feat(asm): encode the arm64 SVE gather loads and scatter stores
Assisted-by: GLM 5.3
This commit is contained in:
petrbalvin committed
1 parent
6faf850793
commit
104bea036b
5 files changed
+1273
-2
No files matched your search
@@ -562,3 +562,212 @@ func TestArm64AssembleExtensionStage3Refusals(t *testing.T) {
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// TestArm64AssembleExtensionGatherGolden drives the gather/scatter family
|
||||
// through the full assembler: every corpus line of the scope the layer
|
||||
// covers, corpus text in, corpus word out. The scope is the plain,
|
||||
// sign-extended and first-fault loads and the stores over the core
|
||||
// addressing modes: the two-general-register pair, the vector base with a
|
||||
// general offset, the extended 32-bit offsets, the immediate offset on a
|
||||
// vector base and the lone vector base of the stores.
|
||||
func TestArm64AssembleExtensionGatherGolden(t *testing.T) {
|
||||
tests := []struct {
|
||||
stmt string
|
||||
want uint32
|
||||
}{
|
||||
{"ZLD1B (R6)(R14), P4.Z, [Z13.B]", 0xa40651cd},
|
||||
{"ZLD1B (R6)(R14), P4.Z, [Z13.H]", 0xa42651cd},
|
||||
{"ZLD1B (R6)(R14), P4.Z, [Z13.S]", 0xa44651cd},
|
||||
{"ZLD1B (R6)(R14), P4.Z, [Z13.D]", 0xa46651cd},
|
||||
{"ZLD1B (Z10.D)(R19), P3.Z, [Z15.D]", 0xc44ace6f},
|
||||
{"ZLD1B (Z4.S.UXTW)(R3), P3.Z, [Z4.S]", 0x84044c64},
|
||||
{"ZLD1B (Z6.D)(R14), P4.Z, [Z13.D]", 0xc446d1cd},
|
||||
{"ZLD1D (R6<<3)(R14), P4.Z, [Z13.D]", 0xa5e651cd},
|
||||
{"ZLD1D (R6<<3)(R14), P4.Z, [Z13.Q]", 0xa58691cd},
|
||||
{"ZLD1D (Z23.D<<3)(R24), P1.Z, [Z22.D]", 0xc5f7c716},
|
||||
{"ZLD1D (Z10.D)(R19), P3.Z, [Z15.D]", 0xc5cace6f},
|
||||
{"ZLD1D (Z6.D<<3)(R14), P4.Z, [Z13.D]", 0xc5e6d1cd},
|
||||
{"ZLD1D (Z6.D)(R14), P4.Z, [Z13.D]", 0xc5c6d1cd},
|
||||
{"ZLD1H (R6<<1)(R14), P4.Z, [Z13.H]", 0xa4a651cd},
|
||||
{"ZLD1H (R6<<1)(R14), P4.Z, [Z13.S]", 0xa4c651cd},
|
||||
{"ZLD1H (R6<<1)(R14), P4.Z, [Z13.D]", 0xa4e651cd},
|
||||
{"ZLD1H (Z4.S.UXTW<<1)(R3), P3.Z, [Z4.S]", 0x84a44c64},
