHECK-NEXT:    orl %edi, %esi
-; CHECK-NEXT:    shldl $7, %edi, %esi
 ; CHECK-NEXT:    setne %al
 ; CHECK-NEXT:    retq
   %or = or i32 %x, %y
@@ -388,8 +384,8 @@ define i1 @fshl_or_ne_0(i32 %x, i32 %y) {
 define i1 @fshl_or_commute_ne_0(i32 %x, i32 %y) {
 ; CHECK-LABEL: fshl_or_commute_ne_0:
 ; CHECK:       # %bb.0:
+; CHECK-NEXT:    shll $7, %esi
 ; CHECK-NEXT:    orl %edi, %esi
-; CHECK-NEXT:    shldl $7, %edi, %esi
 ; CHECK-NEXT:    setne %al
 ; CHECK-NEXT:    retq
   %or = or i32 %y, %x
@@ -401,12 +397,10 @@ define i1 @fshl_or_commute_ne_0(i32 %x, i32 %y) {
 define <4 x i1> @fshl_or2_ne_0(<4 x i32> %x, <4 x i32> %y) {
 ; CHECK-LABEL: fshl_or2_ne_0:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    por %xmm0, %xmm1
+; CHECK-NEXT:    pxor %xmm2, %xmm2
 ; CHECK-NEXT:    psrld $27, %xmm1
-; CHECK-NEXT:    pslld $5, %xmm0
 ; CHECK-NEXT:    por %xmm1, %xmm0
-; CHECK-NEXT:    pxor %xmm1, %xmm1
-; CHECK-NEXT:    pcmpeqd %xmm1, %xmm0
+; CHECK-NEXT:    pcmpeqd %xmm2, %xmm0
 ; CHECK-NEXT:    pcmpeqd %xmm1, %xmm1
 ; CHECK-NEXT:    pxor %xmm1, %xmm0
 ; CHECK-NEXT:    retq
@@ -419,12 +413,10 @@ define <4 x i1> @fshl_or2_ne_0(<4 x i32> %x, <4 x i32> %y) {
 define <4 x i1> @fshl_or2_commute_ne_0(<4 x i32> %x, <4 x i32> %y) {
 ; CHECK-LABEL: fshl_or2_commute_ne_0:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    por %xmm0, %xmm1
+; CHECK-NEXT:    pxor %xmm2, %xmm2
 ; CHECK-NEXT:    psrld $27, %xmm1
-; CHECK-NEXT:    pslld $5, %xmm0
 ; CHECK-NEXT:    por %xmm1, %xmm0
-; CHECK-NEXT:    pxor %xmm1, %xmm1
-; CHECK-NEXT:    pcmpeqd %xmm1, %xmm0
+; CHECK-NEXT:    pcmpeqd %xmm2, %xmm0
 ; CHECK-NEXT:    pcmpeqd %xmm1, %xmm1
 ; CHECK-NEXT:    pxor %xmm1, %xmm0
 ; CHECK-NEXT:    retq
@@ -437,8 +429,8 @@ define <4 x i1> @fshl_or2_commute_ne_0(<4 x i32> %x, <4 x i32> %y) {
 define i1 @fshr_or_ne_0(i64 %x, i64 %y) {
 ; CHECK-LABEL: fshr_or_ne_0:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    orl %edi, %esi
-; CHECK-NEXT:    shldq $63, %rdi, %rsi
+; CHECK-NEXT:    shlq $63, %rsi
+; CHECK-NEXT:    orq %rdi, %rsi
 ; CHECK-NEXT:    setne %al
 ; CHECK-NEXT:    retq
   %or = or i64 %x, %y
@@ -450,8 +442,8 @@ define i1 @fshr_or_ne_0(i64 %x, i64 %y) {
 define i1 @fshr_or_commute_ne_0(i64 %x, i64 %y) {
 ; CHECK-LABEL: fshr_or_commute_ne_0:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    orl %edi, %esi
-; CHECK-NEXT:    shldq $63, %rdi, %rsi
+; CHECK-NEXT:    shlq $63, %rsi
+; CHECK-NEXT:    orq %rdi, %rsi
 ; CHECK-NEXT:    setne %al
 ; CHECK-NEXT:    retq
   %or = or i64 %y, %x
@@ -463,8 +455,9 @@ define i1 @fshr_or_commute_ne_0(i64 %x, i64 %y) {
 define i1 @fshr_or2_ne_0(i16 %x, i16 %y) {
 ; CHECK-LABEL: fshr_or2_ne_0:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    orl %edi, %esi
-; CHECK-NEXT:    shrdw $2, %di, %si
+; CHECK-NEXT:    movzwl %si, %eax
+; CHECK-NEXT:    shrl $2, %eax
+; CHECK-NEXT:    orw %di, %ax
 ; CHECK-NEXT:    setne %al
 ; CHECK-NEXT:    retq
   %or = or i16 %x, %y
@@ -476,8 +469,9 @@ define i1 @fshr_or2_ne_0(i16 %x, i16 %y) {
 define i1 @fshr_or2_commute_ne_0(i16 %x, i16 %y) {
 ; CHECK-LABEL: fshr_or2_commute_ne_0:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    orl %edi, %esi
-; CHECK-NEXT:    shrdw $2, %di, %si
+; CHECK-NEXT:    movzwl %si, %eax
+; CHECK-NEXT:    shrl $2, %eax
+; CHECK-NEXT:    orw %di, %ax
 ; CHECK-NEXT:    setne %al
 ; CHECK-NEXT:    retq
   %or = or i16 %y, %x