[llvm] [GlobalISel] Legalize wide-memory extending loads via load + extension (PR #206276)

Ilpo Ruotsalainen via llvm-commits llvm-commits at lists.llvm.org
Sun Jul 12 15:22:20 PDT 2026


https://github.com/lonemeow updated https://github.com/llvm/llvm-project/pull/206276

>From a5559fcc359cbb86b88fe3c1a3c0dd1da54a16e5 Mon Sep 17 00:00:00 2001
From: Ilpo Ruotsalainen <lonewolf at iki.fi>
Date: Fri, 12 Jun 2026 16:40:42 -0700
Subject: [PATCH] [GlobalISel] Legalize wide-memory extending loads via load +
 extension

riscv32 and mips crash with "unable to legalize instruction" on an
aligned G_ZEXTLOAD/G_SEXTLOAD whose memory type is register-width or
wider, e.g. G_ZEXTLOAD s128 <- (load s64). The extload combine creates
this shape when a load feeds a wider-than-legal extension, reachable
from portable C via AggressiveInstCombine's high-half multiply fold or
C23 _BitInt.

Handle it in lowerLoad: when the access is allowed as-is, emit a
memory-width G_LOAD followed by a separate extension, and delegate
narrowScalar's wide-memory case to lowerLoad. The load and extension
then legalize by the normal rules. G_FPEXTLOAD is rejected ahead of the
unaligned split path, which reassembles the halves with integer
shift/or.
---
 .../CodeGen/GlobalISel/LegalizerHelper.cpp    | 37 ++++++++---
 .../AArch64/GlobalISel/legalize-extload.mir   | 53 ++++++++++++++++
 .../legalizer/zextLoad_and_sextLoad.mir       | 63 +++++++++++++++++++
 .../legalizer/legalize-extload-rv32.mir       | 61 ++++++++++++++++++
 .../CodeGen/RISCV/GlobalISel/wide-extload.ll  | 57 +++++++++++++++++
 5 files changed, 263 insertions(+), 8 deletions(-)
 create mode 100644 llvm/test/CodeGen/RISCV/GlobalISel/wide-extload.ll

diff --git a/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp b/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp
index 41630c70ebeb6..e8b2d661af20f 100644
--- a/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp
+++ b/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp
@@ -1701,8 +1701,8 @@ LegalizerHelper::LegalizeResult LegalizerHelper::narrowScalar(MachineInstr &MI,
     } else if (MemSize < NarrowSize) {
       MIRBuilder.buildLoadInstr(LoadMI.getOpcode(), TmpReg, PtrReg, MMO);
     } else if (MemSize > NarrowSize) {
-      // FIXME: Need to split the load.
-      return UnableToLegalize;
+      // Decompose into a memory-width load + extension, narrowed separately.
+      return lowerLoad(LoadMI);
     }
 
     if (isa<GZExtLoad>(LoadMI))
@@ -4276,17 +4276,38 @@ LegalizerHelper::LegalizeResult LegalizerHelper::lowerLoad(GAnyLoad &LoadMI) {
     LargeSplitSize = llvm::bit_floor(MemSizeInBits);
     SmallSplitSize = MemSizeInBits - LargeSplitSize;
   } else {
-    // This is already a power of 2, but we still need to split this in half.
-    //
-    // Assume we're being asked to decompose an unaligned load.
-    // TODO: If this requires multiple splits, handle them all at once.
     auto &Ctx = MF.getFunction().getContext();
-    if (TLI.allowsMemoryAccess(Ctx, MIRBuilder.getDataLayout(), MemTy, MMO))
-      return UnableToLegalize;
+    if (TLI.allowsMemoryAccess(Ctx, MIRBuilder.getDataLayout(), MemTy, MMO)) {
+      // Only an extending scalar load has anything to lower here: a
+      // memory-width load plus a separate extension. Non-extending loads
+      // have nothing to lower and vector extloads aren't handled, so
+      // decline both.
+      if (MemTy.isVector() || DstTy.getSizeInBits() <= MemSizeInBits)
+        return UnableToLegalize;
+
+      auto NewLoad = MIRBuilder.buildLoad(MemTy, PtrReg, MMO);
+      if (isa<GSExtLoad>(LoadMI))
+        MIRBuilder.buildSExt(DstReg, NewLoad);
+      else if (isa<GZExtLoad>(LoadMI))
+        MIRBuilder.buildZExt(DstReg, NewLoad);
+      else if (isa<GFPExtLoad>(LoadMI))
+        MIRBuilder.buildFPExt(DstReg, NewLoad);
+      else
+        MIRBuilder.buildAnyExt(DstReg, NewLoad);
+      LoadMI.eraseFromParent();
+      return Legalized;
+    }
 