|
||||
{"ZLD1H (Z23.D<<1)(R24), P1.Z, [Z22.D]", 0xc4f7c716},
|
||||
{"ZLD1H (Z10.D)(R19), P3.Z, [Z15.D]", 0xc4cace6f},
|
||||
{"ZLD1H (Z4.S.UXTW)(R3), P3.Z, [Z4.S]", 0x84844c64},
|
||||
{"ZLD1H (Z6.D<<1)(R14), P4.Z, [Z13.D]", 0xc4e6d1cd},
|
||||
{"ZLD1H (Z6.D)(R14), P4.Z, [Z13.D]", 0xc4c6d1cd},
|
||||
{"ZLD1SB (R6)(R14), P4.Z, [Z13.H]", 0xa5c651cd},
|
||||
{"ZLD1SB (R6)(R14), P4.Z, [Z13.S]", 0xa5a651cd},
|
||||
{"ZLD1SB (R6)(R14), P4.Z, [Z13.D]", 0xa58651cd},
|
||||
{"ZLD1SB (Z10.D)(R19), P3.Z, [Z15.D]", 0xc44a8e6f},
|
||||
{"ZLD1SB (Z4.S.UXTW)(R3), P3.Z, [Z4.S]", 0x84040c64},
|
||||
{"ZLD1SB (Z6.D)(R14), P4.Z, [Z13.D]", 0xc44691cd},
|
||||
{"ZLD1SH (R6<<1)(R14), P4.Z, [Z13.S]", 0xa52651cd},
|
||||
{"ZLD1SH (R6<<1)(R14), P4.Z, [Z13.D]", 0xa50651cd},
|
||||
{"ZLD1SH (Z4.S.UXTW<<1)(R3), P3.Z, [Z4.S]", 0x84a40c64},
|
||||
{"ZLD1SH (Z23.D<<1)(R24), P1.Z, [Z22.D]", 0xc4f78716},
|
||||
{"ZLD1SH (Z10.D)(R19), P3.Z, [Z15.D]", 0xc4ca8e6f},
|
||||
{"ZLD1SH (Z4.S.UXTW)(R3), P3.Z, [Z4.S]", 0x84840c64},
|
||||
{"ZLD1SH (Z6.D<<1)(R14), P4.Z, [Z13.D]", 0xc4e691cd},
|
||||
{"ZLD1SH (Z6.D)(R14), P4.Z, [Z13.D]", 0xc4c691cd},
|
||||
{"ZLD1SW (R6<<2)(R14), P4.Z, [Z13.D]", 0xa48651cd},
|
||||
{"ZLD1SW (Z23.D<<2)(R24), P1.Z, [Z22.D]", 0xc5778716},
|
||||
{"ZLD1SW (Z10.D)(R19), P3.Z, [Z15.D]", 0xc54a8e6f},
|
||||
{"ZLD1SW (Z6.D<<2)(R14), P4.Z, [Z13.D]", 0xc56691cd},
|
||||
{"ZLD1SW (Z6.D)(R14), P4.Z, [Z13.D]", 0xc54691cd},
|
||||
{"ZLD1W (R6<<2)(R14), P4.Z, [Z13.S]", 0xa54651cd},
|
||||
{"ZLD1W (R6<<2)(R14), P4.Z, [Z13.D]", 0xa56651cd},
|
||||
{"ZLD1W (R6<<2)(R14), P4.Z, [Z13.Q]", 0xa50691cd},
|
||||
{"ZLD1W (Z4.S.UXTW<<2)(R3), P3.Z, [Z4.S]", 0x85244c64},
|
||||
{"ZLD1W (Z23.D<<2)(R24), P1.Z, [Z22.D]", 0xc577c716},
|
||||
{"ZLD1W (Z10.D)(R19), P3.Z, [Z15.D]", 0xc54ace6f},
|
||||
{"ZLD1W (Z4.S.UXTW)(R3), P3.Z, [Z4.S]", 0x85044c64},
|
||||
{"ZLD1W (Z6.D<<2)(R14), P4.Z, [Z13.D]", 0xc566d1cd},
|
||||
{"ZLD1W (Z6.D)(R14), P4.Z, [Z13.D]", 0xc546d1cd},
|
||||
{"ZLDFF1B (R6)(R14), P4.Z, [Z13.B]", 0xa40671cd},
