[InstCombine] Make fptrunc combine use intersection of fast math flags

[john-brawn-arm]

These combines involve swapping the fptrunc with its operand, and using the intersection of fast math flags is the safest option as e.g. if we have (fptrunc (fneg ninf x)) then (fneg ninf (fptrunc x)) will not be correct as if x is a not within the range of the destination type the result of (fptrunc x) will be inf.

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

Author: John Brawn (john-brawn-arm)

Changes

These combines involve swapping the fptrunc with its operand, and using the union of fast math flags is the safest option as e.g. if we have (fptrunc (fneg ninf x)) then (fneg ninf (fptrunc x)) will not be correct as if x is a not within the range of the destination type the result of (fptrunc x) will be inf.

---

Full diff: https://github.com/llvm/llvm-project/pull/118808.diff

3 Files Affected:

• (modified) llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp (+5-4)
• (modified) llvm/test/Transforms/InstCombine/fpcast.ll (+55)
• (modified) llvm/test/Transforms/InstCombine/fptrunc.ll (+24)

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp
index 7221c987b98219..a6c5507c764c76 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp
@@ -1847,15 +1847,16 @@ Instruction *InstCombinerImpl::visitFPTrunc(FPTruncInst &FPT) {
   Value *X;
   Instruction *Op = dyn_cast<Instruction>(FPT.getOperand(0));
   if (Op && Op->hasOneUse()) {
-    // FIXME: The FMF should propagate from the fptrunc, not the source op.
     IRBuilder<>::FastMathFlagGuard FMFG(Builder);
+    FastMathFlags FMF = FPT.getFastMathFlags();
     if (isa<FPMathOperator>(Op))
-      Builder.setFastMathFlags(Op->getFastMathFlags());
+      FMF &= Op->getFastMathFlags();
+    Builder.setFastMathFlags(FMF);
 
     if (match(Op, m_FNeg(m_Value(X)))) {
       Value *InnerTrunc = Builder.CreateFPTrunc(X, Ty);
-
-      return UnaryOperator::CreateFNegFMF(InnerTrunc, Op);
+      Value *Neg = Builder.CreateFNeg(InnerTrunc);
+      return replaceInstUsesWith(FPT, Neg);
     }
 
     // If we are truncating a select that has an extended operand, we can
diff --git a/llvm/test/Transforms/InstCombine/fpcast.ll b/llvm/test/Transforms/InstCombine/fpcast.ll
index 029e513ceafbcd..72bd42e60e0d27 100644
--- a/llvm/test/Transforms/InstCombine/fpcast.ll
+++ b/llvm/test/Transforms/InstCombine/fpcast.ll
@@ -29,6 +29,17 @@ define half @test3(float %a) {
   ret half %c
 }
 
