[llvm] 10def77 - [AArch64][GlobalISel] Tighten up some legal types (#194785)
via llvm-commits
llvm-commits at lists.llvm.org
Tue Apr 28 22:18:37 PDT 2026
Author: David Green
Date: 2026-04-29T06:18:32+01:00
New Revision: 10def779b8785f79b006d1edc0a1adebd696d740
URL: https://github.com/llvm/llvm-project/commit/10def779b8785f79b006d1edc0a1adebd696d740
DIFF: https://github.com/llvm/llvm-project/commit/10def779b8785f79b006d1edc0a1adebd696d740.diff
LOG: [AArch64][GlobalISel] Tighten up some legal types (#194785)
This tightens up some of the legal types from scalar any types to the
correct
integer or floating point types. Some are still not changed, like trunc
and
zext/sext. Type independant operations like loads, stores, vector
operations,
selects etc all still correctly use scalar any types.
Added:
Modified:
llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp
Removed:
################################################################################
diff --git a/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp b/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp
index f47f651f3c581..e8bc74a150b57 100644
--- a/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp
+++ b/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp
@@ -98,6 +98,11 @@ AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST)
const LLT i128 = LLT::integer(128);
+ const LLT nxv16i8 = LLT::scalable_vector(16, i8);
+ const LLT nxv8i16 = LLT::scalable_vector(8, i16);
+ const LLT nxv4i32 = LLT::scalable_vector(4, i32);
+ const LLT nxv2i64 = LLT::scalable_vector(2, i64);
+
std::initializer_list<LLT> PackedVectorAllTypeList = {/* Begin 128bit types */
v16s8, v8s16, v4s32,
v2s64, v2p0,
@@ -177,7 +182,7 @@ AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST)
getActionDefinitionsBuilder({G_ADD, G_SUB, G_AND, G_OR, G_XOR})
.legalFor({i32, i64, v8i8, v16i8, v4i16, v8i16, v2i32, v4i32, v2i64})
- .legalFor(HasSVE, {nxv16s8, nxv8s16, nxv4s32, nxv2s64})
+ .legalFor(HasSVE, {nxv16i8, nxv8i16, nxv4i32, nxv2i64})
.widenScalarToNextPow2(0)
.clampScalar(0, s32, s64)
.clampMaxNumElements(0, s8, 16)
@@ -289,7 +294,7 @@ AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST)
.lower();
getActionDefinitionsBuilder({G_SMULH, G_UMULH})
- .legalFor({s64, v16s8, v8s16, v4s32})
+ .legalFor({i64, v16i8, v8i16, v4i32})
.lower();
getActionDefinitionsBuilder({G_SMIN, G_SMAX, G_UMIN, G_UMAX})
@@ -303,7 +308,7 @@ AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST)
// FIXME: Legal vector types are only legal with NEON.
getActionDefinitionsBuilder(G_ABS)
- .legalFor(HasCSSC, {s32, s64})
+ .legalFor(HasCSSC, {i32, i64})
.legalFor({v16i8, v8i16, v4i32, v2i64, v2p0, v8i8, v4i16, v2i32})
.customIf([=](const LegalityQuery &Q) {
// TODO: Fix suboptimal codegen for 128+ bit types.
@@ -325,12 +330,12 @@ AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST)
getActionDefinitionsBuilder(
{G_ABDS, G_ABDU, G_UAVGFLOOR, G_UAVGCEIL, G_SAVGFLOOR, G_SAVGCEIL})
- .legalFor({v8s8, v16s8, v4s16, v8s16, v2s32, v4s32})
+ .legalFor({v8i8, v16i8, v4i16, v8i16, v2i32, v4i32})
.lower();
getActionDefinitionsBuilder(
{G_SADDE, G_SSUBE, G_UADDE, G_USUBE, G_SADDO, G_SSUBO, G_UADDO, G_USUBO})
- .legalFor({{s32, s32}, {s64, s32}})
+ .legalFor({{i32, i32}, {i64, i32}})
.clampScalar(0, s32, s64)
.clampScalar(1, s32, s64)
.widenScalarToNextPow2(0);
@@ -352,8 +357,8 @@ AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST)
auto always = [=](const LegalityQuery &Q) { return true; };
getActionDefinitionsBuilder(G_CTPOP)
- .legalFor(HasCSSC, {{s32, s32}, {s64, s64}})
- .legalFor({{v8s8, v8s8}, {v16s8, v16s8}})
+ .legalFor(HasCSSC, {{i32, i32}, {i64, i64}})
+ .legalFor({{v8i8, v8i8}, {v16i8, v16i8}})
.customFor(!HasCSSC, {{s32, s32}, {s64, s64}})
.customFor({{s128, s128},