|
||||
{"ZLDFF1B (R6)(R14), P4.Z, [Z13.H]", 0xa42671cd},
|
||||
{"ZLDFF1B (R6)(R14), P4.Z, [Z13.S]", 0xa44671cd},
|
||||
{"ZLDFF1B (R6)(R14), P4.Z, [Z13.D]", 0xa46671cd},
|
||||
{"ZLDFF1B (Z10.D)(R19), P3.Z, [Z15.D]", 0xc44aee6f},
|
||||
{"ZLDFF1B (Z4.S.UXTW)(R3), P3.Z, [Z4.S]", 0x84046c64},
|
||||
{"ZLDFF1B (Z6.D)(R14), P4.Z, [Z13.D]", 0xc446f1cd},
|
||||
{"ZLDFF1D (R6<<3)(R14), P4.Z, [Z13.D]", 0xa5e671cd},
|
||||
{"ZLDFF1D (Z23.D<<3)(R24), P1.Z, [Z22.D]", 0xc5f7e716},
|
||||
{"ZLDFF1D (Z10.D)(R19), P3.Z, [Z15.D]", 0xc5caee6f},
|
||||
{"ZLDFF1D (Z6.D<<3)(R14), P4.Z, [Z13.D]", 0xc5e6f1cd},
|
||||
{"ZLDFF1D (Z6.D)(R14), P4.Z, [Z13.D]", 0xc5c6f1cd},
|
||||
{"ZLDFF1H (R6<<1)(R14), P4.Z, [Z13.H]", 0xa4a671cd},
|
||||
{"ZLDFF1H (R6<<1)(R14), P4.Z, [Z13.S]", 0xa4c671cd},
|
||||
{"ZLDFF1H (R6<<1)(R14), P4.Z, [Z13.D]", 0xa4e671cd},
|
||||
{"ZLDFF1H (Z4.S.UXTW<<1)(R3), P3.Z, [Z4.S]", 0x84a46c64},
|
||||
{"ZLDFF1H (Z23.D<<1)(R24), P1.Z, [Z22.D]", 0xc4f7e716},
|
||||
{"ZLDFF1H (Z10.D)(R19), P3.Z, [Z15.D]", 0xc4caee6f},
|
||||
{"ZLDFF1H (Z4.S.UXTW)(R3), P3.Z, [Z4.S]", 0x84846c64},
|
||||
{"ZLDFF1H (Z6.D<<1)(R14), P4.Z, [Z13.D]", 0xc4e6f1cd},
|
||||
{"ZLDFF1H (Z6.D)(R14), P4.Z, [Z13.D]", 0xc4c6f1cd},
|
||||
{"ZLDFF1SB (R6)(R14), P4.Z, [Z13.H]", 0xa5c671cd},
|
||||
{"ZLDFF1SB (R6)(R14), P4.Z, [Z13.S]", 0xa5a671cd},
|
||||
{"ZLDFF1SB (R6)(R14), P4.Z, [Z13.D]", 0xa58671cd},
|
||||
{"ZLDFF1SB (Z10.D)(R19), P3.Z, [Z15.D]", 0xc44aae6f},
|
||||
{"ZLDFF1SB (Z4.S.UXTW)(R3), P3.Z, [Z4.S]", 0x84042c64},
|
||||
{"ZLDFF1SB (Z6.D)(R14), P4.Z, [Z13.D]", 0xc446b1cd},
|
||||
{"ZLDFF1SH (R6<<1)(R14), P4.Z, [Z13.S]", 0xa52671cd},
|
||||
{"ZLDFF1SH (R6<<1)(R14), P4.Z, [Z13.D]", 0xa50671cd},
|
||||
{"ZLDFF1SH (Z4.S.UXTW<<1)(R3), P3.Z, [Z4.S]", 0x84a42c64},
|
||||
{"ZLDFF1SH (Z23.D<<1)(R24), P1.Z, [Z22.D]", 0xc4f7a716},
|
||||
{"ZLDFF1SH (Z10.D)(R19), P3.Z, [Z15.D]", 0xc4caae6f},
|
||||
{"ZLDFF1SH (Z4.S.UXTW)(R3), P3.Z, [Z4.S]", 0x84842c64},
|
||||
{"ZLDFF1SH (Z6.D<<1)(R14), P4.Z, [Z13.D]", 0xc4e6b1cd},
