[PPC PWR10] Bit extraction opts to use expensive vector code

[Validark]

Godbolt

This code:

export fn extract_bits1(a: u64) u64 {
    return ((a >> 2) & 0b1) | ((a >> 5) & 0b110);
}

This LLVM IR:

define dso_local i64 @extract_bits1(i64 %0) #0 {
1:
  %2 = zext i6 2 to i64
  %3 = lshr i64 %0, %2
  %4 = and i64 %3, 1
  %5 = zext i6 5 to i64
  %6 = lshr i64 %0, %5
  %7 = and i64 %6, 6
  %8 = or i64 %4, %7
  ret i64 %8
}

Gets "optimized" to this LLVM IR when -mcpu=pwr10:

define dso_local i64 @extract_bits1(i64 %0) local_unnamed_addr {
Entry:
  %1 = insertelement <2 x i64> poison, i64 %0, i64 0
  %2 = shufflevector <2 x i64> %1, <2 x i64> poison, <2 x i32> zeroinitializer
  %3 = lshr <2 x i64> %2, <i64 2, i64 5>
  %4 = and <2 x i64> %3, <i64 1, i64 6>
  %shift = shufflevector <2 x i64> %4, <2 x i64> poison, <2 x i32> <i32 1, i32 poison>
  %5 = or <2 x i64> %4, %shift
  %6 = extractelement <2 x i64> %5, i64 0
  ret i64 %6
}

For pwr9, and other architecture/cpu combinations, this does not happen.

This results in this assembly:

.LCPI0_0:
        .long   0
        .long   2
        .long   0
        .long   5
.LCPI0_1:
        .long   0
        .long   1
        .long   0
        .long   6
extract_bits1:
        stwu 1, -64(1)
        stw 4, 32(1)
        stw 3, 16(1)
        xxsplti32dx 34, 0, 66051
        li 3, .LCPI0_0@l
        lis 4, .LCPI0_0@ha
        lxv 35, 32(1)
        lxv 36, 16(1)
        xxsplti32dx 34, 1, 269554195
        vperm 2, 4, 3, 2
        lxvx 35, 4, 3
        li 3, .LCPI0_1@l
        lis 4, .LCPI0_1@ha
        lxvx 0, 4, 3
        vsrd 2, 2, 3
        xxland 1, 34, 0
        xxswapd 35, 1
        xxeval 0, 35, 34, 0, 31
        stxv 0, 48(1)
        lwz 3, 48(1)
        lwz 4, 52(1)
        addi 1, 1, 64
        blr

Compare that to the pwr9 assembly, which at first glance seems a lot better:

extract_bits1:
        stwu 1, -16(1)
        rlwinm 5, 4, 30, 31, 31
        li 3, 0
        rlwimi 5, 4, 27, 29, 30
        mr      4, 5
        addi 1, 1, 16
        blr

Expected Behavior

I expect the optimized LLVM IR to be the same as it is for the pwr9 platform.

define dso_local i64 @extract_bits1(i64 %0) local_unnamed_addr {
Entry:
  %1 = lshr i64 %0, 2
  %2 = and i64 %1, 1
  %3 = lshr i64 %0, 5
  %4 = and i64 %3, 6
  %5 = or i64 %2, %4
  ret i64 %5
}

Originally ziglang/zig#18381

[llvmbot]

@llvm/issue-subscribers-backend-powerpc

Author: Niles Salter (Validark)

[Godbolt](https://zig.godbolt.org/z/M4decT6sc)

This code:

export fn extract_bits1(a: u64) u64 {
    return ((a >> 2) & 0b1) | ((a >> 5) & 0b110);
}

This LLVM IR:

define dso_local i64 @<!-- -->extract_bits1(i64 %0) #<!-- -->0 {
1:
  %2 = zext i6 2 to i64
  %3 = lshr i64 %0, %2
  %4 = and i64 %3, 1
  %5 = zext i6 5 to i64
  %6 = lshr i64 %0, %5
  %7 = and i64 %6, 6
  %8 = or i64 %4, %7
  ret i64 %8
}

Gets "optimized" to this LLVM IR when -mcpu=pwr10:

define dso_local i64 @<!-- -->extract_bits1(i64 %0) local_unnamed_addr {
Entry:
  %1 = insertelement &lt;2 x i64&gt; poison, i64 %0, i64 0
  %2 = shufflevector &lt;2 x i64&gt; %1, &lt;2 x i64&gt; poison, &lt;2 x i32&gt; zeroinitializer
  %3 = lshr &lt;2 x i64&gt; %2, &lt;i64 2, i64 5&gt;
  %4 = and &lt;2 x i64&gt; %3, &lt;i64 1, i64 6&gt;
  %shift = shufflevector &lt;2 x i64&gt; %4, &lt;2 x i64&gt; poison, &lt;2 x i32&gt; &lt;i32 1, i32 poison&gt;
  %5 = or &lt;2 x i64&gt; %4, %shift
  %6 = extractelement &lt;2 x i64&gt; %5, i64 0
  ret i64 %6
}

For pwr9, and other architecture/cpu combinations, this does not happen.

This results in this assembly:

.LCPI0_0:
        .long   0
        .long   2
        .long   0
        .long   5
.LCPI0_1:
        .long   0
        .long   1
        .long   0
        .long   6
extract_bits1:
        stwu 1, -64(1)
        stw 4, 32(1)
        stw 3, 16(1)
        xxsplti32dx 34, 0, 66051
        li 3, .LCPI0_0@<!-- -->l
        lis 4, .LCPI0_0@<!-- -->ha
        lxv 35, 32(1)
        lxv 36, 16(1)
        xxsplti32dx 34, 1, 269554195
        vperm 2, 4, 3, 2
        lxvx 35, 4, 3
        li 3, .LCPI0_1@<!-- -->l
        lis 4, .LCPI0_1@<!-- -->ha
        lxvx 0, 4, 3
        vsrd 2, 2, 3
        xxland 1, 34, 0
        xxswapd 35, 1
        xxeval 0, 35, 34, 0, 31
        stxv 0, 48(1)
        lwz 3, 48(1)
        lwz 4, 52(1)
        addi 1, 1, 64
        blr

Compare that to the pwr9 assembly, which at first glance seems a lot better:

extract_bits1:
        stwu 1, -16(1)
        rlwinm 5, 4, 30, 31, 31
        li 3, 0
        rlwimi 5, 4, 27, 29, 30
        mr      4, 5
        addi 1, 1, 16
        blr

Expected Behavior

I expect the optimized LLVM IR to be the same as it is for the pwr9 platform.

define dso_local i64 @<!-- -->extract_bits1(i64 %0) local_unnamed_addr {
Entry:
  %1 = lshr i64 %0, 2
  %2 = and i64 %1, 1
  %3 = lshr i64 %0, 5
  %4 = and i64 %3, 6
  %5 = or i64 %2, %4
  ret i64 %5
}

Originally ziglang/zig#18381