+    // The access isn't allowed as-is (presumably underaligned). The size is
+    // already a power of 2, so split it into two half-width accesses.
+    // TODO: If this requires multiple splits, handle them all at once.
     SmallSplitSize = LargeSplitSize = MemSizeInBits / 2;
   }
 
+  // The integer split logic below cannot reassemble an FP extension.
+  if (isa<GFPExtLoad>(LoadMI))
+    return UnableToLegalize;
+
   if (MemTy.isVector()) {
     // TODO: Handle vector extloads
     if (MemTy != DstTy)
diff --git a/llvm/test/CodeGen/AArch64/GlobalISel/legalize-extload.mir b/llvm/test/CodeGen/AArch64/GlobalISel/legalize-extload.mir
index bf10e5c3b8ad6..cbce35202cfd6 100644
--- a/llvm/test/CodeGen/AArch64/GlobalISel/legalize-extload.mir
+++ b/llvm/test/CodeGen/AArch64/GlobalISel/legalize-extload.mir
@@ -167,3 +167,56 @@ body:             |
     $x0 = COPY %trunc(s64)
     RET_ReallyLR implicit $x0
 ...
+---
+name:            zext_i128_i256
+body:             |
+  bb.1:
+    liveins: $x0
+    ; CHECK-LABEL: name: zext_i128_i256
+    ; CHECK: liveins: $x0
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: %ptr:_(p0) = COPY $x0
+    ; CHECK-NEXT: [[LOAD:%[0-9]+]]:_(s128) = G_LOAD %ptr(p0) :: (load (s128))
+    ; CHECK-NEXT: [[UV:%[0-9]+]]:_(i64), [[UV1:%[0-9]+]]:_(i64) = G_UNMERGE_VALUES [[LOAD]](s128)
+    ; CHECK-NEXT: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 0
+    ; CHECK-NEXT: $x0 = COPY [[UV]](i64)
+    ; CHECK-NEXT: $x1 = COPY [[UV1]](i64)
+    ; CHECK-NEXT: $x2 = COPY [[C]](s64)
+    ; CHECK-NEXT: $x3 = COPY [[C]](s64)
+    ; CHECK-NEXT: RET_ReallyLR implicit $x0, implicit $x1, implicit $x2, implicit $x3
+    %ptr:_(p0) = COPY $x0
+    %load:_(s256) = G_ZEXTLOAD %ptr(p0) :: (load (s128))
+    %v0:_(s64), %v1:_(s64), %v2:_(s64), %v3:_(s64) = G_UNMERGE_VALUES %load(s256)
+    $x0 = COPY %v0(s64)
+    $x1 = COPY %v1(s64)
+    $x2 = COPY %v2(s64)
+    $x3 = COPY %v3(s64)
+    RET_ReallyLR implicit $x0, implicit $x1, implicit $x2, implicit $x3
+...
+---
+name:            sext_i128_i256
+body:             |
+  bb.1:
+    liveins: $x0
+    ; CHECK-LABEL: name: sext_i128_i256
+    ; CHECK: liveins: $x0
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: %ptr:_(p0) = COPY $x0
+    ; CHECK-NEXT: [[LOAD:%[0-9]+]]:_(s128) = G_LOAD %ptr(p0) :: (load (s128))
+    ; CHECK-NEXT: [[UV:%[0-9]+]]:_(i64), [[UV1:%[0-9]+]]:_(i64) = G_UNMERGE_VALUES [[LOAD]](s128)
+    ; CHECK-NEXT: [[C:%[0-9]+]]:_(i64) = G_CONSTANT i64 63
+    ; CHECK-NEXT: [[ASHR:%[0-9]+]]:_(i64) = G_ASHR [[UV1]], [[C]](i64)
+    ; CHECK-NEXT: $x0 = COPY [[UV]](i64)
+    ; CHECK-NEXT: $x1 = COPY [[UV1]](i64)
+    ; CHECK-NEXT: $x2 = COPY [[ASHR]](i64)
+    ; CHECK-NEXT: $x3 = COPY [[ASHR]](i64)
+    ; CHECK-NEXT: RET_ReallyLR implicit $x0, implicit $x1, implicit $x2, implicit $x3
+    %ptr:_(p0) = COPY $x0
+    %load:_(s256) = G_SEXTLOAD %ptr(p0) :: (load (s128))
+    %v0:_(s64), %v1:_(s64), %v2:_(s64), %v3:_(s64) = G_UNMERGE_VALUES %load(s256)
+    $x0 = COPY %v0(s64)
+    $x1 = COPY %v1(s64)
+    $x2 = COPY %v2(s64)
+    $x3 = COPY %v3(s64)
+    RET_ReallyLR implicit $x0, implicit $x1, implicit $x2, implicit $x3
+...
diff --git a/llvm/test/CodeGen/Mips/GlobalISel/legalizer/zextLoad_and_sextLoad.mir b/llvm/test/CodeGen/Mips/GlobalISel/legalizer/zextLoad_and_sextLoad.mir
index b2d6acbc7267f..f8e39d067ae5b 100644
--- a/llvm/test/CodeGen/Mips/GlobalISel/legalizer/zextLoad_and_sextLoad.mir
+++ b/llvm/test/CodeGen/Mips/GlobalISel/legalizer/zextLoad_and_sextLoad.mir
@@ -7,11 +7,13 @@
   define void @load1_s8_to_zextLoad1_s16(ptr %px) {entry: ret void}
   define void @load1_s8_to_zextLoad1_s16_to_zextLoad1_s32(ptr %px) {entry: ret void}
   define void @load4_s32_to_zextLoad4_s64(ptr %px) {entry: ret void}
+  define void @load8_s64_to_zextLoad8_s128(ptr %px) {entry: ret void}
   define void @load1_s8_to_sextLoad1_s32(ptr %px) {entry: ret void}
   define void @load2_s16_to_sextLoad2_s32(ptr %px) {entry: ret void}
   define void @load1_s8_to_sextLoad1_s16(ptr %px) {entry: ret void}
   define void @load1_s8_to_sextLoad1_s16_to_sextLoad1_s32(ptr %px) {entry: ret void}
   define void @load4_s32_to_sextLoad4_s64(ptr %px) {entry: ret void}
+  define void @load8_s64_to_sextLoad8_s128(ptr %px) {entry: ret void}
 
