fix(transform): fix fcmp folding

dtcxzyw · Aug 21, 2023 · c7b254e · c7b254e
1 parent 648819e
commit c7b254e
Show file tree

Hide file tree

Showing 19 changed files with 1,927 additions and 539 deletions.
diff --git a/cmmc/Target/RISCV/RISCV.yml b/cmmc/Target/RISCV/RISCV.yml
@@ -1787,14 +1787,71 @@ InstSelection:
             Lhs: $X
             Rhs: $Y
         Rhs: $Z
-        $Predicate: isAggressiveOptEnabled()
+        $Predicate: isAggressiveOptEnabled() && ctx.hasOneUse($MulDst)
     Replace:
       FMADD_S:
         Rd: $Dst
         Rs1: $X
         Rs2: $Y
         Rs3: $Z
 
+  - Pattern:
+      InstFSub:
+        Dst: $Dst
+        Lhs:
+          InstFMul:
+            Dst: $MulDst
+            Lhs: $X
+            Rhs: $Y
+        Rhs: $Z
+        $Predicate: isAggressiveOptEnabled() && ctx.hasOneUse($MulDst)
+    Replace:
+      FMSUB_S:
+        Rd: $Dst
+        Rs1: $X
+        Rs2: $Y
+        Rs3: $Z
+
+  - Pattern:
+      InstFAdd:
+        Dst: $Dst
+        Lhs:
+          InstFNeg:
+            Dst: $NegDst
+            Src:
+              InstFMul:
+                Dst: $MulDst
+                Lhs: $X
+                Rhs: $Y
+        Rhs: $Z
+        $Predicate: isAggressiveOptEnabled() && ctx.hasOneUse($MulDst) && ctx.hasOneUse($NegDst)
+    Replace:
+      FNMSUB_S:
+        Rd: $Dst
+        Rs1: $X
+        Rs2: $Y
+        Rs3: $Z
+
+  - Pattern:
+      InstFSub:
+        Dst: $Dst
+        Lhs:
+          InstFNeg:
+            Dst: $NegDst
+            Src:
+              InstFMul:
+                Dst: $MulDst
+                Lhs: $X
+                Rhs: $Y
+        Rhs: $Z
+        $Predicate: isAggressiveOptEnabled() && ctx.hasOneUse($MulDst) && ctx.hasOneUse($NegDst)
+    Replace:
+      FNMADD_S:
+        Rd: $Dst
+        Rs1: $X
+        Rs2: $Y
+        Rs3: $Z
+
   - Pattern:
       $Template:
         Dst: $Dst

diff --git a/cmmc/Transforms/Combine/ArithmeticReduce.cpp b/cmmc/Transforms/Combine/ArithmeticReduce.cpp
@@ -1327,28 +1327,30 @@ class ArithmeticReduce final : public TransformPass<Function> {
                     return val;
             }
 
-            // TODO: handle nan
             // x <= f1 || x >= f2
             if(or_(fcmp(cmp1, any(v1), fp_(f1)), fcmp(cmp2, exactly(v1), fp_(f2)))(matchCtx)) {
                 if(cmp1 > cmp2) {
                     std::swap(cmp1, cmp2);
                     std::swap(f1, f2);
                 }
+                auto isNotNan = [&](Value* v) {
+                    return builder.makeOp<CompareInst>(InstructionID::FCmp, CompareOp::FCmpOrderedEqual, v, v);
+                };
                 if(cmp1 == CompareOp::FCmpOrderedLessEqual && cmp2 == CompareOp::FCmpOrderedGreaterEqual) {
                     if(f1 >= f2)
-                        return builder.getTrue();
+                        return isNotNan(v1);
                 }
                 if(cmp1 == CompareOp::FCmpOrderedLessEqual && cmp2 == CompareOp::FCmpOrderedGreaterThan) {
                     if(f1 >= f2)
-                        return builder.getTrue();
+                        return isNotNan(v1);
                 }
                 if(cmp1 == CompareOp::FCmpOrderedLessThan && cmp2 == CompareOp::FCmpOrderedGreaterEqual) {
                     if(f1 >= f2)
-                        return builder.getTrue();
+                        return isNotNan(v1);
                 }
                 if(cmp1 == CompareOp::FCmpOrderedLessThan && cmp2 == CompareOp::FCmpOrderedGreaterThan) {
                     if(f1 > f2)
-                        return builder.getTrue();
+                        return isNotNan(v1);
                     // TODO: f1 == f2 -> x une f1
                 }
             }

diff --git a/tests/Regression/CodeGen/fma.arm.s b/tests/Regression/CodeGen/fma.arm.s
diff --git a/tests/Regression/CodeGen/fma.mips.s b/tests/Regression/CodeGen/fma.mips.s
diff --git a/tests/Regression/CodeGen/fma.riscv.s b/tests/Regression/CodeGen/fma.riscv.s
diff --git a/tests/Regression/CodeGen/fma.sy b/tests/Regression/CodeGen/fma.sy
@@ -4,3 +4,12 @@ float fmadd1(float x, float y, float z) {
 float fmadd2(float x, float y, float z) {
     return z + x * y;
 }
+float fmsub(float x, float y, float z) {
+    return x * y - z;
+}
+float fnmadd(float x, float y, float z) {
+    return - (x * y) - z;
+}
+float fnmsub(float x, float y, float z) {
+    return - (x * y) + z;
+}
diff --git a/tests/SysY2022/performance/large_loop_array_1.out b/tests/SysY2022/performance/large_loop_array_1.out
@@ -1,2 +1,2 @@
+0x1.45378p+50
 0
-0