[RISCV] Fix incorrect codegen for Zfa with negated forms of constants in the lookup table

[asb]

The logic in RISCVLoadFPImm::getLoadFPImm recognises that the only supported negative value is -1.0, but due to a typo returns false otherwise (entry 0, which is -1.0) rather than returning -1 (indicating no match found).

[llvmbot]

@llvm/pr-subscribers-backend-risc-v

Changes

The logic in RISCVLoadFPImm::getLoadFPImm recognises that the only supported negative value is -1.0, but due to a typo returns false otherwise (entry 0, which is -1.0) rather than returning -1 (indicating no match found).

---

Full diff: https://github.com/llvm/llvm-project/pull/68026.diff

4 Files Affected:

• (modified) llvm/lib/Target/RISCV/MCTargetDesc/RISCVBaseInfo.cpp (+1-1)
• (modified) llvm/test/CodeGen/RISCV/double-zfa.ll (+2-2)
• (modified) llvm/test/CodeGen/RISCV/float-zfa.ll (+2-2)
• (modified) llvm/test/CodeGen/RISCV/half-zfa-fli.ll (+2-2)

diff --git a/llvm/lib/Target/RISCV/MCTargetDesc/RISCVBaseInfo.cpp b/llvm/lib/Target/RISCV/MCTargetDesc/RISCVBaseInfo.cpp
index 0a42c6faee29008..d71efc11e6a9fcf 100644
--- a/llvm/lib/Target/RISCV/MCTargetDesc/RISCVBaseInfo.cpp
+++ b/llvm/lib/Target/RISCV/MCTargetDesc/RISCVBaseInfo.cpp
@@ -273,7 +273,7 @@ int RISCVLoadFPImm::getLoadFPImm(APFloat FPImm) {
   if (Sign) {
     if (Entry == 16)
       return 0;
-    return false;
+    return -1;
   }
 
   return Entry;
diff --git a/llvm/test/CodeGen/RISCV/double-zfa.ll b/llvm/test/CodeGen/RISCV/double-zfa.ll
index 00481fb367136f0..d81b7ddafd81851 100644
--- a/llvm/test/CodeGen/RISCV/double-zfa.ll
+++ b/llvm/test/CodeGen/RISCV/double-zfa.ll
@@ -143,11 +143,11 @@ define double @loadfpimm16() {
 
 ; Ensure fli isn't incorrectly used for negated versions of numbers in the fli
 ; table.
-; FIXME: Codegen is incorrect.
 define double @loadfpimm17() {
 ; CHECK-LABEL: loadfpimm17:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    fli.d fa0, -1.0
+; CHECK-NEXT:    lui a0, %hi(.LCPI15_0)
+; CHECK-NEXT:    fld fa0, %lo(.LCPI15_0)(a0)
 ; CHECK-NEXT:    ret
   ret double -2.0
 }
diff --git a/llvm/test/CodeGen/RISCV/float-zfa.ll b/llvm/test/CodeGen/RISCV/float-zfa.ll
index e41aa631d38a6ee..3b9043b0afe24b6 100644
--- a/llvm/test/CodeGen/RISCV/float-zfa.ll
+++ b/llvm/test/CodeGen/RISCV/float-zfa.ll
@@ -97,11 +97,11 @@ define float @loadfpimm11() {
 
 ; Ensure fli isn't incorrectly used for negated versions of numbers in the fli
 ; table.
-; FIXME: Codegen is incorrect.
 define float @loadfpimm12() {
 ; CHECK-LABEL: loadfpimm12:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    fli.s fa0, -1.0
+; CHECK-NEXT:    lui a0, 786432
+; CHECK-NEXT:    fmv.w.x fa0, a0
 ; CHECK-NEXT:    ret
   ret float -2.0
 }
diff --git a/llvm/test/CodeGen/RISCV/half-zfa-fli.ll b/llvm/test/CodeGen/RISCV/half-zfa-fli.ll
index 90d5bb4b53ba46d..cab15e58b0bb823 100644
--- a/llvm/test/CodeGen/RISCV/half-zfa-fli.ll
+++ b/llvm/test/CodeGen/RISCV/half-zfa-fli.ll
@@ -197,11 +197,11 @@ define half @loadfpimm13() {
 
 ; Ensure fli isn't incorrectly used for negated versions of numbers in the fli
 ; table.
-; FIXME: Codegen is incorrect when Zfa is enabled.
 define half @loadfpimm14() {
 ; CHECK-LABEL: loadfpimm14:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    fli.h fa0, -1.0
+; CHECK-NEXT:    lui a0, 1048572
+; CHECK-NEXT:    fmv.h.x fa0, a0
 ; CHECK-NEXT:    ret
 ;
 ; ZFHMIN-LABEL: loadfpimm14:

[topperc]

LGTM

[asb]

It looks like this was the cause of all the test failures I saw with the GCC torture suite and Zfa (not that the torture suite exercises it particularly well - only ~50 tests see changed codegen). I'll do a quick test with a few other simple code bases, but after that I don't think there are blockers on following up on #67964 with a PR to mark Zfa as non-experimental.

I wasn't planning on suggesting this for 17.0.2 as I don't think we have a history or expectation of backporting fixes to experimental extensions (and it might send the wrong message about the degree of support we offer for them), but let me know if you feel it would be worth doing.