T0b1/wasmtime
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Infer REX prefixes for SIMD load_extend

Browse Source
This commit is contained in:
Andrew Brown
2020-04-01 09:57:14 -07:00
parent 9336884db5
commit dc874a5b3b
4 changed files with 40 additions and 29 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
cranelift/codegen/meta/src/isa/x86/encodings.rs
View File

@@ -1945,15 +1945,8 @@ fn define_simd(
for recipe in &[rec_fld, rec_fldDisp8, rec_fldDisp32] { for recipe in &[rec_fld, rec_fldDisp8, rec_fldDisp32] {
let inst = *inst; let inst = *inst;
let template = recipe.opcodes(*opcodes); let template = recipe.opcodes(*opcodes);
e.enc32_maybe_isap(inst.clone().bind(I32), template.clone(), isap); e.enc_both_inferred_maybe_isap(inst.clone().bind(I32), template.clone(), isap);
// REX-less encoding must come after REX encoding so we don't use it by e.enc64_maybe_isap(inst.bind(I64), template.infer_rex(), isap);
// default. Otherwise reg-alloc would never use r8 and up.
e.enc64_maybe_isap(inst.clone().bind(I32), template.clone().rex(), isap);
e.enc64_maybe_isap(inst.clone().bind(I32), template.clone(), isap);
// Similar to above; TODO some of this duplication can be cleaned up by infer_rex()
// tracked in https://github.com/bytecodealliance/cranelift/issues/1090
e.enc64_maybe_isap(inst.clone().bind(I64), template.clone().rex(), isap);
e.enc64_maybe_isap(inst.bind(I64), template, isap);
} }
} }

6
cranelift/codegen/meta/src/isa/x86/recipes.rs
View File

@@ -2087,7 +2087,7 @@ pub(crate) fn define<'shared>(
); );
// XX /r float load with 8-bit offset. // XX /r float load with 8-bit offset.
recipes.add_template_recipe( recipes.add_template_inferred(
EncodingRecipeBuilder::new("fldDisp8", &formats.load, 2) EncodingRecipeBuilder::new("fldDisp8", &formats.load, 2)
.operands_in(vec![gpr]) .operands_in(vec![gpr])
.operands_out(vec![fpr]) .operands_out(vec![fpr])
@@ -2110,6 +2110,7 @@ pub(crate) fn define<'shared>(
sink.put1(offset as u8); sink.put1(offset as u8);
"#, "#,
), ),
"size_plus_maybe_sib_for_inreg_0_plus_rex_prefix_for_inreg0_outreg0",
); );
let has_big_offset = let has_big_offset =
@@ -2142,7 +2143,7 @@ pub(crate) fn define<'shared>(
); );
// XX /r float load with 32-bit offset. // XX /r float load with 32-bit offset.
recipes.add_template_recipe( recipes.add_template_inferred(
EncodingRecipeBuilder::new("fldDisp32", &formats.load, 5) EncodingRecipeBuilder::new("fldDisp32", &formats.load, 5)
.operands_in(vec![gpr]) .operands_in(vec![gpr])
.operands_out(vec![fpr]) .operands_out(vec![fpr])
@@ -2165,6 +2166,7 @@ pub(crate) fn define<'shared>(
sink.put4(offset as u32); sink.put4(offset as u32);
"#, "#,
), ),
"size_plus_maybe_sib_for_inreg_0_plus_rex_prefix_for_inreg0_outreg0",
); );
} }

16
cranelift/codegen/src/isa/x86/enc_tables.rs
View File

@@ -156,6 +156,22 @@ fn size_plus_maybe_sib_or_offset_for_inreg_0_plus_rex_prefix_for_inreg0_outreg0(
+ if needs_rex { 1 } else { 0 } + if needs_rex { 1 } else { 0 }
} }
/// Calculates the size while inferring if the first input register (inreg0) and first output
/// register (outreg0) require a dynamic REX and if the first input register (inreg0) requires a
/// SIB.
fn size_plus_maybe_sib_for_inreg_0_plus_rex_prefix_for_inreg0_outreg0(
sizing: &RecipeSizing,
enc: Encoding,
inst: Inst,
divert: &RegDiversions,
func: &Function,
) -> u8 {
let needs_rex = (EncodingBits::from(enc.bits()).rex_w() != 0)
|| test_input(0, inst, divert, func, is_extended_reg)
|| test_result(0, inst, divert, func, is_extended_reg);
size_plus_maybe_sib_for_inreg_0(sizing, enc, inst, divert, func) + if needs_rex { 1 } else { 0 }
}
/// Infers whether a dynamic REX prefix will be emitted, for use with one input reg. /// Infers whether a dynamic REX prefix will be emitted, for use with one input reg.
/// ///
/// A REX prefix is known to be emitted if either: /// A REX prefix is known to be emitted if either:

36
cranelift/filetests/filetests/isa/x86/simd-memory-binemit.clif
View File

@@ -40,29 +40,29 @@ block0:
function %uload_extend() { function %uload_extend() {
block0: block0:
[-,%rdx] v1 = iconst.i64 0x0123_4567_89ab_cdef [-,%rdx] v1 = iconst.i64 0x0123_4567_89ab_cdef
[-,%xmm2] v3 = uload8x8 v1+0 ; bin: heap_oob 66 40 0f 38 30 12 [-,%xmm2] v3 = uload8x8 v1+0 ; bin: heap_oob 66 0f 38 30 12
[-,%xmm2] v4 = uload8x8 v1+20 ; bin: heap_oob 66 40 0f 38 30 52 14 [-,%xmm2] v4 = uload8x8 v1+20 ; bin: heap_oob 66 0f 38 30 52 14
[-,%xmm2] v5 = uload8x8 v1+256 ; bin: heap_oob 66 40 0f 38 30 92 00000100 [-,%xmm2] v5 = uload8x8 v1+256 ; bin: heap_oob 66 0f 38 30 92 00000100
[-,%xmm2] v6 = uload16x4 v1+0 ; bin: heap_oob 66 40 0f 38 33 12 [-,%xmm2] v6 = uload16x4 v1+0 ; bin: heap_oob 66 0f 38 33 12
[-,%xmm2] v7 = uload16x4 v1+20 ; bin: heap_oob 66 40 0f 38 33 52 14 [-,%xmm2] v7 = uload16x4 v1+20 ; bin: heap_oob 66 0f 38 33 52 14
[-,%xmm2] v8 = uload16x4 v1+256 ; bin: heap_oob 66 40 0f 38 33 92 00000100 [-,%xmm2] v8 = uload16x4 v1+256 ; bin: heap_oob 66 0f 38 33 92 00000100
[-,%xmm2] v9 = uload32x2 v1+0 ; bin: heap_oob 66 40 0f 38 35 12 [-,%xmm10] v9 = uload32x2 v1+0 ; bin: heap_oob 66 44 0f 38 35 12
[-,%xmm2] v10 = uload32x2 v1+20 ; bin: heap_oob 66 40 0f 38 35 52 14 [-,%xmm10] v10 = uload32x2 v1+20 ; bin: heap_oob 66 44 0f 38 35 52 14
[-,%xmm2] v11 = uload32x2 v1+256 ; bin: heap_oob 66 40 0f 38 35 92 00000100 [-,%xmm10] v11 = uload32x2 v1+256 ; bin: heap_oob 66 44 0f 38 35 92 00000100
return return
} }
function %sload_extend() { function %sload_extend() {
block0: block0:
[-,%rdx] v1 = iconst.i64 0x0123_4567_89ab_cdef [-,%rdx] v1 = iconst.i64 0x0123_4567_89ab_cdef
[-,%xmm2] v3 = sload8x8 v1+0 ; bin: heap_oob 66 40 0f 38 20 12 [-,%xmm2] v3 = sload8x8 v1+0 ; bin: heap_oob 66 0f 38 20 12
[-,%xmm2] v4 = sload8x8 v1+20 ; bin: heap_oob 66 40 0f 38 20 52 14 [-,%xmm2] v4 = sload8x8 v1+20 ; bin: heap_oob 66 0f 38 20 52 14
[-,%xmm2] v5 = sload8x8 v1+256 ; bin: heap_oob 66 40 0f 38 20 92 00000100 [-,%xmm2] v5 = sload8x8 v1+256 ; bin: heap_oob 66 0f 38 20 92 00000100
[-,%xmm2] v6 = sload16x4 v1+0 ; bin: heap_oob 66 40 0f 38 23 12 [-,%xmm10] v6 = sload16x4 v1+0 ; bin: heap_oob 66 44 0f 38 23 12
[-,%xmm2] v7 = sload16x4 v1+20 ; bin: heap_oob 66 40 0f 38 23 52 14 [-,%xmm10] v7 = sload16x4 v1+20 ; bin: heap_oob 66 44 0f 38 23 52 14
[-,%xmm2] v8 = sload16x4 v1+256 ; bin: heap_oob 66 40 0f 38 23 92 00000100 [-,%xmm10] v8 = sload16x4 v1+256 ; bin: heap_oob 66 44 0f 38 23 92 00000100
[-,%xmm2] v9 = sload32x2 v1+0 ; bin: heap_oob 66 40 0f 38 25 12 [-,%xmm2] v9 = sload32x2 v1+0 ; bin: heap_oob 66 0f 38 25 12
[-,%xmm2] v10 = sload32x2 v1+20 ; bin: heap_oob 66 40 0f 38 25 52 14 [-,%xmm2] v10 = sload32x2 v1+20 ; bin: heap_oob 66 0f 38 25 52 14
[-,%xmm2] v11 = sload32x2 v1+256 ; bin: heap_oob 66 40 0f 38 25 92 00000100 [-,%xmm2] v11 = sload32x2 v1+256 ; bin: heap_oob 66 0f 38 25 92 00000100
return return
} }