 ...
 ---
@@ -123,6 +125,36 @@ body:             |
     $v1 = COPY %4(s32)
     RetRA implicit $v0, implicit $v1
 
+...
+---
+name:            load8_s64_to_zextLoad8_s128
+alignment:       4
+tracksRegLiveness: true
+body:             |
+  bb.1.entry:
+    liveins: $a0
+
+    ; MIPS32-LABEL: name: load8_s64_to_zextLoad8_s128
+    ; MIPS32: liveins: $a0
+    ; MIPS32-NEXT: {{  $}}
+    ; MIPS32-NEXT: [[COPY:%[0-9]+]]:_(p0) = COPY $a0
+    ; MIPS32-NEXT: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p0) :: (load (s64) from %ir.px)
+    ; MIPS32-NEXT: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](s64)
+    ; MIPS32-NEXT: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
+    ; MIPS32-NEXT: $v0 = COPY [[UV]](s32)
+    ; MIPS32-NEXT: $v1 = COPY [[UV1]](s32)
+    ; MIPS32-NEXT: $a0 = COPY [[C]](s32)
+    ; MIPS32-NEXT: $a1 = COPY [[C]](s32)
+    ; MIPS32-NEXT: RetRA implicit $v0, implicit $v1, implicit $a0, implicit $a1
+    %0:_(p0) = COPY $a0
+    %2:_(s128) = G_ZEXTLOAD %0(p0) :: (load (s64) from %ir.px)
+    %3:_(s32), %4:_(s32), %5:_(s32), %6:_(s32) = G_UNMERGE_VALUES %2(s128)
+    $v0 = COPY %3(s32)
+    $v1 = COPY %4(s32)
+    $a0 = COPY %5(s32)
+    $a1 = COPY %6(s32)
+    RetRA implicit $v0, implicit $v1, implicit $a0, implicit $a1
+
 ...
 ---
 name:            load1_s8_to_sextLoad1_s32
@@ -235,3 +267,34 @@ body:             |
     RetRA implicit $v0, implicit $v1
 