|
||||
{"ZLDFF1SH (Z6.D)(R14), P4.Z, [Z13.D]", 0xc4c6b1cd},
|
||||
{"ZLDFF1SW (R6<<2)(R14), P4.Z, [Z13.D]", 0xa48671cd},
|
||||
{"ZLDFF1SW (Z23.D<<2)(R24), P1.Z, [Z22.D]", 0xc577a716},
|
||||
{"ZLDFF1SW (Z10.D)(R19), P3.Z, [Z15.D]", 0xc54aae6f},
|
||||
{"ZLDFF1SW (Z6.D<<2)(R14), P4.Z, [Z13.D]", 0xc566b1cd},
|
||||
{"ZLDFF1SW (Z6.D)(R14), P4.Z, [Z13.D]", 0xc546b1cd},
|
||||
{"ZLDFF1W (R6<<2)(R14), P4.Z, [Z13.S]", 0xa54671cd},
|
||||
{"ZLDFF1W (R6<<2)(R14), P4.Z, [Z13.D]", 0xa56671cd},
|
||||
{"ZLDFF1W (Z4.S.UXTW<<2)(R3), P3.Z, [Z4.S]", 0x85246c64},
|
||||
{"ZLDFF1W (Z23.D<<2)(R24), P1.Z, [Z22.D]", 0xc577e716},
|
||||
{"ZLDFF1W (Z10.D)(R19), P3.Z, [Z15.D]", 0xc54aee6f},
|
||||
{"ZLDFF1W (Z4.S.UXTW)(R3), P3.Z, [Z4.S]", 0x85046c64},
|
||||
{"ZLDFF1W (Z6.D<<2)(R14), P4.Z, [Z13.D]", 0xc566f1cd},
|
||||
{"ZLDFF1W (Z6.D)(R14), P4.Z, [Z13.D]", 0xc546f1cd},
|
||||
{"ZST1B [Z7.D], P4, (R21)(R7)", 0xe47550e7},
|
||||
{"ZST1B [Z6.D], P4, (Z21.D.UXTW)(R7)", 0xe41590e6},
|
||||
{"ZST1B [Z6.S], P4, (Z21.S.UXTW)(R7)", 0xe45590e6},
|
||||
{"ZST1B [Z8.D], P3, (Z6.D)(RSP)", 0xe406afe8},
|
||||
{"ZST1D [Z8.D], P3, (R6<<3)(RSP)", 0xe5e64fe8},
|
||||
{"ZST1D [Z8.Q], P3, (R6<<3)(RSP)", 0xe5c64fe8},
|
||||
{"ZST1D [Z6.D], P4, (Z21.D.UXTW<<3)(R7)", 0xe5b590e6},
|
||||
{"ZST1D [Z6.D], P4, (Z21.D.UXTW)(R7)", 0xe59590e6},
|
||||
{"ZST1D [Z8.D], P3, (Z6.D<<3)(RSP)", 0xe5a6afe8},
|
||||
{"ZST1D [Z8.D], P3, (Z6.D)(RSP)", 0xe586afe8},
|
||||
{"ZST1H [Z7.D], P4, (R21<<1)(R7)", 0xe4f550e7},
|
||||
{"ZST1H [Z6.S], P4, (Z21.S.UXTW<<1)(R7)", 0xe4f590e6},
|
||||
{"ZST1H [Z6.D], P4, (Z21.D.UXTW<<1)(R7)", 0xe4b590e6},
|
||||
{"ZST1H [Z6.D], P4, (Z21.D.UXTW)(R7)", 0xe49590e6},
|
||||
{"ZST1H [Z6.S], P4, (Z21.S.UXTW)(R7)", 0xe4d590e6},
|
||||
{"ZST1H [Z8.D], P3, (Z6.D<<1)(RSP)", 0xe4a6afe8},
|
||||
{"ZST1H [Z8.D], P3, (Z6.D)(RSP)", 0xe486afe8},
|
||||
{"ZST1W [Z7.D], P4, (R21<<2)(R7)", 0xe57550e7},
|
||||
{"ZST1W [Z8.Q], P3, (R6<<2)(RSP)", 0xe5064fe8},
|
||||
{"ZST1W [Z6.S], P4, (Z21.S.UXTW<<2)(R7)", 0xe57590e6},