+define half @test3_fast(float %a) {
+; CHECK-LABEL: @test3_fast(
+; CHECK-NEXT:    [[TMP1:%.*]] = fptrunc float [[A:%.*]] to half
+; CHECK-NEXT:    [[C:%.*]] = call half @llvm.fabs.f16(half [[TMP1]])
+; CHECK-NEXT:    ret half [[C]]
+;
+  %b = call float @llvm.fabs.f32(float %a)
+  %c = fptrunc fast float %b to half
+  ret half %c
+}
+
 define half @fneg_fptrunc(float %a) {
 ; CHECK-LABEL: @fneg_fptrunc(
 ; CHECK-NEXT:    [[TMP1:%.*]] = fptrunc float [[A:%.*]] to half
@@ -78,6 +89,28 @@ define half @test4-fast(float %a) {
 ; CHECK-NEXT:    [[TMP1:%.*]] = fptrunc fast float [[A:%.*]] to half
 ; CHECK-NEXT:    [[C:%.*]] = fneg fast half [[TMP1]]
 ; CHECK-NEXT:    ret half [[C]]
+;
+  %b = fsub fast float -0.0, %a
+  %c = fptrunc fast float %b to half
+  ret half %c
+}
+
+define half @test4-mixed-fast-1(float %a) {
+; CHECK-LABEL: @test4-mixed-fast-1(
+; CHECK-NEXT:    [[TMP1:%.*]] = fptrunc float [[A:%.*]] to half
+; CHECK-NEXT:    [[C:%.*]] = fneg half [[TMP1]]
+; CHECK-NEXT:    ret half [[C]]
+;
+  %b = fsub float -0.0, %a
+  %c = fptrunc fast float %b to half
+  ret half %c
+}
+
+define half @test4-mixed-fast-2(float %a) {
+; CHECK-LABEL: @test4-mixed-fast-2(
+; CHECK-NEXT:    [[TMP1:%.*]] = fptrunc float [[A:%.*]] to half
+; CHECK-NEXT:    [[C:%.*]] = fneg half [[TMP1]]
+; CHECK-NEXT:    ret half [[C]]
 ;
   %b = fsub fast float -0.0, %a
   %c = fptrunc float %b to half
@@ -89,6 +122,28 @@ define half @test4_unary_fneg-fast(float %a) {
 ; CHECK-NEXT:    [[TMP1:%.*]] = fptrunc fast float [[A:%.*]] to half
 ; CHECK-NEXT:    [[C:%.*]] = fneg fast half [[TMP1]]
 ; CHECK-NEXT:    ret half [[C]]
+;
+  %b = fneg fast float %a
+  %c = fptrunc fast float %b to half
+  ret half %c
+}
+
+define half @test4_unary_fneg-mixed-fast-1(float %a) {
+; CHECK-LABEL: @test4_unary_fneg-mixed-fast-1(
+; CHECK-NEXT:    [[TMP1:%.*]] = fptrunc float [[A:%.*]] to half
+; CHECK-NEXT:    [[C:%.*]] = fneg half [[TMP1]]
+; CHECK-NEXT:    ret half [[C]]
+;
+  %b = fneg float %a
+  %c = fptrunc fast float %b to half
+  ret half %c
+}
+
+define half @test4_unary_fneg-mixed-fast-2(float %a) {
+; CHECK-LABEL: @test4_unary_fneg-mixed-fast-2(
+; CHECK-NEXT:    [[TMP1:%.*]] = fptrunc float [[A:%.*]] to half
+; CHECK-NEXT:    [[C:%.*]] = fneg half [[TMP1]]
+; CHECK-NEXT:    ret half [[C]]
 ;
   %b = fneg fast float %a
   %c = fptrunc float %b to half
diff --git a/llvm/test/Transforms/InstCombine/fptrunc.ll b/llvm/test/Transforms/InstCombine/fptrunc.ll
index a4296a326c4bc6..0b5d8b3cd06e07 100644
--- a/llvm/test/Transforms/InstCombine/fptrunc.ll
+++ b/llvm/test/Transforms/InstCombine/fptrunc.ll
@@ -61,6 +61,18 @@ define float @fptrunc_select_true_val(float %x, double %y, i1 %cond) {
   ret float %r
 }
 
+define float @fptrunc_fast_select_true_val(float %x, double %y, i1 %cond) {
+; CHECK-LABEL: @fptrunc_fast_select_true_val(
+; CHECK-NEXT:    [[TMP1:%.*]] = fptrunc fast double [[Y:%.*]] to float
+; CHECK-NEXT:    [[NARROW_SEL:%.*]] = select i1 [[COND:%.*]], float [[TMP1]], float [[X:%.*]]
+; CHECK-NEXT:    ret float [[NARROW_SEL]]
+;
+  %e = fpext float %x to double
+  %sel = select fast i1 %cond, double %y, double %e
+  %r = fptrunc fast double %sel to float
+  ret float %r
+}
+
 define <2 x float> @fptrunc_select_false_val(<2 x float> %x, <2 x double> %y, <2 x i1> %cond) {
 ; CHECK-LABEL: @fptrunc_select_false_val(
 ; CHECK-NEXT:    [[TMP1:%.*]] = fptrunc <2 x double> [[Y:%.*]] to <2 x float>
@@ -73,6 +85,18 @@ define <2 x float> @fptrunc_select_false_val(<2 x float> %x, <2 x double> %y, <2
   ret <2 x float> %r
 }
 
+define <2 x float> @fptrunc_nnan_select_false_val(<2 x float> %x, <2 x double> %y, <2 x i1> %cond) {
+; CHECK-LABEL: @fptrunc_nnan_select_false_val(
+; CHECK-NEXT:    [[TMP1:%.*]] = fptrunc nnan <2 x double> [[Y:%.*]] to <2 x float>
+; CHECK-NEXT:    [[NARROW_SEL:%.*]] = select <2 x i1> [[COND:%.*]], <2 x float> [[X:%.*]], <2 x float> [[TMP1]]
+; CHECK-NEXT:    ret <2 x float> [[NARROW_SEL]]
+;
+  %e = fpext <2 x float> %x to <2 x double>
+  %sel = select nnan <2 x i1> %cond, <2 x double> %e, <2 x double> %y
+  %r = fptrunc nnan <2 x double> %sel to <2 x float>
+  ret <2 x float> %r
+}
+
 declare void @use(float)
 
 define half @fptrunc_select_true_val_extra_use(half %x, float %y, i1 %cond) {

[arsenm]

Description says union but this is intersection.Also dyn_cast to FPMathOperator and query flags from that, this is effectively isa + cast

[john-brawn-arm]

Yes, I meant intersection.

[arsenm]

Test a vector case, and with a subset of flags

[dtcxzyw]

LGTM.