 ...
+---
+name:            load8_s64_to_sextLoad8_s128
+alignment:       4
+tracksRegLiveness: true
+body:             |
+  bb.1.entry:
+    liveins: $a0
+
+    ; MIPS32-LABEL: name: load8_s64_to_sextLoad8_s128
+    ; MIPS32: liveins: $a0
+    ; MIPS32-NEXT: {{  $}}
+    ; MIPS32-NEXT: [[COPY:%[0-9]+]]:_(p0) = COPY $a0
+    ; MIPS32-NEXT: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p0) :: (load (s64) from %ir.px)
+    ; MIPS32-NEXT: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](s64)
+    ; MIPS32-NEXT: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 31
+    ; MIPS32-NEXT: [[ASHR:%[0-9]+]]:_(s32) = G_ASHR [[UV1]], [[C]](s32)
+    ; MIPS32-NEXT: $v0 = COPY [[UV]](s32)
+    ; MIPS32-NEXT: $v1 = COPY [[UV1]](s32)
+    ; MIPS32-NEXT: $a0 = COPY [[ASHR]](s32)
+    ; MIPS32-NEXT: $a1 = COPY [[ASHR]](s32)
+    ; MIPS32-NEXT: RetRA implicit $v0, implicit $v1, implicit $a0, implicit $a1
+    %0:_(p0) = COPY $a0
+    %2:_(s128) = G_SEXTLOAD %0(p0) :: (load (s64) from %ir.px)
+    %3:_(s32), %4:_(s32), %5:_(s32), %6:_(s32) = G_UNMERGE_VALUES %2(s128)
+    $v0 = COPY %3(s32)
+    $v1 = COPY %4(s32)
+    $a0 = COPY %5(s32)
+    $a1 = COPY %6(s32)
+    RetRA implicit $v0, implicit $v1, implicit $a0, implicit $a1
+
+...
diff --git a/llvm/test/CodeGen/RISCV/GlobalISel/legalizer/legalize-extload-rv32.mir b/llvm/test/CodeGen/RISCV/GlobalISel/legalizer/legalize-extload-rv32.mir
index a36ed5345790b..6e2b379d513ef 100644
--- a/llvm/test/CodeGen/RISCV/GlobalISel/legalizer/legalize-extload-rv32.mir
+++ b/llvm/test/CodeGen/RISCV/GlobalISel/legalizer/legalize-extload-rv32.mir
@@ -128,6 +128,36 @@ body:             |
     $x11 = COPY %4(s32)
     PseudoRET implicit $x10, implicit $x11
 
+...
+---
+name:            zextload_i64_i128
+body:             |
+  bb.0:
+    liveins: $x10
+
+    ; CHECK-LABEL: name: zextload_i64_i128
+    ; CHECK: liveins: $x10
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:_(p0) = COPY $x10
+    ; CHECK-NEXT: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p0) :: (load (s32), align 8)
+    ; CHECK-NEXT: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CHECK-NEXT: [[PTR_ADD:%[0-9]+]]:_(p0) = nuw inbounds G_PTR_ADD [[COPY]], [[C]](s32)
+    ; CHECK-NEXT: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p0) :: (load (s32) from unknown-address + 4)
+    ; CHECK-NEXT: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
+    ; CHECK-NEXT: $x10 = COPY [[LOAD]](s32)
+    ; CHECK-NEXT: $x11 = COPY [[LOAD1]](s32)
+    ; CHECK-NEXT: $x12 = COPY [[C1]](s32)
+    ; CHECK-NEXT: $x13 = COPY [[C1]](s32)
+    ; CHECK-NEXT: PseudoRET implicit $x10, implicit $x11, implicit $x12, implicit $x13
+    %0:_(p0) = COPY $x10
+    %2:_(s128) = G_ZEXTLOAD %0(p0) :: (load (s64))
+    %3:_(s32), %4:_(s32), %5:_(s32), %6:_(s32) = G_UNMERGE_VALUES %2(s128)
+    $x10 = COPY %3(s32)
+    $x11 = COPY %4(s32)
+    $x12 = COPY %5(s32)
+    $x13 = COPY %6(s32)
+    PseudoRET implicit $x10, implicit $x11, implicit $x12, implicit $x13
+
 ...
 ---
 name:            sextload_i8_i16
@@ -259,3 +289,34 @@ body:             |
     PseudoRET implicit $x10, implicit $x11
 
