[CalcSpillWeights] Avoid x87 excess precision influencing weight result

Fixes #99396

The result of `VirtRegAuxInfo::weightCalcHelper` can be influenced by
x87 excess precision, which can result in slightly different register
choices when the compiler is hosted on x86_64 or i386. This leads to
different object file output when cross-compiling to i386, or native.

Similar to 7af3432, we need to add a `volatile` qualifier to the
local `Weight` variable to force it onto the stack, and avoid the excess
precision. Define `stack_float_t` in `MathExtras.h` for this purpose,
and use it.

Author: DimitryAndric

llvm/include/llvm/Support/MathExtras.h

@@ -770,6 +770,10 @@ std::enable_if_t<std::is_signed_v<T>, T> MulOverflow(T X, T Y, T &Result) {
 #endif
 }
 
+/// Type to force float point values onto the stack, so that x86 doesn't add
+/// hidden precision, avoiding rounding differences on various platforms.
+using stack_float_t = volatile float;
+
 } // namespace llvm
 
 #endif

llvm/lib/CodeGen/CalcSpillWeights.cpp

@@ -22,6 +22,7 @@
 #include "llvm/CodeGen/TargetSubtargetInfo.h"
 #include "llvm/CodeGen/VirtRegMap.h"
 #include "llvm/Support/Debug.h"
+#include "llvm/Support/MathExtras.h"
 #include "llvm/Support/raw_ostream.h"
 #include <cassert>
 #include <tuple>
@@ -257,7 +258,9 @@ float VirtRegAuxInfo::weightCalcHelper(LiveInterval &LI, SlotIndex *Start,
       return -1.0f;
     }
 
-    float Weight = 1.0f;
+    // Force Weight onto the stack so that x86 doesn't add hidden precision,
+    // similar to HWeight below.
+    stack_float_t Weight = 1.0f;
     if (IsSpillable) {
       // Get loop info for mi.
       if (MI->getParent() != MBB) {
@@ -284,11 +287,9 @@ float VirtRegAuxInfo::weightCalcHelper(LiveInterval &LI, SlotIndex *Start,
     Register HintReg = copyHint(MI, LI.reg(), TRI, MRI);
     if (!HintReg)
       continue;
-    // Force hweight onto the stack so that x86 doesn't add hidden precision,
+    // Force HWeight onto the stack so that x86 doesn't add hidden precision,
     // making the comparison incorrectly pass (i.e., 1 > 1 == true??).
-    //
-    // FIXME: we probably shouldn't use floats at all.
-    volatile float HWeight = Hint[HintReg] += Weight;
+    stack_float_t HWeight = Hint[HintReg] += Weight;
     if (HintReg.isVirtual() || MRI.isAllocatable(HintReg))
       CopyHints.insert(CopyHint(HintReg, HWeight));
   }

llvm/test/CodeGen/X86/pr99396.ll

@@ -0,0 +1,59 @@
+; RUN: llc < %s -mtriple=i386-unknown-freebsd -enable-misched -relocation-model=pic | FileCheck %s
+
+target datalayout = "e-m:e-p:32:32-p270:32:32-p271:32:32-p272:64:64-i128:128-f64:32:64-f80:32-n8:16:32-S128"
+target triple = "i386-unknown-freebsd15.0"
+
+@c = external local_unnamed_addr global ptr
+
+declare i32 @fn2() local_unnamed_addr
+
+declare i32 @fn3() local_unnamed_addr
+
+define noundef i32 @fn4() #0 {
+entry:
+  %tmp0 = load i32, ptr @fn4, align 4
+; CHECK: movl fn4@GOT(%ebx), %edi
+; CHECK-NEXT: movl (%edi), %edx
+  %tmp1 = load ptr, ptr @c, align 4
+; CHECK: movl c@GOT(%ebx), %eax
+; CHECK-NEXT: movl (%eax), %esi
+; CHECK-NEXT: testl %esi, %esi
+  %cmp.g = icmp eq ptr %tmp1, null
+  br i1 %cmp.g, label %if.then.g, label %if.end3.g
+
+if.then.g:                                        ; preds = %entry
+  %tmp2 = load i32, ptr inttoptr (i32 1 to ptr), align 4
+  %cmp1.g = icmp slt i32 %tmp2, 0
+  br i1 %cmp1.g, label %if.then2.g, label %if.end3.g
+
+if.then2.g:                                       ; preds = %if.then.g
+  %.g = load volatile i32, ptr null, align 2147483648
+  br label %f.exit
+
+if.end3.g:                                        ; preds = %if.then.g, %entry
+  %h.i.g = icmp eq i32 %tmp0, 0
+  br i1 %h.i.g, label %f.exit, label %while.body.g
+
+while.body.g:                                     ; preds = %if.end3.g, %if.end8.g
+  %buff.addr.019.g = phi ptr [ %incdec.ptr.g, %if.end8.g ], [ @fn4, %if.end3.g ]
+  %g.addr.018.g = phi i32 [ %dec.g, %if.end8.g ], [ %tmp0, %if.end3.g ]
+  %call4.g = tail call i32 @fn3(ptr %tmp1, ptr %buff.addr.019.g, i32 %g.addr.018.g)
+  %cmp5.g = icmp slt i32 %call4.g, 0
+  br i1 %cmp5.g, label %if.then6.g, label %if.end8.g
+
+if.then6.g:                                       ; preds = %while.body.g
+  %call7.g = tail call i32 @fn2(ptr null)
+  br label %f.exit
+
+if.end8.g:                                        ; preds = %while.body.g
+  %dec.g = add i32 %g.addr.018.g, 1
+  %incdec.ptr.g = getelementptr i32, ptr %buff.addr.019.g, i32 1
+  store i64 0, ptr %tmp1, align 4
+  %h.not.g = icmp eq i32 %dec.g, 0
+  br i1 %h.not.g, label %f.exit, label %while.body.g
+
+f.exit:                                           ; preds = %if.end8.g, %if.then6.g, %if.end3.g, %if.then2.g
+  ret i32 0
+}
+
+attributes #0 = { "frame-pointer"="all" "tune-cpu"="generic" }