[x86] make 8-bit shl undesirable

[spatel]

Summary

I was looking at a potential DAGCombiner fix for 1 of the regressions in D60278, and it caused severe regression test pain because x86 TLI lies about the desirability of 8-bit shift ops.

We've hinted at making all 8-bit ops undesirable for the reason in the code comment:

// TODO: Almost no 8-bit ops are desirable because they have no actual
//       size/speed advantages vs. 32-bit ops, but they do have a major
//       potential disadvantage by causing partial register stalls.

...but that leads to massive diffs and exposes all kinds of optimization holes itself. So I'm trying for 1 tiny opcode step at a time. :)

[andreadb]

Looks good to me.

[RKSimon]

Little steps..... LGTM (@andreadb any comments?)

[craig.topper]

llvm/test/CodeGen/X86/scheduler-backtracking.ll

19

We should get this down to addl %esi, %esi.

llvm/lib/Target/X86/X86ISelLowering.cpp

 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
   if (!isTypeLegal(VT))
     return false;
 
   // There are no vXi8 shifts.
   if (Opc == ISD::SHL && VT.isVector() && VT.getVectorElementType() == MVT::i8)
     return false;
 
-  // 8-bit multiply is probably not much cheaper than 32-bit multiply, and
-  // we have specializations to turn 32-bit multiply into LEA or other ops.
+  // TODO: Almost no 8-bit ops are desirable because they have no actual
+  //       size/speed advantages vs. 32-bit ops, but they do have a major
+  //       potential disadvantage by causing partial register stalls.
+  //
+  // 8-bit multiply/shl is probably not cheaper than 32-bit multiply/shl, and
+  // we have specializations to turn 32-bit multiply/shl into LEA or other ops.
   // Also, see the comment in "IsDesirableToPromoteOp" - where we additionally
   // check for a constant operand to the multiply.
-  if (Opc == ISD::MUL && VT == MVT::i8)
+  if ((Opc == ISD::MUL || Opc == ISD::SHL) && VT == MVT::i8)
     return false;
 
   // i16 instruction encodings are longer and some i16 instructions are slow,
   // so those are not desirable.
   if (VT == MVT::i16) {
     switch (Opc) {
     default:
       break;

llvm/test/CodeGen/X86/bt.ll

; <label>:16:
   ret void
 }
 
 ; Make sure we can simplify bt when the shift amount has known zeros in it
 ; which cause the and mask to have bits removed.
 define zeroext i1 @demanded_with_known_zeroes(i32 %bit, i32 %bits) {
 ; X86-LABEL: demanded_with_known_zeroes:
 ; X86:       # %bb.0: # %entry
-; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
-; X86-NEXT:    movb {{[0-9]+}}(%esp), %cl
-; X86-NEXT:    shlb $2, %cl
-; X86-NEXT:    movzbl %cl, %ecx
-; X86-NEXT:    btl %ecx, %eax
+; X86-NEXT:    movb {{[0-9]+}}(%esp), %al
+; X86-NEXT:    shlb $2, %al
+; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X86-NEXT:    movzbl %al, %eax
+; X86-NEXT:    btl %eax, %ecx
 ; X86-NEXT:    setb %al
 ; X86-NEXT:    retl
 ;
 ; X64-LABEL: demanded_with_known_zeroes:
 ; X64:       # %bb.0: # %entry
-; X64-NEXT:    shlb $2, %dil
-; X64-NEXT:    movzbl %dil, %eax
-; X64-NEXT:    btl %eax, %esi
+; X64-NEXT:    shll $2, %edi
+; X64-NEXT:    btl %edi, %esi
 ; X64-NEXT:    setb %al
 ; X64-NEXT:    retq
 entry:
   %bit2 = shl i32 %bit, 2
   %and = and i32 %bit2, 31
   %shl = shl i32 1, %and
   %and1 = and i32 %shl, %bits
   %tobool = icmp ne i32 %and1, 0
   ret i1 %tobool
 }

llvm/test/CodeGen/X86/btc_bts_btr.ll

; X86-NEXT:    retl
   store i64 %3, i64* %x
   ret void
 }
 
 define i32 @btr_32_mask_zeros(i32 %x, i32 %n) {
 ; X64-LABEL: btr_32_mask_zeros:
 ; X64:       # %bb.0:
 ; X64-NEXT:    movl %edi, %eax
-; X64-NEXT:    shlb $2, %sil
+; X64-NEXT:    shll $2, %esi
 ; X64-NEXT:    btrl %esi, %eax
 ; X64-NEXT:    retq
 ;
 ; X86-LABEL: btr_32_mask_zeros:
 ; X86:       # %bb.0:
-; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
 ; X86-NEXT:    movb {{[0-9]+}}(%esp), %cl
 ; X86-NEXT:    shlb $2, %cl
+; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
 ; X86-NEXT:    btrl %ecx, %eax
 ; X86-NEXT:    retl
   %1 = shl i32 %n, 2
   %2 = and i32 %1, 31
   %3 = shl i32 1, %2
   %4 = xor i32 %3, -1
   %5 = and i32 %x, %4
   ret i32 %5
 }
 
