[LLVM][IR] Add textual shorthand for specifying constant vector splats. #74620
Conversation
|
@llvm/pr-subscribers-llvm-ir Author: Paul Walker (paulwalker-arm) ChangesAdd LL parsing for The intent is to extend ConstantInt/FP classes to support vector types rather than redirecting to other constant classes as the get() methods do today. This patch gives IR writers the convenience of using the shorthand today, thus allowing existing tests to be ported. Full diff: https://github.com/llvm/llvm-project/pull/74620.diff 5 Files Affected:
diff --git a/llvm/docs/LangRef.rst b/llvm/docs/LangRef.rst
index f5e8065ca1dc6..f7977488a54fd 100644
--- a/llvm/docs/LangRef.rst
+++ b/llvm/docs/LangRef.rst
@@ -4299,6 +4299,11 @@ constants and smaller complex constants.
"``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants
must have :ref:`vector type <t_vector>`, and the number and types of
elements must match those specified by the type.
+
+ When creating a vector whose elements have the same constant value, the
+ prefered syntax is ``splat (<Ty> Val)``. For example: "``splat (i32 11)``".
+ These vector constants must have ::ref:`vector type <t_vector>` with an
+ element type that matches the ``splat`` operand.
**Zero initialization**
The string '``zeroinitializer``' can be used to zero initialize a
value to zero of *any* type, including scalar and
diff --git a/llvm/include/llvm/AsmParser/LLToken.h b/llvm/include/llvm/AsmParser/LLToken.h
index 0aa0093e8efbd..147cf56c821aa 100644
--- a/llvm/include/llvm/AsmParser/LLToken.h
+++ b/llvm/include/llvm/AsmParser/LLToken.h
@@ -336,6 +336,7 @@ enum Kind {
kw_extractelement,
kw_insertelement,
kw_shufflevector,
+ kw_splat,
kw_extractvalue,
kw_insertvalue,
kw_blockaddress,
diff --git a/llvm/lib/AsmParser/LLLexer.cpp b/llvm/lib/AsmParser/LLLexer.cpp
index bf01b39e6f971..919c69fe2783e 100644
--- a/llvm/lib/AsmParser/LLLexer.cpp
+++ b/llvm/lib/AsmParser/LLLexer.cpp
@@ -698,6 +698,7 @@ lltok::Kind LLLexer::LexIdentifier() {
KEYWORD(uinc_wrap);
KEYWORD(udec_wrap);
+ KEYWORD(splat);
KEYWORD(vscale);
KEYWORD(x);
KEYWORD(blockaddress);
diff --git a/llvm/lib/AsmParser/LLParser.cpp b/llvm/lib/AsmParser/LLParser.cpp
index 5c7d551b2b31a..76c6694ae2b2a 100644
--- a/llvm/lib/AsmParser/LLParser.cpp
+++ b/llvm/lib/AsmParser/LLParser.cpp
@@ -3981,6 +3981,31 @@ bool LLParser::parseValID(ValID &ID, PerFunctionState *PFS, Type *ExpectedTy) {
return false;
}
+ case lltok::kw_splat: {
+ Lex.Lex();
+ if (parseToken(lltok::lparen, "expected '(' after vector splat"))
+ return true;
+ Constant *C;
+ if (parseGlobalTypeAndValue(C))
+ return true;
+ if (parseToken(lltok::rparen, "expected ')' at end of vector splat"))
+ return true;
+
+ if (auto *CI = dyn_cast<ConstantInt>(C)) {
+ ID.APSIntVal = CI->getValue();
+ ID.Kind = ValID::t_APSInt;
+ return false;
+ }
+
+ if (auto *CFP = dyn_cast<ConstantFP>(C)) {
+ ID.APFloatVal = CFP->getValue();
+ ID.Kind = ValID::t_APFloat;
+ return false;
+ }
+
+ return tokError("invalid splat operand");
+ }
+
case lltok::kw_getelementptr:
case lltok::kw_shufflevector:
case lltok::kw_insertelement:
@@ -5740,14 +5765,15 @@ bool LLParser::convertValIDToValue(Type *Ty, ValID &ID, Value *&V,
V = NoCFIValue::get(cast<GlobalValue>(V));
return V == nullptr;
case ValID::t_APSInt:
- if (!Ty->isIntegerTy())
+ if (!Ty->isIntOrIntVectorTy())
return error(ID.Loc, "integer constant must have integer type");
- ID.APSIntVal = ID.APSIntVal.extOrTrunc(Ty->getPrimitiveSizeInBits());
- V = ConstantInt::get(Context, ID.APSIntVal);
+ ID.APSIntVal = ID.APSIntVal.extOrTrunc(Ty->getScalarSizeInBits());
+ V = ConstantInt::get(Ty, ID.APSIntVal);
return false;
- case ValID::t_APFloat:
- if (!Ty->isFloatingPointTy() ||
- !ConstantFP::isValueValidForType(Ty, ID.APFloatVal))
+ case ValID::t_APFloat: {
+ Type *ScalarTy = Ty->getScalarType();
+ if (!ScalarTy->isFloatingPointTy() ||
+ !ConstantFP::isValueValidForType(ScalarTy, ID.APFloatVal))
return error(ID.Loc, "floating point constant invalid for type");
// The lexer has no type info, so builds all half, bfloat, float, and double
@@ -5756,13 +5782,13 @@ bool LLParser::convertValIDToValue(Type *Ty, ValID &ID, Value *&V,
// Check for signaling before potentially converting and losing that info.
bool IsSNAN = ID.APFloatVal.isSignaling();
bool Ignored;
- if (Ty->isHalfTy())
+ if (ScalarTy->isHalfTy())
ID.APFloatVal.convert(APFloat::IEEEhalf(), APFloat::rmNearestTiesToEven,
&Ignored);
- else if (Ty->isBFloatTy())
+ else if (ScalarTy->isBFloatTy())
ID.APFloatVal.convert(APFloat::BFloat(), APFloat::rmNearestTiesToEven,
&Ignored);
- else if (Ty->isFloatTy())
+ else if (ScalarTy->isFloatTy())
ID.APFloatVal.convert(APFloat::IEEEsingle(), APFloat::rmNearestTiesToEven,
&Ignored);
if (IsSNAN) {
@@ -5774,13 +5800,15 @@ bool LLParser::convertValIDToValue(Type *Ty, ValID &ID, Value *&V,
ID.APFloatVal.isNegative(), &Payload);
}
}
- V = ConstantFP::get(Context, ID.APFloatVal);
- if (V->getType() != Ty)
+ if (Type::getFloatingPointTy(Context, ID.APFloatVal.getSemantics()) !=
+ ScalarTy)
return error(ID.Loc, "floating point constant does not have type '" +
getTypeString(Ty) + "'");
+ V = ConstantFP::get(Ty, ID.APFloatVal);
return false;
+ }
case ValID::t_Null:
if (!Ty->isPointerTy())
return error(ID.Loc, "null must be a pointer type");
diff --git a/llvm/test/Assembler/constant-splat.ll b/llvm/test/Assembler/constant-splat.ll
new file mode 100644
index 0000000000000..25d83f4a1488c
--- /dev/null
+++ b/llvm/test/Assembler/constant-splat.ll
@@ -0,0 +1,57 @@
+; RUN: llvm-as < %s | llvm-dis | FileCheck %s
+
+; NOTE: Tests the expansion of the "splat" shorthand method to create vector
+; constants. Future work will change how "splat" is expanded, ultimately
+; leading to a point where "splat" is emitted as the disassembly.
+
+; CHECK: @constant.splat.i1 = constant <1 x i1> <i1 true>
+@constant.splat.i1 = constant <1 x i1> splat (i1 true)
+
+; CHECK: @constant.splat.i32 = constant <5 x i32> <i32 7, i32 7, i32 7, i32 7, i32 7>
+@constant.splat.i32 = constant <5 x i32> splat (i32 7)
+
+; CHECK: @constant.splat.i128 = constant <2 x i128> <i128 85070591730234615870450834276742070272, i128 85070591730234615870450834276742070272>
+@constant.splat.i128 = constant <2 x i128> splat (i128 85070591730234615870450834276742070272)
+
+; CHECK: @constant.splat.f16 = constant <4 x half> <half 0xHBC00, half 0xHBC00, half 0xHBC00, half 0xHBC00>
+@constant.splat.f16 = constant <4 x half> splat (half 0xHBC00)
+
+; CHECK: @constant.splat.f32 = constant <5 x float> <float -2.000000e+00, float -2.000000e+00, float -2.000000e+00, float -2.000000e+00, float -2.000000e+00>
+@constant.splat.f32 = constant <5 x float> splat (float -2.000000e+00)
+
+; CHECK: @constant.splat.f64 = constant <3 x double> <double -3.000000e+00, double -3.000000e+00, double -3.000000e+00>
+@constant.splat.f64 = constant <3 x double> splat (double -3.000000e+00)
+
+; CHECK: @constant.splat.128 = constant <2 x fp128> <fp128 0xL00000000000000018000000000000000, fp128 0xL00000000000000018000000000000000>
+@constant.splat.128 = constant <2 x fp128> splat (fp128 0xL00000000000000018000000000000000)
+
+; CHECK: @constant.splat.bf16 = constant <4 x bfloat> <bfloat 0xRC0A0, bfloat 0xRC0A0, bfloat 0xRC0A0, bfloat 0xRC0A0>
+@constant.splat.bf16 = constant <4 x bfloat> splat (bfloat 0xRC0A0)
+
+; CHECK: @constant.splat.x86_fp80 = constant <3 x x86_fp80> <x86_fp80 0xK4000C8F5C28F5C28F800, x86_fp80 0xK4000C8F5C28F5C28F800, x86_fp80 0xK4000C8F5C28F5C28F800>
+@constant.splat.x86_fp80 = constant <3 x x86_fp80> splat (x86_fp80 0xK4000C8F5C28F5C28F800)
+
+; CHECK: @constant.splat.ppc_fp128 = constant <1 x ppc_fp128> <ppc_fp128 0xM80000000000000000000000000000000>
+@constant.splat.ppc_fp128 = constant <1 x ppc_fp128> splat (ppc_fp128 0xM80000000000000000000000000000000)
+
+define void @add_fixed_lenth_vector_splat_i32(<4 x i32> %a) {
+; CHECK: %add = add <4 x i32> %a, <i32 137, i32 137, i32 137, i32 137>
+ %add = add <4 x i32> %a, splat (i32 137)
+ ret void
+}
+
+define <4 x i32> @ret_fixed_lenth_vector_splat_i32() {
+; CHECK: ret <4 x i32> <i32 56, i32 56, i32 56, i32 56>
+ ret <4 x i32> splat (i32 56)
+}
+
+define void @add_fixed_lenth_vector_splat_double(<vscale x 2 x double> %a) {
+; CHECK: %add = fadd <vscale x 2 x double> %a, shufflevector (<vscale x 2 x double> insertelement (<vscale x 2 x double> poison, double 5.700000e+00, i64 0), <vscale x 2 x double> poison, <vscale x 2 x i32> zeroinitializer)
+ %add = fadd <vscale x 2 x double> %a, splat (double 5.700000e+00)
+ ret void
+}
+
+define <vscale x 4 x i32> @ret_scalable_vector_splat_i32() {
+; CHECK: ret <vscale x 4 x i32> shufflevector (<vscale x 4 x i32> insertelement (<vscale x 4 x i32> poison, i32 78, i64 0), <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer)
+ ret <vscale x 4 x i32> splat (i32 78)
+}
|
llvm/docs/LangRef.rst
Outdated
| @@ -4299,6 +4299,11 @@ constants and smaller complex constants. | |||
| "``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants | |||
| must have :ref:`vector type <t_vector>`, and the number and types of | |||
| elements must match those specified by the type. | |||
|
|
|||
| When creating a vector whose elements have the same constant value, the | |||
| prefered syntax is ``splat (<Ty> Val)``. For example: "``splat (i32 11)``". | |||
There was a problem hiding this comment.
Am I being too adventurous here by suggesting this be the preferred syntax?
|
My main question here is whether this should be limited to immediates, or work with everything that Ultimately we will need some way to represent all splats we can currently represent -- I assume that the |
|
It depends how far down the rabbit hole we want to go. From a textual point of view the change looks easy but given the goal is to move away from For today perhaps it's best I extend the textual support to fallback to |
paulwalker-arm
force-pushed
the
llvm-ir-splat-macro
branch
from
836f7d3
to
fa97697
Compare
My general thought here was that this is necessary for "closure", so that the |
|
✅ With the latest revision this PR passed the C/C++ code formatter. |
Add LL parsing for `<N x ty> splat(ty <imm>)` that lowers onto ConstantInt::get() for integer types and ConstantFP::get() for floating-point types. The intent is to extend ConstantInt/FP classes to support vector types rather than redirecting to other constant classes as the get() methods do today. This patch gives IR writers the convenience of using the shorthand today, thus allowing existing tests to be ported.
paulwalker-arm
force-pushed
the
llvm-ir-splat-macro
branch
from
59bdd36
to
4e175dc
Compare
…. NFC A handy shorthand for specifying the shufflevector(insertelement(poison, foo, 0), poison, zeroinitializer) splat pattern was introduced in llvm#74620. Some of the RISC-V tests were converted over to use this new form in dbb65dd, this patch handles the rest. This converts some constant expressions to the new form, but it also converts instruction sequences that weren't previously constant expressions to constant expressions as well. In some cases this causes some minor code motion in the test output, but it's not significant to the test itself. E.g. see llvm/test/CodeGen/RISCV/rvv/fixed-vectors-masked-{gather,scatter}.ll
…#87616) A handy shorthand for specifying the shufflevector(insertelement(poison, foo, 0), poison, zeroinitializer) splat pattern was introduced in #74620. Some of the RISC-V tests were converted over to use this new form in dbb65dd, this patch handles the rest which didn't have any codegen diffs. This not only converts some constant expressions to the new form, but also instruction sequences that weren't previously constant expressions to constant expressions as well. In some cases this affects codegen, but these have been omitted here and will be handled in a separate PR.
Add LL parsing for
<N x ty> splat(ty <imm>)that lowers onto ConstantInt::get() for integer types and ConstantFP::get() for floating-point types.The intent is to extend ConstantInt/FP classes to support vector types rather than redirecting to other constant classes as the get() methods do today.
This patch gives IR writers the convenience of using the shorthand today, thus allowing existing tests to be ported.