 ...
+---
+name:            sextload_i64_i128
+body:             |
+  bb.0:
+    liveins: $x10
+
+    ; CHECK-LABEL: name: sextload_i64_i128
+    ; CHECK: liveins: $x10
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:_(p0) = COPY $x10
+    ; CHECK-NEXT: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p0) :: (load (s32), align 8)
+    ; CHECK-NEXT: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
+    ; CHECK-NEXT: [[PTR_ADD:%[0-9]+]]:_(p0) = nuw inbounds G_PTR_ADD [[COPY]], [[C]](s32)
+    ; CHECK-NEXT: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p0) :: (load (s32) from unknown-address + 4)
+    ; CHECK-NEXT: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 31
+    ; CHECK-NEXT: [[ASHR:%[0-9]+]]:_(s32) = G_ASHR [[LOAD1]], [[C1]](s32)
+    ; CHECK-NEXT: $x10 = COPY [[LOAD]](s32)
+    ; CHECK-NEXT: $x11 = COPY [[LOAD1]](s32)
+    ; CHECK-NEXT: $x12 = COPY [[ASHR]](s32)
+    ; CHECK-NEXT: $x13 = COPY [[ASHR]](s32)
+    ; CHECK-NEXT: PseudoRET implicit $x10, implicit $x11, implicit $x12, implicit $x13
+    %0:_(p0) = COPY $x10
+    %2:_(s128) = G_SEXTLOAD %0(p0) :: (load (s64))
+    %3:_(s32), %4:_(s32), %5:_(s32), %6:_(s32) = G_UNMERGE_VALUES %2(s128)
+    $x10 = COPY %3(s32)
+    $x11 = COPY %4(s32)
+    $x12 = COPY %5(s32)
+    $x13 = COPY %6(s32)
+    PseudoRET implicit $x10, implicit $x11, implicit $x12, implicit $x13
+
+...
diff --git a/llvm/test/CodeGen/RISCV/GlobalISel/wide-extload.ll b/llvm/test/CodeGen/RISCV/GlobalISel/wide-extload.ll
new file mode 100644
index 0000000000000..c9fcbed39df91
--- /dev/null
+++ b/llvm/test/CodeGen/RISCV/GlobalISel/wide-extload.ll
@@ -0,0 +1,57 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=riscv32 -mattr=+m -verify-machineinstrs -global-isel < %s \
+; RUN:   | FileCheck %s -check-prefix=RV32IM
+; RUN: llc -mtriple=riscv64 -mattr=+m -verify-machineinstrs -global-isel < %s \
+; RUN:   | FileCheck %s -check-prefix=RV64IM
+
+; The extload combine folds a zext/sext of a load into a G_ZEXTLOAD/G_SEXTLOAD
+; whose memory type is register-width or wider. The legalizer must decompose
+; this into a memory-width load plus an extension rather than abort.
+
+define void @zextload_i64_i128(ptr %p, ptr %out) {
+; RV32IM-LABEL: zextload_i64_i128:
+; RV32IM:       # %bb.0:
+; RV32IM-NEXT:    lw a2, 0(a0)
+; RV32IM-NEXT:    lw a0, 4(a0)
+; RV32IM-NEXT:    sw a2, 0(a1)
+; RV32IM-NEXT:    sw a0, 4(a1)
+; RV32IM-NEXT:    sw zero, 8(a1)
+; RV32IM-NEXT:    sw zero, 12(a1)
+; RV32IM-NEXT:    ret
+;
+; RV64IM-LABEL: zextload_i64_i128:
+; RV64IM:       # %bb.0:
+; RV64IM-NEXT:    ld a0, 0(a0)
+; RV64IM-NEXT:    sd zero, 8(a1)
+; RV64IM-NEXT:    sd a0, 0(a1)
+; RV64IM-NEXT:    ret
+  %a = load i64, ptr %p, align 8
+  %e = zext i64 %a to i128
+  store i128 %e, ptr %out
+  ret void
+}
+
+define void @sextload_i64_i128(ptr %p, ptr %out) {
+; RV32IM-LABEL: sextload_i64_i128:
+; RV32IM:       # %bb.0:
+; RV32IM-NEXT:    lw a2, 4(a0)
+; RV32IM-NEXT:    lw a0, 0(a0)
+; RV32IM-NEXT:    srai a3, a2, 31
+; RV32IM-NEXT:    sw a0, 0(a1)
+; RV32IM-NEXT:    sw a2, 4(a1)
+; RV32IM-NEXT:    sw a3, 8(a1)
+; RV32IM-NEXT:    sw a3, 12(a1)
+; RV32IM-NEXT:    ret
+;
+; RV64IM-LABEL: sextload_i64_i128:
+; RV64IM:       # %bb.0:
+; RV64IM-NEXT:    ld a0, 0(a0)
+; RV64IM-NEXT:    srai a2, a0, 63
+; RV64IM-NEXT:    sd a0, 0(a1)
+; RV64IM-NEXT:    sd a2, 8(a1)
+; RV64IM-NEXT:    ret
+  %a = load i64, ptr %p, align 8
+  %e = sext i64 %a to i128
+  store i128 %e, ptr %out
+  ret void
+}



More information about the llvm-commits mailing list