 define i32 @bts_32_mask_zeros(i32 %x, i32 %n) {
 ; X64-LABEL: bts_32_mask_zeros:
 ; X64:       # %bb.0:
 ; X64-NEXT:    movl %edi, %eax
-; X64-NEXT:    shlb $2, %sil
+; X64-NEXT:    shll $2, %esi
 ; X64-NEXT:    btsl %esi, %eax
 ; X64-NEXT:    retq
 ;
 ; X86-LABEL: bts_32_mask_zeros:
 ; X86:       # %bb.0:
-; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
 ; X86-NEXT:    movb {{[0-9]+}}(%esp), %cl
 ; X86-NEXT:    shlb $2, %cl
+; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
 ; X86-NEXT:    btsl %ecx, %eax
 ; X86-NEXT:    retl
   %1 = shl i32 %n, 2
   %2 = and i32 %1, 31
   %3 = shl i32 1, %2
   %4 = or i32 %x, %3
   ret i32 %4
 }
 
 define i32 @btc_32_mask_zeros(i32 %x, i32 %n) {
 ; X64-LABEL: btc_32_mask_zeros:
 ; X64:       # %bb.0:
 ; X64-NEXT:    movl %edi, %eax
-; X64-NEXT:    shlb $2, %sil
+; X64-NEXT:    shll $2, %esi
 ; X64-NEXT:    btcl %esi, %eax
 ; X64-NEXT:    retq
 ;
 ; X86-LABEL: btc_32_mask_zeros:
 ; X86:       # %bb.0:
-; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
 ; X86-NEXT:    movb {{[0-9]+}}(%esp), %cl
 ; X86-NEXT:    shlb $2, %cl
+; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
 ; X86-NEXT:    btcl %ecx, %eax
 ; X86-NEXT:    retl
   %1 = shl i32 %n, 2
   %2 = and i32 %1, 31
   %3 = shl i32 1, %2
   %4 = xor i32 %x, %3
   ret i32 %4
 }
 
 define i64 @btr_64_mask_zeros(i64 %x, i64 %n) {
 ; X64-LABEL: btr_64_mask_zeros:
 ; X64:       # %bb.0:
 ; X64-NEXT:    movq %rdi, %rax
-; X64-NEXT:    shlb $2, %sil
+; X64-NEXT:    shlq $2, %rsi
 ; X64-NEXT:    btrq %rsi, %rax
 ; X64-NEXT:    retq
 ;
 ; X86-LABEL: btr_64_mask_zeros:
 ; X86:       # %bb.0:
-; X86-NEXT:    movb {{[0-9]+}}(%esp), %cl
-; X86-NEXT:    shlb $2, %cl
+; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X86-NEXT:    shll $2, %ecx
 ; X86-NEXT:    movl $1, %eax
 ; X86-NEXT:    xorl %edx, %edx
 ; X86-NEXT:    shldl %cl, %eax, %edx
 ; X86-NEXT:    shll %cl, %eax
 ; X86-NEXT:    testb $32, %cl
 ; X86-NEXT:    je .LBB39_2
 ; X86-NEXT:  # %bb.1:
 ; X86-NEXT:    movl %eax, %edx
; X86-NEXT:    retl
   %5 = and i64 %x, %4
   ret i64 %5
 }
 