|
||||
{"ZST1W [Z6.D], P4, (Z21.D.UXTW<<2)(R7)", 0xe53590e6},
|
||||
{"ZST1W [Z6.D], P4, (Z21.D.UXTW)(R7)", 0xe51590e6},
|
||||
{"ZST1W [Z6.S], P4, (Z21.S.UXTW)(R7)", 0xe55590e6},
|
||||
{"ZST1W [Z8.D], P3, (Z6.D<<2)(RSP)", 0xe526afe8},
|
||||
{"ZST1W [Z8.D], P3, (Z6.D)(RSP)", 0xe506afe8},
|
||||
{"ZLD1B 6(Z7.S), P4.Z, [Z13.S]", 0x8426d0ed},
|
||||
{"ZLD1B 6(Z7.D), P4.Z, [Z13.D]", 0xc426d0ed},
|
||||
{"ZLD1D 8(Z14.D), P6.Z, [Z6.D]", 0xc5a1d9c6},
|
||||
{"ZLD1H 6(Z7.S), P4.Z, [Z13.S]", 0x84a3d0ed},
|
||||
{"ZLD1H 6(Z7.D), P4.Z, [Z13.D]", 0xc4a3d0ed},
|
||||
{"ZLD1SB 6(Z7.S), P4.Z, [Z13.S]", 0x842690ed},
|
||||
{"ZLD1SB 6(Z7.D), P4.Z, [Z13.D]", 0xc42690ed},
|
||||
{"ZLD1SH 6(Z7.S), P4.Z, [Z13.S]", 0x84a390ed},
|
||||
{"ZLD1SH 6(Z7.D), P4.Z, [Z13.D]", 0xc4a390ed},
|
||||
{"ZLD1SW 12(Z0.D), P6.Z, [Z25.D]", 0xc5239819},
|
||||
{"ZLD1W 12(Z0.S), P6.Z, [Z25.S]", 0x8523d819},
|
||||
{"ZLD1W 12(Z0.D), P6.Z, [Z25.D]", 0xc523d819},
|
||||
{"ZLDFF1B 6(Z7.S), P4.Z, [Z13.S]", 0x8426f0ed},
|
||||
{"ZLDFF1B 6(Z7.D), P4.Z, [Z13.D]", 0xc426f0ed},
|
||||
{"ZLDFF1D 8(Z14.D), P6.Z, [Z6.D]", 0xc5a1f9c6},
|
||||
{"ZLDFF1H 6(Z7.S), P4.Z, [Z13.S]", 0x84a3f0ed},
|
||||
{"ZLDFF1H 6(Z7.D), P4.Z, [Z13.D]", 0xc4a3f0ed},
|
||||
{"ZLDFF1SB 6(Z7.S), P4.Z, [Z13.S]", 0x8426b0ed},
|
||||
{"ZLDFF1SB 6(Z7.D), P4.Z, [Z13.D]", 0xc426b0ed},
|
||||
{"ZLDFF1SH 6(Z7.S), P4.Z, [Z13.S]", 0x84a3b0ed},
|
||||
{"ZLDFF1SH 6(Z7.D), P4.Z, [Z13.D]", 0xc4a3b0ed},
|
||||
{"ZLDFF1SW 12(Z0.D), P6.Z, [Z25.D]", 0xc523b819},
|
||||
{"ZLDFF1W 12(Z0.S), P6.Z, [Z25.S]", 0x8523f819},
|
||||
{"ZLDFF1W 12(Z0.D), P6.Z, [Z25.D]", 0xc523f819},
|
||||
{"ZST1B [Z8.S], P3, 6(Z15.S)", 0xe466ade8},
|
||||
{"ZST1B [Z8.D], P3, 6(Z15.D)", 0xe446ade8},
|
||||
{"ZST1D [Z16.D], P2, (Z5.D)", 0xe5c0a8b0},
|
||||
{"ZST1H [Z8.S], P3, 6(Z15.S)", 0xe4e3ade8},
|
||||
{"ZST1H [Z8.D], P3, 6(Z15.D)", 0xe4c3ade8},
|
||||
{"ZST1W [Z16.S], P2, (Z5.S)", 0xe560a8b0},
|
||||
{"ZST1W [Z16.D], P2, (Z5.D)", 0xe540a8b0},
|
||||
}
|
||||
for _, tt := range tests {
