[AArch64][GlobalISel] Use the IR pointer type in MMO to guess G_LOAD regbanks

[aemerson]

Summary

Results in large perf improvements on FP benchmarks and some minor improvements in code size too.

[arsenm]

llvm/lib/Target/AArch64/AArch64RegisterBankInfo.cpp

689–690

This is going to break opaque pointers

[aemerson]

llvm/lib/Target/AArch64/AArch64RegisterBankInfo.cpp

689–690

Indeed opaque pointers will break this. Which is why I was going to propose some other form of propagating type information on vregs. RFC incoming...

[aemerson]

Abandoning, we'll probably use the FP type info in LLT when that get's added eventually.

llvm/lib/Target/AArch64/AArch64RegisterBankInfo.cpp

case TargetOpcode::G_LOAD:
     // Loading in vector unit is slightly more expensive.
     // This is actually only true for the LD1R and co instructions,
     // but anyway for the fast mode this number does not matter and
     // for the greedy mode the cost of the cross bank copy will
     // offset this number.
     // FIXME: Should be derived from the scheduling model.
     if (OpRegBankIdx[0] != PMI_FirstGPR)
       Cost = 2;
-    else
-      // Check if that load feeds fp instructions.
-      // In that case, we want the default mapping to be on FPR
-      // instead of blind map every scalar to GPR.
-      for (const MachineInstr &UseMI :
-           MRI.use_nodbg_instructions(MI.getOperand(0).getReg())) {
-        // If we have at least one direct use in a FP instruction,
-        // assume this was a floating point load in the IR.
-        // If it was not, we would have had a bitcast before
-        // reaching that instruction.
-        if (onlyUsesFP(UseMI, MRI, TRI)) {
-          OpRegBankIdx[0] = PMI_FirstFPR;
-          break;
-        }
-      }
+    else {
+      // Check if the MMO on the load indicates the IR type is an FP type.
+      const auto &MMO = **MI.memoperands_begin();
+      const Value *LdVal = MMO.getValue();
+      if (LdVal) {
+        PointerType *PTy = cast<PointerType>(LdVal->getType());
+        if (PTy->getElementType()->isFPOrFPVectorTy()) {
+          OpRegBankIdx[0] = PMI_FirstFPR;
+          break;
+        }
+      }
+    }
+    // If we're going to assign GPR, have a second attempt to check for FP
+    // users, in case one of them is a COPY that writes to an FPR register.
     break;
   case TargetOpcode::G_STORE:
     // Check if that store is fed by fp instructions.
     if (OpRegBankIdx[0] == PMI_FirstGPR) {
       Register VReg = MI.getOperand(0).getReg();
       if (!VReg)
         break;
       MachineInstr *DefMI = MRI.getVRegDef(VReg);

llvm/test/CodeGen/AArch64/GlobalISel/regbankselect-fp-loads.mir

+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -O0 -mtriple=aarch64-apple-ios -run-pass=regbankselect -verify-machineinstrs %s -o - | FileCheck %s -check-prefix=CHECK -check-prefix=IOS
+
+--- |
+  target datalayout = "e-m:o-i64:64-i128:128-n32:64-S128"
+
+  @var_fp = global float 0.0
+  @var_int = global i32 0
+
+  define float @fp_load_phi() { ret float undef }
+  define i32 @int_load_phi() { ret i32 undef }
+
+...
+---
+name:            fp_load_phi
+legalized:       true
+regBankSelected: false
+tracksRegLiveness: true
+body:             |
+  ; CHECK-LABEL: name: fp_load_phi
+  ; CHECK: bb.0:
+  ; CHECK:   successors: %bb.1(0x80000000)
+  ; CHECK:   liveins: $w0
+  ; CHECK:   [[COPY:%[0-9]+]]:gpr(s32) = COPY $w0
+  ; CHECK:   [[GV:%[0-9]+]]:gpr(p0) = G_GLOBAL_VALUE @var_fp
+  ; CHECK:   %fp_load:fpr(s32) = G_LOAD [[GV]](p0) :: (load 4 from @var_fp)
+  ; CHECK: bb.1:
+  ; CHECK:   successors: %bb.1(0x40000000), %bb.2(0x40000000)
+  ; CHECK:   [[PHI:%[0-9]+]]:fpr(s32) = PHI %fp_load(s32), %bb.0, [[PHI]](s32), %bb.1
+  ; CHECK:   G_BRCOND [[COPY]](s32), %bb.1
+  ; CHECK: bb.2:
+  ; CHECK:   $s0 = COPY [[PHI]](s32)
+  ; CHECK:   RET_ReallyLR implicit $s0
+  ; Here we're checking that the load is assigned an FPR bank, since it's
+  ; loading from an fp type in the IR.
+  bb.0:
+    liveins: $w0
+    successors: %bb.1
+    %0:_(s32) = COPY $w0
+    %1:_(p0) = G_GLOBAL_VALUE @var_fp
+    %fp_load:_(s32) = G_LOAD %1 :: (load 4 from @var_fp)
+
+  bb.1:
+    successors: %bb.1, %bb.2
+    %2:_(s32) = PHI %fp_load, %bb.0, %2, %bb.1
+    G_BRCOND %0, %bb.1
+
+  bb.2:
+    $s0 = COPY %2
+    RET_ReallyLR implicit $s0
+...
+
+---
+name:            int_load_phi
+legalized:       true
+regBankSelected: false
+tracksRegLiveness: true
+body:             |
+  ; CHECK-LABEL: name: int_load_phi
+  ; CHECK: bb.0:
+  ; CHECK:   successors: %bb.1(0x80000000)
+  ; CHECK:   liveins: $w0
+  ; CHECK:   [[COPY:%[0-9]+]]:gpr(s32) = COPY $w0
+  ; CHECK:   [[GV:%[0-9]+]]:gpr(p0) = G_GLOBAL_VALUE @var_fp
+  ; CHECK:   %fp_load:gpr(s32) = G_LOAD [[GV]](p0) :: (load 4 from @var_int)
+  ; CHECK: bb.1:
+  ; CHECK:   successors: %bb.1(0x40000000), %bb.2(0x40000000)
+  ; CHECK:   [[PHI:%[0-9]+]]:gpr(s32) = PHI %fp_load(s32), %bb.0, [[PHI]](s32), %bb.1
+  ; CHECK:   G_BRCOND [[COPY]](s32), %bb.1
+  ; CHECK: bb.2:
+  ; CHECK:   $s0 = COPY [[PHI]](s32)
+  ; CHECK:   RET_ReallyLR implicit $s0
+  bb.0:
+    liveins: $w0
+    successors: %bb.1
+    %0:_(s32) = COPY $w0
+    %1:_(p0) = G_GLOBAL_VALUE @var_fp
+    %fp_load:_(s32) = G_LOAD %1 :: (load 4 from @var_int)
+
+  bb.1:
+    successors: %bb.1, %bb.2
+    %2:_(s32) = PHI %fp_load, %bb.0, %2, %bb.1
+    G_BRCOND %0, %bb.1
+
+  bb.2:
+    $s0 = COPY %2
+    RET_ReallyLR implicit $s0
+...