 define i64 @bts_64_mask_zeros(i64 %x, i64 %n) {
 ; X64-LABEL: bts_64_mask_zeros:
 ; X64:       # %bb.0:
 ; X64-NEXT:    movq %rdi, %rax
-; X64-NEXT:    shlb $2, %sil
+; X64-NEXT:    shlq $2, %rsi
 ; X64-NEXT:    btsq %rsi, %rax
 ; X64-NEXT:    retq
 ;
 ; X86-LABEL: bts_64_mask_zeros:
 ; X86:       # %bb.0:
-; X86-NEXT:    movb {{[0-9]+}}(%esp), %cl
-; X86-NEXT:    shlb $2, %cl
+; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X86-NEXT:    shll $2, %ecx
 ; X86-NEXT:    movl $1, %eax
 ; X86-NEXT:    xorl %edx, %edx
 ; X86-NEXT:    shldl %cl, %eax, %edx
 ; X86-NEXT:    shll %cl, %eax
 ; X86-NEXT:    testb $32, %cl
 ; X86-NEXT:    je .LBB40_2
 ; X86-NEXT:  # %bb.1:
 ; X86-NEXT:    movl %eax, %edx
 ; X86-NEXT:    xorl %eax, %eax
 ; X86-NEXT:  .LBB40_2:
 ; X86-NEXT:    orl {{[0-9]+}}(%esp), %edx
 ; X86-NEXT:    orl {{[0-9]+}}(%esp), %eax
 ; X86-NEXT:    retl
   %1 = shl i64 %n, 2
   %2 = and i64 %1, 63
   %3 = shl i64 1, %2
   %4 = or i64 %x, %3
   ret i64 %4
 }
 
 define i64 @btc_64_mask_zeros(i64 %x, i64 %n) {
 ; X64-LABEL: btc_64_mask_zeros:
 ; X64:       # %bb.0:
 ; X64-NEXT:    movq %rdi, %rax
-; X64-NEXT:    shlb $2, %sil
+; X64-NEXT:    shlq $2, %rsi
 ; X64-NEXT:    btcq %rsi, %rax
 ; X64-NEXT:    retq
 ;
 ; X86-LABEL: btc_64_mask_zeros:
 ; X86:       # %bb.0:
-; X86-NEXT:    movb {{[0-9]+}}(%esp), %cl
-; X86-NEXT:    shlb $2, %cl
+; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X86-NEXT:    shll $2, %ecx
 ; X86-NEXT:    movl $1, %eax
 ; X86-NEXT:    xorl %edx, %edx
 ; X86-NEXT:    shldl %cl, %eax, %edx
 ; X86-NEXT:    shll %cl, %eax
 ; X86-NEXT:    testb $32, %cl
 ; X86-NEXT:    je .LBB41_2
 ; X86-NEXT:  # %bb.1:
 ; X86-NEXT:    movl %eax, %edx

llvm/test/CodeGen/X86/rotate4.ll

; X64-NEXT:    retq
   %7 = lshr i32 %0, %6
   %8 = or i32 %7, %4
   ret i32 %8
 }
 
 define i32 @rotate_demanded_bits_3(i32, i32) {
 ; X86-LABEL: rotate_demanded_bits_3:
 ; X86:       # %bb.0:
-; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
 ; X86-NEXT:    movb {{[0-9]+}}(%esp), %cl
 ; X86-NEXT:    addb %cl, %cl
+; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
 ; X86-NEXT:    roll %cl, %eax
 ; X86-NEXT:    retl
 ;
 ; X64-LABEL: rotate_demanded_bits_3:
 ; X64:       # %bb.0:
 ; X64-NEXT:    # kill: def $esi killed $esi def $rsi
 ; X64-NEXT:    movl %edi, %eax
 ; X64-NEXT:    leal (%rsi,%rsi), %ecx

llvm/test/CodeGen/X86/scheduler-backtracking.ll

 
 define i256 @test1(i256 %a) nounwind {
 ; ILP-LABEL: test1:
 ; ILP:       # %bb.0:
 ; ILP-NEXT:    pushq %r14
 ; ILP-NEXT:    pushq %rbx
 ; ILP-NEXT:    movq %rdi, %rax
 ; ILP-NEXT:    xorl %r8d, %r8d
-; ILP-NEXT:    addb %sil, %sil
+; ILP-NEXT:    addq %rsi, %rsi
 ; ILP-NEXT:    addb $2, %sil
 ; ILP-NEXT:    orb $1, %sil
 ; ILP-NEXT:    movl $1, %r10d
 ; ILP-NEXT:    xorl %r14d, %r14d
 ; ILP-NEXT:    movl %esi, %ecx
 ; ILP-NEXT:    shldq %cl, %r10, %r14
 ; ILP-NEXT:    movl $1, %edx
 ; ILP-NEXT:    shlq %cl, %rdx