|
||||
words := assembleArm64Words(t, arm64ExtProbeHead+"\t"+tt.stmt+"\n\tRET\n")
|
||||
if len(words) != 2 {
|
||||
t.Fatalf("%s: got %d words, want the statement and RET", tt.stmt, len(words))
|
||||
}
|
||||
if words[0] != tt.want {
|
||||
t.Errorf("%s:\n got %08x\n want %08x", tt.stmt, words[0], tt.want)
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// TestArm64AssembleExtensionGatherRefusals pins the diagnostics the
|
||||
// gather/scatter statements get from the layer: the addressing-mode
|
||||
// mismatches, the locked list arrangements, the zeroing qualifier of the
|
||||
// loads against the bare predicates of the stores, and the immediate
|
||||
// scaling.
|
||||
func TestArm64AssembleExtensionGatherRefusals(t *testing.T) {
|
||||
tests := []struct {
|
||||
stmt string
|
||||
want string
|
||||
}{
|
||||
{"ZLD1B (R6<<1)(R14), P4.Z, [Z13.B]", "scaled by <<0"},
|
||||
{"ZLD1B (Z6.D)(R14), P4.Z, [Z13.B]", "wants two general registers"},
|
||||
{"ZLD1B (Z6.S)(R14), P4.Z, [Z13.S]", "wants two general registers"},
|
||||
{"ZLD1B (R6)(R14), P4/M, [Z13.B]", "wants the zeroing qualifier"},
|
||||
{"ZLD1B (R6)(R14), P9.Z, [Z13.B]", "outside P0-P7"},
|
||||
{"ZLD1B (R6)(ZR), P4.Z, [Z13.B]", "is not an extended-layer operand"},
|
||||
{"ZLD1B (RSP)(R14), P4.Z, [Z13.B]", "is not an extended-layer operand"},
|
||||
{"ZLD1D 12(Z14.D), P6.Z, [Z6.D]", "wants two general registers"},
|
||||
{"ZLD1D 264(Z14.D), P6.Z, [Z6.D]", "wants two general registers"},
|
||||
{"ZST1B [Z7.D], P4/Z, (R21)(R7)", "takes no qualifier"},
|
||||
{"ZST1B [Z7.S], P4, (R21)(R7)", "carries arrangement .S, want .D"},
|
||||
{"ZST1B [Z7.D], P4, (Z21.S)(R7)", "wants two general registers"},
|
||||
}
|
||||
for _, tt := range tests {
|
||||
got := assembleArm64SrcError(t, arm64ExtProbeHead+"\t"+tt.stmt+"\n\tRET\n")
|
||||
if !strings.Contains(got, tt.want) {
|
||||
t.Errorf("%s: error %q does not name %q", tt.stmt, got, tt.want)
|
||||
}
|
||||
}
|
||||
}
|
||||
Reference in new issue
Block a user