{v4s16, v4s16},
@@ -374,14 +379,14 @@ AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST)
.scalarizeIf(scalarOrEltWiderThan(0, 64), 0);
getActionDefinitionsBuilder({G_CTLZ, G_CTLS})
- .legalFor({{s32, s32},
- {s64, s64},
- {v8s8, v8s8},
- {v16s8, v16s8},
- {v4s16, v4s16},
- {v8s16, v8s16},
- {v2s32, v2s32},
- {v4s32, v4s32}})
+ .legalFor({{i32, i32},
+ {i64, i64},
+ {v8i8, v8i8},
+ {v16i8, v16i8},
+ {v4i16, v4i16},
+ {v8i16, v8i16},
+ {v2i32, v2i32},
+ {v4i32, v4i32}})
.widenScalarToNextPow2(1, /*Min=*/32)
.clampScalar(1, s32, s64)
.widenScalarOrEltToNextPow2OrMinSize(1, /*Min=*/8)
@@ -405,7 +410,7 @@ AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST)
getActionDefinitionsBuilder(G_CTTZ_ZERO_UNDEF).lower();
getActionDefinitionsBuilder(G_BITREVERSE)
- .legalFor({s32, s64, v8s8, v16s8})
+ .legalFor({i32, i64, v8i8, v16i8})
.widenScalarToNextPow2(0, /*Min = */ 32)
.widenScalarOrEltToNextPow2OrMinSize(0, 8)
.clampScalar(0, s32, s64)
@@ -418,7 +423,7 @@ AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST)
.lower();
getActionDefinitionsBuilder(G_BSWAP)
- .legalFor({s32, s64, v4s16, v8s16, v2s32, v4s32, v2s64})
+ .legalFor({i32, i64, v4i16, v8i16, v2i32, v4i32, v2i64})
.widenScalarOrEltToNextPow2(0, 16)
.clampScalar(0, s32, s64)
.clampNumElements(0, v4s16, v8s16)
@@ -427,8 +432,8 @@ AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST)
.moreElementsToNextPow2(0);
getActionDefinitionsBuilder({G_UADDSAT, G_SADDSAT, G_USUBSAT, G_SSUBSAT})
- .legalFor({v8s8, v16s8, v4s16, v8s16, v2s32, v4s32, v2s64})
- .legalFor(HasSVE, {nxv16s8, nxv8s16, nxv4s32, nxv2s64})
+ .legalFor({v8i8, v16i8, v4i16, v8i16, v2i32, v4i32, v2i64})
+ .legalFor(HasSVE, {nxv16i8, nxv8i16, nxv4i32, nxv2i64})
.clampNumElements(0, v8s8, v16s8)
.clampNumElements(0, v4s16, v8s16)
.clampNumElements(0, v2s32, v4s32)
@@ -482,14 +487,14 @@ AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST)
.libcallFor({{f32, i32}, {f64, i32}, {f128, i32}});
getActionDefinitionsBuilder({G_LROUND, G_INTRINSIC_LRINT})
- .legalFor({{s32, s32}, {s32, s64}, {s64, s32}, {s64, s64}})
- .legalFor(HasFP16, {{s32, s16}, {s64, s16}})
+ .legalFor({{i32, f32}, {i32, f64}, {i64, f32}, {i64, f64}})
+ .legalFor(HasFP16, {{i32, f16}, {i64, f16}})
.minScalar(1, s32)
.libcallFor({{s64, s128}})
.lower();
getActionDefinitionsBuilder({G_LLROUND, G_INTRINSIC_LLRINT})
- .legalFor({{s64, s32}, {s64, s64}})
- .legalFor(HasFP16, {{s64, s16}})
+ .legalFor({{i64, f32}, {i64, f64}})
+ .legalFor(HasFP16, {{i64, f16}})
.minScalar(0, s64)
.minScalar(1, s32)
.libcallFor({{s64, s128}})
@@ -759,12 +764,12 @@ AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST)
.customIf(isVector(0));
getActionDefinitionsBuilder(G_FCMP)
- .legalFor({{s32, f32},
- {s32, f64},
- {v4s32, v4f32},
- {v2s32, v2f32},
- {v2s64, v2f64}})
- .legalFor(HasFP16, {{s32, f16}, {v4s16, v4f16}, {v8s16, v8f16}})
+ .legalFor({{i32, f32},
+ {i32, f64},
+ {v4i32, v4f32},
+ {v2i32, v2f32},
+ {v2i64, v2f64}})
+ .legalFor(HasFP16, {{i32, f16}, {v4i16, v4f16}, {v8i16, v8f16}})
.widenScalarOrEltToNextPow2(1)
.clampScalar(0, s32, s32)
.minScalarOrElt(1, MinFPScalar)
@@ -845,12 +850,11 @@ AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST)
.alwaysLegal();
getActionDefinitionsBuilder({G_TRUNC_SSAT_S, G_TRUNC_SSAT_U, G_TRUNC_USAT_U})
- .legalFor({{v8s8, v8s16}, {v4s16, v4s32}, {v2s32, v2s64}})
+ .legalFor({{v8i8, v8i16}, {v4i16, v4i32}, {v2i32, v2i64}})
.clampNumElements(0, v2s32, v2s32);
getActionDefinitionsBuilder(G_SEXT_INREG)
- .legalFor({s32, s64})
- .legalFor(PackedVectorAllTypeList)
+ .legalFor({i32, i64, v8i8, v16i8, v4i16, v8i16, v2i32, v4i32, v2i64})
.maxScalar(0, s64)
.clampNumElements(0, v8s8, v16s8)
.clampNumElements(0, v4s16, v8s16)
More information about the llvm-commits
mailing list