 ; ILP-NEXT:    movq %rbx, 16(%rax)
 ; ILP-NEXT:    popq %rbx
 ; ILP-NEXT:    popq %r14
 ; ILP-NEXT:    retq
 ;
 ; HYBRID-LABEL: test1:
 ; HYBRID:       # %bb.0:
 ; HYBRID-NEXT:    movq %rdi, %rax
-; HYBRID-NEXT:    addb %sil, %sil
+; HYBRID-NEXT:    addq %rsi, %rsi
 ; HYBRID-NEXT:    addb $2, %sil
 ; HYBRID-NEXT:    orb $1, %sil
 ; HYBRID-NEXT:    movb $-128, %cl
 ; HYBRID-NEXT:    subb %sil, %cl
 ; HYBRID-NEXT:    xorl %r8d, %r8d
 ; HYBRID-NEXT:    movl $1, %r11d
 ; HYBRID-NEXT:    movl $1, %r9d
 ; HYBRID-NEXT:    shrdq %cl, %r8, %r9

 ; HYBRID-NEXT:    cmovnsq %r9, %rdx
 ; HYBRID-NEXT:    cmoveq %r8, %rdx
 ; HYBRID-NEXT:    movq %rdx, 16(%rax)
 ; HYBRID-NEXT:    retq
 ;
 ; BURR-LABEL: test1:
 ; BURR:       # %bb.0:
 ; BURR-NEXT:    movq %rdi, %rax
-; BURR-NEXT:    addb %sil, %sil
+; BURR-NEXT:    addq %rsi, %rsi
 ; BURR-NEXT:    addb $2, %sil
 ; BURR-NEXT:    orb $1, %sil
 ; BURR-NEXT:    movb $-128, %cl
 ; BURR-NEXT:    subb %sil, %cl
 ; BURR-NEXT:    xorl %r8d, %r8d
 ; BURR-NEXT:    movl $1, %r11d
 ; BURR-NEXT:    movl $1, %r9d
 ; BURR-NEXT:    shrdq %cl, %r8, %r9

 ; BURR-NEXT:    cmoveq %r8, %rdx
 ; BURR-NEXT:    movq %rdx, 16(%rax)
 ; BURR-NEXT:    retq
 ;
 ; SRC-LABEL: test1:
 ; SRC:       # %bb.0:
 ; SRC-NEXT:    pushq %rbx
 ; SRC-NEXT:    movq %rdi, %rax
-; SRC-NEXT:    addb %sil, %sil
+; SRC-NEXT:    addq %rsi, %rsi
 ; SRC-NEXT:    addb $2, %sil
 ; SRC-NEXT:    orb $1, %sil
 ; SRC-NEXT:    movb $-128, %cl
 ; SRC-NEXT:    subb %sil, %cl
 ; SRC-NEXT:    xorl %r8d, %r8d
 ; SRC-NEXT:    movl $1, %edi
 ; SRC-NEXT:    movl $1, %r10d
 ; SRC-NEXT:    shrdq %cl, %r8, %r10

 ; SRC-NEXT:    popq %rbx
 ; SRC-NEXT:    retq
 ;
 ; LIN-LABEL: test1:
 ; LIN:       # %bb.0:
 ; LIN-NEXT:    movq %rdi, %rax
 ; LIN-NEXT:    xorl %r9d, %r9d
 ; LIN-NEXT:    movl $1, %r8d
-; LIN-NEXT:    addb %sil, %sil
+; LIN-NEXT:    addq %rsi, %rsi
 ; LIN-NEXT:    addb $2, %sil
 ; LIN-NEXT:    orb $1, %sil
 ; LIN-NEXT:    movl $1, %edx
 ; LIN-NEXT:    movl %esi, %ecx
 ; LIN-NEXT:    shlq %cl, %rdx
 ; LIN-NEXT:    testb $64, %sil
 ; LIN-NEXT:    movq %rdx, %rcx
 ; LIN-NEXT:    cmovneq %r9, %rcx

llvm/test/CodeGen/X86/select_const.ll

 ; CHECK-NEXT:    retq
   %sel = select i1 %cond, i16 7, i16 71
   ret i16 %sel
 }
 
 define i32 @select_pow2_diff_neg(i1 zeroext %cond) {
 ; CHECK-LABEL: select_pow2_diff_neg:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    shlb $4, %dil
-; CHECK-NEXT:    movzbl %dil, %eax
-; CHECK-NEXT:    orl $-25, %eax
+; CHECK-NEXT:    # kill: def $edi killed $edi def $rdi
+; CHECK-NEXT:    shll $4, %edi
+; CHECK-NEXT:    leal -25(%rdi), %eax
 ; CHECK-NEXT:    retq
   %sel = select i1 %cond, i32 -9, i32 -25
   ret i32 %sel
 }
 
 define i64 @select_pow2_diff_neg_invert(i1 zeroext %cond) {
 ; CHECK-LABEL: select_pow2_diff_neg_invert:
 ; CHECK:       # %bb.0: